Fix SPDX (#283)

Fix license field to be valid SPDX expression

.github/workflows/release.yaml

@@ -252,7 +252,7 @@ jobs:
           name: sqlite-vec-iossimulator-x86_64-extension
           path: dist/iossimulator-x86_64
       - run: make sqlite-vec.h
-      - uses: asg017/setup-sqlite-dist@73e37b2ffb0b51e64a64eb035da38c958b9ff6c6
+      - uses: asg017/setup-sqlite-dist@fadb0183a6ec70c3f1942de7d232b087ff2bacd1
       - run: sqlite-dist build --set-version $(cat VERSION)
       - run: |
           gh release upload ${{ github.ref_name }} \

.gitignore

@@ -31,3 +31,6 @@ poetry.lock
 
 memstat.c
 memstat.*
+
+
+.DS_Store

Makefile

@@ -37,11 +37,18 @@ endif
 
 ifndef OMIT_SIMD
 	ifeq ($(shell uname -sm),Darwin x86_64)
-	CFLAGS += -mavx -DSQLITE_VEC_ENABLE_AVX
+	CFLAGS += -mavx -mavx2 -DSQLITE_VEC_ENABLE_AVX
 	endif
 	ifeq ($(shell uname -sm),Darwin arm64)
 	CFLAGS += -mcpu=apple-m1 -DSQLITE_VEC_ENABLE_NEON
 	endif
+	ifeq ($(shell uname -s),Linux)
+	ifeq ($(findstring android,$(CC)),)
+	ifneq ($(filter avx,$(shell grep -o 'avx[^ ]*' /proc/cpuinfo 2>/dev/null | head -1)),)
+	CFLAGS += -mavx -mavx2 -DSQLITE_VEC_ENABLE_AVX
+	endif
+	endif
+	endif
 endif
 
 ifdef USE_BREW_SQLITE
@@ -155,6 +162,13 @@ clean:
 	rm -rf dist
 
 
+TARGET_AMALGAMATION=$(prefix)/sqlite-vec.c
+
+amalgamation: $(TARGET_AMALGAMATION)
+
+$(TARGET_AMALGAMATION): sqlite-vec.c $(wildcard sqlite-vec-*.c) scripts/amalgamate.py $(prefix)
+	python3 scripts/amalgamate.py sqlite-vec.c > $@
+
 FORMAT_FILES=sqlite-vec.h sqlite-vec.c
 format: $(FORMAT_FILES)
 	clang-format -i $(FORMAT_FILES)
@@ -174,7 +188,7 @@ evidence-of:
 test:
 	sqlite3 :memory: '.read test.sql'
 
-.PHONY: version loadable static test clean gh-release evidence-of install uninstall
+.PHONY: version loadable static test clean gh-release evidence-of install uninstall amalgamation
 
 publish-release:
 	./scripts/publish-release.sh
@@ -190,7 +204,22 @@ test-loadable-watch:
 	watchexec --exts c,py,Makefile --clear -- make test-loadable
 
 test-unit:
-	$(CC) -DSQLITE_CORE -DSQLITE_VEC_TEST tests/test-unit.c sqlite-vec.c vendor/sqlite3.c -I./ -Ivendor -o $(prefix)/test-unit && $(prefix)/test-unit
+	$(CC) -DSQLITE_CORE -DSQLITE_VEC_TEST -DSQLITE_VEC_ENABLE_RESCORE -DSQLITE_VEC_ENABLE_DISKANN=1 tests/test-unit.c sqlite-vec.c vendor/sqlite3.c -I./ -Ivendor $(CFLAGS) -o $(prefix)/test-unit && $(prefix)/test-unit
+
+# Standalone sqlite3 CLI with vec0 compiled in. Useful for benchmarking,
+# profiling (has debug symbols), and scripting without .load_extension.
+#   make cli
+#   dist/sqlite3 :memory: "SELECT vec_version()"
+#   dist/sqlite3 < script.sql
+cli: sqlite-vec.h $(prefix)
+	$(CC) -O2 -g \
+	-DSQLITE_CORE \
+	-DSQLITE_EXTRA_INIT=core_init \
+	-DSQLITE_THREADSAFE=0 \
+	-Ivendor/ -I./ \
+	$(CFLAGS) \
+	vendor/sqlite3.c vendor/shell.c sqlite-vec.c examples/sqlite3-cli/core_init.c \
+	-ldl -lm -o $(prefix)/sqlite3
 
 fuzz-build:
 	$(MAKE) -C tests/fuzz all

TODO.md

@@ -0,0 +1,73 @@
+# TODO: `ann` base branch + consolidated benchmarks
+
+## 1. Create `ann` branch with shared code
+
+### 1.1 Branch setup
+- [x] `git checkout -B ann origin/main`
+- [x] Cherry-pick `624f998` (vec0_distance_full shared distance dispatch)
+- [x] Cherry-pick stdint.h fix for test header
+- [ ] Pull NEON cosine optimization from ivf-yolo3 into shared code
+  - Currently only in ivf branch but is general-purpose (benefits all distance calcs)
+  - Lives in `distance_cosine_float()` — ~57 lines of ARM NEON vectorized cosine
+
+### 1.2 Benchmark infrastructure (`benchmarks-ann/`)
+- [x] Seed data pipeline (`seed/Makefile`, `seed/build_base_db.py`)
+- [x] Ground truth generator (`ground_truth.py`)
+- [x] Results schema (`schema.sql`)
+- [x] Benchmark runner with `INDEX_REGISTRY` extension point (`bench.py`)
+  - Baseline configs (float, int8-rescore, bit-rescore) implemented
+  - Index branches register their types via `INDEX_REGISTRY` dict
+- [x] Makefile with baseline targets
+- [x] README
+
+### 1.3 Rebase feature branches onto `ann`
+- [x] Rebase `diskann-yolo2` onto `ann` (1 commit: DiskANN implementation)
+- [x] Rebase `ivf-yolo3` onto `ann` (1 commit: IVF implementation)
+- [x] Rebase `annoy-yolo2` onto `ann` (2 commits: Annoy implementation + schema fix)
+- [x] Verify each branch has only its index-specific commits remaining
+- [ ] Force-push all 4 branches to origin
+
+---
+
+## 2. Per-branch: register index type in benchmarks
+
+Each index branch should add to `benchmarks-ann/` when rebased onto `ann`:
+
+### 2.1 Register in `bench.py`
+
+Add an `INDEX_REGISTRY` entry. Each entry provides:
+- `defaults` — default param values
+- `create_table_sql(params)` — CREATE VIRTUAL TABLE with INDEXED BY clause
+- `insert_sql(params)` — custom insert SQL, or None for default
+- `post_insert_hook(conn, params)` — training/building step, returns time
+- `run_query(conn, params, query, k)` — custom query, or None for default MATCH
+- `describe(params)` — one-line description for report output
+
+### 2.2 Add configs to `Makefile`
+
+Append index-specific config variables and targets. Example pattern:
+
+```makefile
+DISKANN_CONFIGS = \
+    "diskann-R48-binary:type=diskann,R=48,L=128,quantizer=binary" \
+    ...
+
+ALL_CONFIGS += $(DISKANN_CONFIGS)
+
+bench-diskann: seed
+    $(BENCH) --subset-size 10000 -k 10 -o runs/diskann $(BASELINES) $(DISKANN_CONFIGS)
+    ...
+```
+
+### 2.3 Migrate existing benchmark results/docs
+
+- Move useful results docs (RESULTS.md, etc.) into `benchmarks-ann/results/`
+- Delete redundant per-branch benchmark directories once consolidated infra is proven
+
+---
+
+## 3. Future improvements
+
+- [ ] Reporting script (`report.py`) — query results.db, produce markdown comparison tables
+- [ ] Profiling targets in Makefile (lift from ivf-yolo3's Instruments/perf wrappers)
+- [ ] Pre-computed ground truth integration (use GT DB files instead of on-the-fly brute-force)

VERSION

@@ -1 +1 @@
-0.1.7
+0.1.10-alpha.3

benchmarks-ann/.gitignore

@@ -0,0 +1,8 @@
+*.db
+*.db-shm
+*.db-wal
+*.parquet
+runs/
+
+viewer/
+searcher/

benchmarks-ann/Makefile

@@ -0,0 +1,85 @@
+BENCH = python bench.py
+BASE_DB = cohere1m/base.db
+EXT = ../dist/vec0
+
+# --- Baseline (brute-force) configs ---
+BASELINES = \
+	"brute-float:type=vec0-flat,variant=float" \
+	"brute-int8:type=vec0-flat,variant=int8" \
+	"brute-bit:type=vec0-flat,variant=bit"
+
+# --- IVF configs ---
+IVF_CONFIGS = \
+	"ivf-n32-p8:type=ivf,nlist=32,nprobe=8" \
+	"ivf-n128-p16:type=ivf,nlist=128,nprobe=16" \
+	"ivf-n512-p32:type=ivf,nlist=512,nprobe=32"
+
+RESCORE_CONFIGS = \
+	"rescore-bit-os8:type=rescore,quantizer=bit,oversample=8" \
+	"rescore-bit-os16:type=rescore,quantizer=bit,oversample=16" \
+	"rescore-int8-os8:type=rescore,quantizer=int8,oversample=8"
+
+# --- DiskANN configs ---
+DISKANN_CONFIGS = \
+	"diskann-R48-binary:type=diskann,R=48,L=128,quantizer=binary" \
+	"diskann-R72-binary:type=diskann,R=72,L=128,quantizer=binary" \
+	"diskann-R72-int8:type=diskann,R=72,L=128,quantizer=int8" \
+	"diskann-R72-L256:type=diskann,R=72,L=256,quantizer=binary"
+
+ALL_CONFIGS = $(BASELINES) $(RESCORE_CONFIGS) $(IVF_CONFIGS) $(DISKANN_CONFIGS)
+
+.PHONY: seed ground-truth bench-smoke bench-rescore bench-ivf bench-diskann bench-10k bench-50k bench-100k bench-all \
+        report clean
+
+# --- Data preparation ---
+seed:
+	$(MAKE) -C cohere1m
+
+ground-truth: seed
+	python ground_truth.py --subset-size 10000
+	python ground_truth.py --subset-size 50000
+	python ground_truth.py --subset-size 100000
+
+# --- Quick smoke test ---
+bench-smoke: seed
+	$(BENCH) --subset-size 5000 -k 10 -n 20 --dataset cohere1m -o runs \
+		"brute-float:type=vec0-flat,variant=float" \
+		"ivf-quick:type=ivf,nlist=16,nprobe=4" \
+		"diskann-quick:type=diskann,R=48,L=64,quantizer=binary"
+
+bench-rescore: seed
+	$(BENCH) --subset-size 10000 -k 10 --dataset cohere1m -o runs \
+		$(RESCORE_CONFIGS)
+
+
+# --- Standard sizes ---
+bench-10k: seed
+	$(BENCH) --subset-size 10000 -k 10 --dataset cohere1m -o runs $(ALL_CONFIGS)
+
+bench-50k: seed
+	$(BENCH) --subset-size 50000 -k 10 --dataset cohere1m -o runs $(ALL_CONFIGS)
+
+bench-100k: seed
+	$(BENCH) --subset-size 100000 -k 10 --dataset cohere1m -o runs $(ALL_CONFIGS)
+
+bench-all: bench-10k bench-50k bench-100k
+
+# --- IVF across sizes ---
+bench-ivf: seed
+	$(BENCH) --subset-size 10000 -k 10 --dataset cohere1m -o runs $(BASELINES) $(IVF_CONFIGS)
+	$(BENCH) --subset-size 50000 -k 10 --dataset cohere1m -o runs $(BASELINES) $(IVF_CONFIGS)
+	$(BENCH) --subset-size 100000 -k 10 --dataset cohere1m -o runs $(BASELINES) $(IVF_CONFIGS)
+
+# --- DiskANN across sizes ---
+bench-diskann: seed
+	$(BENCH) --subset-size 10000 -k 10 --dataset cohere1m -o runs $(BASELINES) $(DISKANN_CONFIGS)
+	$(BENCH) --subset-size 50000 -k 10 --dataset cohere1m -o runs $(BASELINES) $(DISKANN_CONFIGS)
+	$(BENCH) --subset-size 100000 -k 10 --dataset cohere1m -o runs $(BASELINES) $(DISKANN_CONFIGS)
+
+# --- Report ---
+report:
+	@echo "Use: sqlite3 runs/cohere1m/<size>/results.db 'SELECT run_id, config_name, status, recall FROM runs JOIN run_results USING(run_id)'"
+
+# --- Cleanup ---
+clean:
+	rm -rf runs/

benchmarks-ann/README.md

@@ -0,0 +1,111 @@
+# KNN Benchmarks for sqlite-vec
+
+Benchmarking infrastructure for vec0 KNN configurations. Includes brute-force
+baselines (float, int8, bit), rescore, IVF, and DiskANN index types.
+
+## Datasets
+
+Each dataset is a subdirectory containing a `Makefile` and `build_base_db.py`
+that produce a `base.db`. The benchmark runner auto-discovers any subdirectory
+with a `base.db` file.
+
+```
+cohere1m/           # Cohere 768d cosine, 1M vectors
+  Makefile          # downloads parquets from Zilliz, builds base.db
+  build_base_db.py
+  base.db           # (generated)
+
+cohere10m/          # Cohere 768d cosine, 10M vectors (10 train shards)
+  Makefile          # make -j12 download to fetch all shards in parallel
+  build_base_db.py
+  base.db           # (generated)
+```
+
+Every `base.db` has the same schema:
+
+| Table | Columns | Description |
+|-------|---------|-------------|
+| `train` | `id INTEGER PRIMARY KEY, vector BLOB` | Indexed vectors (f32 blobs) |
+| `query_vectors` | `id INTEGER PRIMARY KEY, vector BLOB` | Query vectors for KNN evaluation |
+| `neighbors` | `query_vector_id INTEGER, rank INTEGER, neighbors_id TEXT` | Ground-truth nearest neighbors |
+
+To add a new dataset, create a directory with a `Makefile` that builds `base.db`
+with the tables above. It will be available via `--dataset <dirname>` automatically.
+
+### Building datasets
+
+```bash
+# Cohere 1M
+cd cohere1m && make download && make && cd ..
+
+# Cohere 10M (parallel download recommended — 10 train shards + test + neighbors)
+cd cohere10m && make -j12 download && make && cd ..
+```
+
+## Prerequisites
+
+- Built `dist/vec0` extension (run `make loadable` from repo root)
+- Python 3.10+
+- `uv`
+
+## Quick start
+
+```bash
+# 1. Build a dataset
+cd cohere1m && make && cd ..
+
+# 2. Quick smoke test (5k vectors)
+make bench-smoke
+
+# 3. Full benchmark at 10k
+make bench-10k
+```
+
+## Usage
+
+```bash
+uv run python bench.py --subset-size 10000 -k 10 -n 50 --dataset cohere1m \
+  "brute-float:type=baseline,variant=float" \
+  "rescore-bit-os8:type=rescore,quantizer=bit,oversample=8"
+```
+
+### Config format
+
+`name:type=<index_type>,key=val,key=val`
+
+| Index type | Keys |
+|-----------|------|
+| `baseline` | `variant` (float/int8/bit), `oversample` |
+| `rescore` | `quantizer` (bit/int8), `oversample` |
+| `ivf` | `nlist`, `nprobe` |
+| `diskann` | `R`, `L`, `quantizer` (binary/int8), `buffer_threshold` |
+
+### Make targets
+
+| Target | Description |
+|--------|-------------|
+| `make seed` | Download and build default dataset |
+| `make bench-smoke` | Quick 5k test (3 configs) |
+| `make bench-10k` | All configs at 10k vectors |
+| `make bench-50k` | All configs at 50k vectors |
+| `make bench-100k` | All configs at 100k vectors |
+| `make bench-all` | 10k + 50k + 100k |
+| `make bench-ivf` | Baselines + IVF across 10k/50k/100k |
+| `make bench-diskann` | Baselines + DiskANN across 10k/50k/100k |
+
+## Results DB
+
+Each run writes to `runs/<dataset>/<subset_size>/results.db` (SQLite, WAL mode).
+Progress is written continuously — query from another terminal to monitor:
+
+```bash
+sqlite3 runs/cohere1m/10000/results.db "SELECT run_id, config_name, status FROM runs"
+```
+
+See `results_schema.sql` for the full schema (tables: `runs`, `run_results`,
+`insert_batches`, `queries`).
+
+## Adding an index type
+
+Add an entry to `INDEX_REGISTRY` in `bench.py` and append configs to
+`ALL_CONFIGS` in the `Makefile`. See existing entries for the pattern.

benchmarks-ann/bench-delete/.gitignore

@@ -0,0 +1,3 @@
+runs/
+*.db
+__pycache__/

benchmarks-ann/bench-delete/Makefile

@@ -0,0 +1,41 @@
+BENCH = python bench_delete.py
+EXT = ../../dist/vec0
+
+# --- Configs to test ---
+FLAT = "flat:type=vec0-flat,variant=float"
+RESCORE_BIT = "rescore-bit:type=rescore,quantizer=bit,oversample=8"
+RESCORE_INT8 = "rescore-int8:type=rescore,quantizer=int8,oversample=8"
+DISKANN_R48 = "diskann-R48:type=diskann,R=48,L=128,quantizer=binary"
+DISKANN_R72 = "diskann-R72:type=diskann,R=72,L=128,quantizer=binary"
+
+ALL_CONFIGS = $(FLAT) $(RESCORE_BIT) $(RESCORE_INT8) $(DISKANN_R48) $(DISKANN_R72)
+
+DELETE_PCTS = 5,10,25,50,75,90
+
+.PHONY: smoke bench-10k bench-50k bench-all report clean
+
+# Quick smoke test (small dataset, few queries)
+smoke:
+	$(BENCH) --subset-size 5000 --delete-pct 10,50 -k 10 -n 20 \
+		--dataset cohere1m --ext $(EXT) \
+		$(FLAT) $(DISKANN_R48)
+
+# Standard benchmarks
+bench-10k:
+	$(BENCH) --subset-size 10000 --delete-pct $(DELETE_PCTS) -k 10 -n 50 \
+		--dataset cohere1m --ext $(EXT) $(ALL_CONFIGS)
+
+bench-50k:
+	$(BENCH) --subset-size 50000 --delete-pct $(DELETE_PCTS) -k 10 -n 50 \
+		--dataset cohere1m --ext $(EXT) $(ALL_CONFIGS)
+
+bench-all: bench-10k bench-50k
+
+# Query saved results
+report:
+	@echo "Query results:"
+	@echo "  sqlite3 runs/cohere1m/10000/delete_results.db \\"
+	@echo "    \"SELECT config_name, delete_pct, recall, query_mean_ms, vacuum_size_mb FROM delete_runs ORDER BY config_name, delete_pct\""
+
+clean:
+	rm -rf runs/

benchmarks-ann/bench-delete/README.md

@@ -0,0 +1,69 @@
+# bench-delete: Recall degradation after random deletion
+
+Measures how KNN recall changes after deleting a random percentage of rows
+from different index types (flat, rescore, DiskANN).
+
+## Quick start
+
+```bash
+# Ensure dataset exists
+make -C ../datasets/cohere1m
+
+# Ensure extension is built
+make -C ../.. loadable
+
+# Quick smoke test
+make smoke
+
+# Full benchmark at 10k vectors
+make bench-10k
+```
+
+## Usage
+
+```bash
+python bench_delete.py --subset-size 10000 --delete-pct 10,25,50,75 \
+  "flat:type=vec0-flat,variant=float" \
+  "diskann-R72:type=diskann,R=72,L=128,quantizer=binary" \
+  "rescore-bit:type=rescore,quantizer=bit,oversample=8"
+```
+
+## What it measures
+
+For each config and delete percentage:
+
+| Metric | Description |
+|--------|-------------|
+| **recall** | KNN recall@k after deletion (ground truth recomputed over surviving rows) |
+| **delta** | Recall change vs 0% baseline |
+| **query latency** | Mean/median query time after deletion |
+| **db_size_mb** | DB file size before VACUUM |
+| **vacuum_size_mb** | DB file size after VACUUM (space reclaimed) |
+| **delete_time_s** | Wall time for the DELETE operations |
+
+## How it works
+
+1. Build index with N vectors (one copy per config)
+2. Measure recall at k=10 (pre-delete baseline)
+3. For each delete %:
+   - Copy the master DB
+   - Delete a random selection of rows (deterministic seed)
+   - Measure recall (ground truth recomputed over surviving rows only)
+   - VACUUM and measure size savings
+4. Print comparison table
+
+## Expected behavior
+
+- **Flat index**: Recall should be 1.0 at all delete percentages (brute-force is always exact)
+- **Rescore**: Recall should stay close to baseline (quantized scan + rescore is robust)
+- **DiskANN**: Recall may degrade at high delete % due to graph fragmentation (dangling edges, broken connectivity)
+
+## Results DB
+
+Results are stored in `runs/<dataset>/<subset_size>/delete_results.db`:
+
+```sql
+SELECT config_name, delete_pct, recall, vacuum_size_mb
+FROM delete_runs
+ORDER BY config_name, delete_pct;
+```

benchmarks-ann/bench-delete/bench_delete.py

@@ -0,0 +1,593 @@
+#!/usr/bin/env python3
+"""Benchmark: measure recall degradation after random row deletion.
+
+Given a dataset and index config, this script:
+  1. Builds the index (flat + ANN)
+  2. Measures recall at k=10 (pre-delete baseline)
+  3. Deletes a random % of rows
+  4. Measures recall again (post-delete)
+  5. Records DB size before/after deletion, recall delta, timings
+
+Usage:
+  python bench_delete.py --subset-size 10000 --delete-pct 25 \
+    "diskann-R48:type=diskann,R=48,L=128,quantizer=binary"
+
+  # Multiple delete percentages in one run:
+  python bench_delete.py --subset-size 10000 --delete-pct 10,25,50,75 \
+    "diskann-R48:type=diskann,R=48,L=128,quantizer=binary"
+"""
+import argparse
+import json
+import os
+import random
+import shutil
+import sqlite3
+import statistics
+import struct
+import time
+
+_SCRIPT_DIR = os.path.dirname(os.path.abspath(__file__))
+_BENCH_DIR = os.path.join(_SCRIPT_DIR, "..")
+_ROOT_DIR = os.path.join(_BENCH_DIR, "..")
+
+EXT_PATH = os.path.join(_ROOT_DIR, "dist", "vec0")
+DATASETS_DIR = os.path.join(_BENCH_DIR, "datasets")
+
+DATASETS = {
+    "cohere1m": {"base_db": os.path.join(DATASETS_DIR, "cohere1m", "base.db"), "dimensions": 768},
+    "cohere10m": {"base_db": os.path.join(DATASETS_DIR, "cohere10m", "base.db"), "dimensions": 768},
+    "nyt": {"base_db": os.path.join(DATASETS_DIR, "nyt", "base.db"), "dimensions": 256},
+    "nyt-768": {"base_db": os.path.join(DATASETS_DIR, "nyt-768", "base.db"), "dimensions": 768},
+    "nyt-1024": {"base_db": os.path.join(DATASETS_DIR, "nyt-1024", "base.db"), "dimensions": 1024},
+    "nyt-384": {"base_db": os.path.join(DATASETS_DIR, "nyt-384", "base.db"), "dimensions": 384},
+}
+
+INSERT_BATCH_SIZE = 1000
+
+
+# ============================================================================
+# Timing helpers
+# ============================================================================
+
+def now_ns():
+    return time.time_ns()
+
+def ns_to_s(ns):
+    return ns / 1_000_000_000
+
+def ns_to_ms(ns):
+    return ns / 1_000_000
+
+
+# ============================================================================
+# Index registry (subset of bench.py — only types relevant to deletion)
+# ============================================================================
+
+def _vec0_flat_create(p):
+    dims = p["dimensions"]
+    variant = p.get("variant", "float")
+    col = f"embedding float[{dims}]"
+    if variant == "int8":
+        col = f"embedding int8[{dims}]"
+    elif variant == "bit":
+        col = f"embedding bit[{dims}]"
+    return f"CREATE VIRTUAL TABLE vec_items USING vec0(id INTEGER PRIMARY KEY, {col})"
+
+def _rescore_create(p):
+    dims = p["dimensions"]
+    q = p.get("quantizer", "bit")
+    os_val = p.get("oversample", 8)
+    return (
+        f"CREATE VIRTUAL TABLE vec_items USING vec0("
+        f"id INTEGER PRIMARY KEY, "
+        f"embedding float[{dims}] indexed by rescore(quantizer={q}, oversample={os_val}))"
+    )
+
+def _diskann_create(p):
+    dims = p["dimensions"]
+    R = p.get("R", 72)
+    L = p.get("L", 128)
+    q = p.get("quantizer", "binary")
+    bt = p.get("buffer_threshold", 0)
+    sl_insert = p.get("search_list_size_insert", 0)
+    sl_search = p.get("search_list_size_search", 0)
+    parts = [
+        f"neighbor_quantizer={q}",
+        f"n_neighbors={R}",
+        f"buffer_threshold={bt}",
+    ]
+    if sl_insert or sl_search:
+        # Per-path overrides — don't also set search_list_size
+        if sl_insert:
+            parts.append(f"search_list_size_insert={sl_insert}")
+        if sl_search:
+            parts.append(f"search_list_size_search={sl_search}")
+    else:
+        parts.append(f"search_list_size={L}")
+    opts = ", ".join(parts)
+    return (
+        f"CREATE VIRTUAL TABLE vec_items USING vec0("
+        f"id INTEGER PRIMARY KEY, "
+        f"embedding float[{dims}] indexed by diskann({opts}))"
+    )
+
+def _ivf_create(p):
+    dims = p["dimensions"]
+    nlist = p.get("nlist", 128)
+    nprobe = p.get("nprobe", 16)
+    q = p.get("quantizer", "none")
+    os_val = p.get("oversample", 1)
+    parts = [f"nlist={nlist}", f"nprobe={nprobe}"]
+    if q != "none":
+        parts.append(f"quantizer={q}")
+    if os_val > 1:
+        parts.append(f"oversample={os_val}")
+    opts = ", ".join(parts)
+    return (
+        f"CREATE VIRTUAL TABLE vec_items USING vec0("
+        f"id INTEGER PRIMARY KEY, "
+        f"embedding float[{dims}] indexed by ivf({opts}))"
+    )
+
+
+INDEX_REGISTRY = {
+    "vec0-flat": {
+        "defaults": {"variant": "float"},
+        "create_table_sql": _vec0_flat_create,
+        "post_insert_hook": None,
+    },
+    "rescore": {
+        "defaults": {"quantizer": "bit", "oversample": 8},
+        "create_table_sql": _rescore_create,
+        "post_insert_hook": None,
+    },
+    "ivf": {
+        "defaults": {"nlist": 128, "nprobe": 16, "quantizer": "none",
+                      "oversample": 1},
+        "create_table_sql": _ivf_create,
+        "post_insert_hook": lambda conn, params: _ivf_train(conn),
+    },
+    "diskann": {
+        "defaults": {"R": 72, "L": 128, "quantizer": "binary",
+                      "buffer_threshold": 0},
+        "create_table_sql": _diskann_create,
+        "post_insert_hook": None,
+    },
+}
+
+
+def _ivf_train(conn):
+    """Trigger built-in k-means training for IVF."""
+    t0 = now_ns()
+    conn.execute("INSERT INTO vec_items(vec_items) VALUES ('compute-centroids')")
+    conn.commit()
+    return ns_to_s(now_ns() - t0)
+
+
+# ============================================================================
+# Config parsing (same format as bench.py)
+# ============================================================================
+
+INT_KEYS = {"R", "L", "oversample", "nlist", "nprobe", "buffer_threshold",
+            "search_list_size_insert", "search_list_size_search"}
+
+def parse_config(spec):
+    if ":" not in spec:
+        raise ValueError(f"Config must be 'name:key=val,...': {spec}")
+    name, rest = spec.split(":", 1)
+    params = {}
+    for kv in rest.split(","):
+        k, v = kv.split("=", 1)
+        k = k.strip()
+        v = v.strip()
+        if k in INT_KEYS:
+            v = int(v)
+        params[k] = v
+    index_type = params.pop("type", None)
+    if not index_type or index_type not in INDEX_REGISTRY:
+        raise ValueError(f"Unknown index type: {index_type}")
+    params["index_type"] = index_type
+    merged = dict(INDEX_REGISTRY[index_type]["defaults"])
+    merged.update(params)
+    return name, merged
+
+
+# ============================================================================
+# DB helpers
+# ============================================================================
+
+def create_bench_db(db_path, ext_path, base_db, page_size=4096):
+    if os.path.exists(db_path):
+        os.remove(db_path)
+    conn = sqlite3.connect(db_path)
+    conn.execute(f"PRAGMA page_size={page_size}")
+    conn.execute("PRAGMA journal_mode=WAL")
+    conn.enable_load_extension(True)
+    conn.load_extension(ext_path)
+    conn.execute(f"ATTACH DATABASE '{base_db}' AS base")
+    return conn
+
+
+def load_query_vectors(base_db, n):
+    conn = sqlite3.connect(base_db)
+    rows = conn.execute(
+        "SELECT id, vector FROM query_vectors LIMIT ?", (n,)
+    ).fetchall()
+    conn.close()
+    return rows
+
+
+def insert_loop(conn, subset_size, label, start_from=0):
+    insert_sql = (
+        "INSERT INTO vec_items(id, embedding) "
+        "SELECT id, vector FROM base.train "
+        "WHERE id >= :lo AND id < :hi"
+    )
+    total = 0
+    for lo in range(start_from, subset_size, INSERT_BATCH_SIZE):
+        hi = min(lo + INSERT_BATCH_SIZE, subset_size)
+        conn.execute(insert_sql, {"lo": lo, "hi": hi})
+        conn.commit()
+        total += hi - lo
+        if total % 5000 == 0 or total == subset_size - start_from:
+            print(f"    [{label}] inserted {total + start_from}/{subset_size}", flush=True)
+
+
+# ============================================================================
+# Recall measurement
+# ============================================================================
+
+def measure_recall(conn, base_db, query_vectors, subset_size, k, alive_ids=None):
+    """Measure KNN recall. If alive_ids is provided, ground truth is computed
+    only over those IDs (to match post-delete state)."""
+    recalls = []
+    times_ms = []
+
+    for qid, query in query_vectors:
+        t0 = now_ns()
+        results = conn.execute(
+            "SELECT id, distance FROM vec_items "
+            "WHERE embedding MATCH :query AND k = :k",
+            {"query": query, "k": k},
+        ).fetchall()
+        t1 = now_ns()
+        times_ms.append(ns_to_ms(t1 - t0))
+
+        result_ids = set(r[0] for r in results)
+
+        # Ground truth: brute-force cosine over surviving rows
+        if alive_ids is not None:
+            # After deletion — compute GT only over alive IDs
+            # Use a temp table for the alive set for efficiency
+            gt_rows = conn.execute(
+                "SELECT id FROM ("
+                "  SELECT id, vec_distance_l2(vector, :query) as dist "
+                "  FROM base.train WHERE id < :n ORDER BY dist LIMIT :k2"
+                ")",
+                {"query": query, "k2": k * 5, "n": subset_size},
+            ).fetchall()
+            # Filter to only alive IDs, take top k
+            gt_alive = [r[0] for r in gt_rows if r[0] in alive_ids][:k]
+            gt_ids = set(gt_alive)
+        else:
+            gt_rows = conn.execute(
+                "SELECT id FROM ("
+                "  SELECT id, vec_distance_l2(vector, :query) as dist "
+                "  FROM base.train WHERE id < :n ORDER BY dist LIMIT :k"
+                ")",
+                {"query": query, "k": k, "n": subset_size},
+            ).fetchall()
+            gt_ids = set(r[0] for r in gt_rows)
+
+        if gt_ids:
+            recalls.append(len(result_ids & gt_ids) / len(gt_ids))
+        else:
+            recalls.append(0.0)
+
+    return {
+        "recall": round(statistics.mean(recalls), 4) if recalls else 0.0,
+        "mean_ms": round(statistics.mean(times_ms), 2) if times_ms else 0.0,
+        "median_ms": round(statistics.median(times_ms), 2) if times_ms else 0.0,
+    }
+
+
+# ============================================================================
+# Delete benchmark core
+# ============================================================================
+
+def run_delete_benchmark(name, params, base_db, ext_path, subset_size, dims,
+                         delete_pcts, k, n_queries, out_dir, seed_val):
+    params["dimensions"] = dims
+    reg = INDEX_REGISTRY[params["index_type"]]
+    create_sql = reg["create_table_sql"](params)
+
+    results = []
+
+    # Build once, copy for each delete %
+    print(f"\n{'='*60}")
+    print(f"Config: {name}  (type={params['index_type']})")
+    print(f"{'='*60}")
+
+    os.makedirs(out_dir, exist_ok=True)
+    master_db_path = os.path.join(out_dir, f"{name}.{subset_size}.db")
+    print(f"  Building index ({subset_size} vectors)...")
+    build_t0 = now_ns()
+    conn = create_bench_db(master_db_path, ext_path, base_db)
+    conn.execute(create_sql)
+    insert_loop(conn, subset_size, name)
+    hook = reg.get("post_insert_hook")
+    if hook:
+        print(f"  Training...")
+        hook(conn, params)
+    conn.close()
+    build_time_s = ns_to_s(now_ns() - build_t0)
+    master_size = os.path.getsize(master_db_path)
+    print(f"  Built in {build_time_s:.1f}s  ({master_size / (1024*1024):.1f} MB)")
+
+    # Load query vectors once
+    query_vectors = load_query_vectors(base_db, n_queries)
+
+    # Measure pre-delete baseline on the master copy
+    print(f"\n  --- 0% deleted (baseline) ---")
+    conn = sqlite3.connect(master_db_path)
+    conn.enable_load_extension(True)
+    conn.load_extension(ext_path)
+    conn.execute(f"ATTACH DATABASE '{base_db}' AS base")
+    baseline = measure_recall(conn, base_db, query_vectors, subset_size, k)
+    conn.close()
+    print(f"  recall={baseline['recall']:.4f}  "
+          f"query={baseline['mean_ms']:.2f}ms")
+
+    results.append({
+        "name": name,
+        "index_type": params["index_type"],
+        "subset_size": subset_size,
+        "delete_pct": 0,
+        "n_deleted": 0,
+        "n_remaining": subset_size,
+        "recall": baseline["recall"],
+        "query_mean_ms": baseline["mean_ms"],
+        "query_median_ms": baseline["median_ms"],
+        "db_size_mb": round(master_size / (1024 * 1024), 2),
+        "build_time_s": round(build_time_s, 1),
+        "delete_time_s": 0.0,
+        "vacuum_size_mb": round(master_size / (1024 * 1024), 2),
+    })
+
+    # All IDs in the dataset
+    all_ids = list(range(subset_size))
+
+    for pct in sorted(delete_pcts):
+        n_delete = int(subset_size * pct / 100)
+        print(f"\n  --- {pct}% deleted ({n_delete} rows) ---")
+
+        # Copy master DB and work on the copy
+        copy_path = os.path.join(out_dir, f"{name}.{subset_size}.del{pct}.db")
+        shutil.copy2(master_db_path, copy_path)
+        # Also copy WAL/SHM if they exist
+        for suffix in ["-wal", "-shm"]:
+            src = master_db_path + suffix
+            if os.path.exists(src):
+                shutil.copy2(src, copy_path + suffix)
+
+        conn = sqlite3.connect(copy_path)
+        conn.enable_load_extension(True)
+        conn.load_extension(ext_path)
+        conn.execute(f"ATTACH DATABASE '{base_db}' AS base")
+
+        # Pick random IDs to delete (deterministic per pct)
+        rng = random.Random(seed_val + pct)
+        to_delete = set(rng.sample(all_ids, n_delete))
+        alive_ids = set(all_ids) - to_delete
+
+        # Delete
+        delete_t0 = now_ns()
+        batch = []
+        for i, rid in enumerate(to_delete):
+            batch.append(rid)
+            if len(batch) >= 500 or i == len(to_delete) - 1:
+                placeholders = ",".join("?" for _ in batch)
+                conn.execute(
+                    f"DELETE FROM vec_items WHERE id IN ({placeholders})",
+                    batch,
+                )
+                conn.commit()
+                batch = []
+        delete_time_s = ns_to_s(now_ns() - delete_t0)
+
+        remaining = conn.execute("SELECT count(*) FROM vec_items").fetchone()[0]
+        pre_vacuum_size = os.path.getsize(copy_path)
+        print(f"  deleted {n_delete} rows in {delete_time_s:.2f}s  "
+              f"({remaining} remaining)")
+
+        # Measure post-delete recall
+        post = measure_recall(conn, base_db, query_vectors, subset_size, k,
+                              alive_ids=alive_ids)
+        print(f"  recall={post['recall']:.4f}  "
+              f"(delta={post['recall'] - baseline['recall']:+.4f})  "
+              f"query={post['mean_ms']:.2f}ms")
+
+        # VACUUM and measure size savings — close fully, reopen without base
+        conn.close()
+        vconn = sqlite3.connect(copy_path)
+        vconn.execute("VACUUM")
+        vconn.close()
+        post_vacuum_size = os.path.getsize(copy_path)
+        saved_mb = (pre_vacuum_size - post_vacuum_size) / (1024 * 1024)
+        print(f"  size: {pre_vacuum_size/(1024*1024):.1f} MB -> "
+              f"{post_vacuum_size/(1024*1024):.1f} MB after VACUUM "
+              f"(saved {saved_mb:.1f} MB)")
+
+        results.append({
+            "name": name,
+            "index_type": params["index_type"],
+            "subset_size": subset_size,
+            "delete_pct": pct,
+            "n_deleted": n_delete,
+            "n_remaining": remaining,
+            "recall": post["recall"],
+            "query_mean_ms": post["mean_ms"],
+            "query_median_ms": post["median_ms"],
+            "db_size_mb": round(pre_vacuum_size / (1024 * 1024), 2),
+            "build_time_s": round(build_time_s, 1),
+            "delete_time_s": round(delete_time_s, 2),
+            "vacuum_size_mb": round(post_vacuum_size / (1024 * 1024), 2),
+        })
+
+    return results
+
+
+# ============================================================================
+# Results DB
+# ============================================================================
+
+RESULTS_SCHEMA = """\
+CREATE TABLE IF NOT EXISTS delete_runs (
+    run_id INTEGER PRIMARY KEY,
+    config_name TEXT NOT NULL,
+    index_type TEXT NOT NULL,
+    params TEXT,
+    dataset TEXT NOT NULL,
+    subset_size INTEGER NOT NULL,
+    delete_pct INTEGER NOT NULL,
+    n_deleted INTEGER NOT NULL,
+    n_remaining INTEGER NOT NULL,
+    k INTEGER NOT NULL,
+    n_queries INTEGER NOT NULL,
+    seed INTEGER NOT NULL,
+    recall REAL,
+    query_mean_ms REAL,
+    query_median_ms REAL,
+    db_size_mb REAL,
+    vacuum_size_mb REAL,
+    build_time_s REAL,
+    delete_time_s REAL,
+    created_at TEXT DEFAULT (datetime('now'))
+);
+"""
+
+def save_results(results, out_dir, dataset, subset_size, params_json, k, n_queries, seed_val):
+    db_path = os.path.join(out_dir, "delete_results.db")
+    db = sqlite3.connect(db_path)
+    db.execute("PRAGMA journal_mode=WAL")
+    db.executescript(RESULTS_SCHEMA)
+    for r in results:
+        db.execute(
+            "INSERT INTO delete_runs "
+            "(config_name, index_type, params, dataset, subset_size, "
+            " delete_pct, n_deleted, n_remaining, k, n_queries, seed, "
+            " recall, query_mean_ms, query_median_ms, "
+            " db_size_mb, vacuum_size_mb, build_time_s, delete_time_s) "
+            "VALUES (?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?)",
+            (
+                r["name"], r["index_type"], params_json, dataset, r["subset_size"],
+                r["delete_pct"], r["n_deleted"], r["n_remaining"], k, n_queries, seed_val,
+                r["recall"], r["query_mean_ms"], r["query_median_ms"],
+                r["db_size_mb"], r["vacuum_size_mb"], r["build_time_s"], r["delete_time_s"],
+            ),
+        )
+    db.commit()
+    db.close()
+    return db_path
+
+
+# ============================================================================
+# Reporting
+# ============================================================================
+
+def print_report(all_results):
+    print(f"\n{'name':>22} {'del%':>5} {'deleted':>8} {'remain':>8} "
+          f"{'recall':>7} {'delta':>7} {'qry(ms)':>8} "
+          f"{'size(MB)':>9} {'vacuumed':>9} {'del(s)':>7}")
+    print("-" * 110)
+
+    # Group by config name
+    configs = {}
+    for r in all_results:
+        configs.setdefault(r["name"], []).append(r)
+
+    for name, rows in configs.items():
+        baseline_recall = rows[0]["recall"]  # 0% delete is always first
+        for r in rows:
+            delta = r["recall"] - baseline_recall
+            delta_str = f"{delta:+.4f}" if r["delete_pct"] > 0 else "-"
+            print(
+                f"{r['name']:>22} {r['delete_pct']:>4}% "
+                f"{r['n_deleted']:>8} {r['n_remaining']:>8} "
+                f"{r['recall']:>7.4f} {delta_str:>7} {r['query_mean_ms']:>8.2f} "
+                f"{r['db_size_mb']:>9.1f} {r['vacuum_size_mb']:>9.1f} "
+                f"{r['delete_time_s']:>7.2f}"
+            )
+        print()
+
+
+# ============================================================================
+# Main
+# ============================================================================
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Benchmark recall degradation after random row deletion",
+        formatter_class=argparse.RawDescriptionHelpFormatter,
+        epilog=__doc__,
+    )
+    parser.add_argument("configs", nargs="+",
+                        help="config specs (name:type=X,key=val,...)")
+    parser.add_argument("--subset-size", type=int, default=10000,
+                        help="number of vectors to build (default: 10000)")
+    parser.add_argument("--delete-pct", type=str, default="10,25,50",
+                        help="comma-separated delete percentages (default: 10,25,50)")
+    parser.add_argument("-k", type=int, default=10, help="KNN k (default 10)")
+    parser.add_argument("-n", type=int, default=50,
+                        help="number of queries (default 50)")
+    parser.add_argument("--dataset", default="cohere1m",
+                        choices=list(DATASETS.keys()))
+    parser.add_argument("--ext", default=EXT_PATH)
+    parser.add_argument("-o", "--out-dir",
+                        default=os.path.join(_SCRIPT_DIR, "runs"))
+    parser.add_argument("--seed", type=int, default=42,
+                        help="random seed for delete selection (default: 42)")
+    args = parser.parse_args()
+
+    ds = DATASETS[args.dataset]
+    base_db = ds["base_db"]
+    dims = ds["dimensions"]
+    if not os.path.exists(base_db):
+        print(f"Error: dataset not found at {base_db}")
+        print(f"Run: make -C {os.path.dirname(base_db)}")
+        return 1
+
+    delete_pcts = [int(x.strip()) for x in args.delete_pct.split(",")]
+    for p in delete_pcts:
+        if not 0 < p < 100:
+            print(f"Error: delete percentage must be 1-99, got {p}")
+            return 1
+
+    out_dir = os.path.join(args.out_dir, args.dataset, str(args.subset_size))
+    os.makedirs(out_dir, exist_ok=True)
+
+    all_results = []
+    for spec in args.configs:
+        name, params = parse_config(spec)
+        params_json = json.dumps(params)
+        results = run_delete_benchmark(
+            name, params, base_db, args.ext, args.subset_size, dims,
+            delete_pcts, args.k, args.n, out_dir, args.seed,
+        )
+        all_results.extend(results)
+
+        save_results(results, out_dir, args.dataset, args.subset_size,
+                     params_json, args.k, args.n, args.seed)
+
+    print_report(all_results)
+
+    results_path = os.path.join(out_dir, "delete_results.db")
+    print(f"\nResults saved to: {results_path}")
+    print(f"Query: sqlite3 {results_path} "
+          f"\"SELECT config_name, delete_pct, recall, vacuum_size_mb "
+          f"FROM delete_runs ORDER BY config_name, delete_pct\"")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

benchmarks-ann/bench-delete/test_smoke.py

@@ -0,0 +1,124 @@
+#!/usr/bin/env python3
+"""Quick self-contained smoke test using a synthetic dataset.
+Creates a tiny base.db in a temp dir, runs the delete benchmark, verifies output.
+"""
+import os
+import random
+import sqlite3
+import struct
+import sys
+import tempfile
+
+_SCRIPT_DIR = os.path.dirname(os.path.abspath(__file__))
+_ROOT_DIR = os.path.join(_SCRIPT_DIR, "..", "..")
+EXT_PATH = os.path.join(_ROOT_DIR, "dist", "vec0")
+
+DIMS = 8
+N_TRAIN = 200
+N_QUERIES = 10
+K_NEIGHBORS = 5
+
+
+def _f32(vals):
+    return struct.pack(f"{len(vals)}f", *vals)
+
+
+def make_synthetic_base_db(path):
+    """Create a minimal base.db with train vectors and query vectors."""
+    rng = random.Random(123)
+    db = sqlite3.connect(path)
+    db.execute("CREATE TABLE train(id INTEGER PRIMARY KEY, vector BLOB)")
+    db.execute("CREATE TABLE query_vectors(id INTEGER PRIMARY KEY, vector BLOB)")
+
+    for i in range(N_TRAIN):
+        vec = [rng.gauss(0, 1) for _ in range(DIMS)]
+        db.execute("INSERT INTO train VALUES (?, ?)", (i, _f32(vec)))
+
+    for i in range(N_QUERIES):
+        vec = [rng.gauss(0, 1) for _ in range(DIMS)]
+        db.execute("INSERT INTO query_vectors VALUES (?, ?)", (i, _f32(vec)))
+
+    db.commit()
+    db.close()
+
+
+def main():
+    if not os.path.exists(EXT_PATH + ".dylib") and not os.path.exists(EXT_PATH + ".so"):
+        # Try bare path (sqlite handles extension)
+        pass
+
+    with tempfile.TemporaryDirectory() as tmpdir:
+        base_db = os.path.join(tmpdir, "base.db")
+        make_synthetic_base_db(base_db)
+
+        # Patch DATASETS to use our synthetic DB
+        import bench_delete
+        bench_delete.DATASETS["synthetic"] = {
+            "base_db": base_db,
+            "dimensions": DIMS,
+        }
+
+        out_dir = os.path.join(tmpdir, "runs")
+
+        # Test flat index
+        print("=== Testing flat index ===")
+        name, params = bench_delete.parse_config("flat:type=vec0-flat,variant=float")
+        params["dimensions"] = DIMS
+        results = bench_delete.run_delete_benchmark(
+            name, params, base_db, EXT_PATH,
+            subset_size=N_TRAIN, dims=DIMS,
+            delete_pcts=[25, 50], k=K_NEIGHBORS, n_queries=N_QUERIES,
+            out_dir=out_dir, seed_val=42,
+        )
+
+        bench_delete.print_report(results)
+
+        # Flat recall should be 1.0 at all delete %
+        for r in results:
+            assert r["recall"] == 1.0, \
+                f"Flat recall should be 1.0, got {r['recall']} at {r['delete_pct']}%"
+        print("\n  PASS: flat recall is 1.0 at all delete percentages\n")
+
+        # Test DiskANN
+        print("=== Testing DiskANN ===")
+        name2, params2 = bench_delete.parse_config(
+            "diskann:type=diskann,R=8,L=32,quantizer=binary"
+        )
+        params2["dimensions"] = DIMS
+        results2 = bench_delete.run_delete_benchmark(
+            name2, params2, base_db, EXT_PATH,
+            subset_size=N_TRAIN, dims=DIMS,
+            delete_pcts=[25, 50], k=K_NEIGHBORS, n_queries=N_QUERIES,
+            out_dir=out_dir, seed_val=42,
+        )
+
+        bench_delete.print_report(results2)
+
+        # DiskANN baseline (0%) should have decent recall
+        baseline = results2[0]
+        assert baseline["recall"] > 0.0, \
+            f"DiskANN baseline recall is zero"
+        print(f"  PASS: DiskANN baseline recall={baseline['recall']}")
+
+        # Test rescore
+        print("\n=== Testing rescore ===")
+        name3, params3 = bench_delete.parse_config(
+            "rescore:type=rescore,quantizer=bit,oversample=4"
+        )
+        params3["dimensions"] = DIMS
+        results3 = bench_delete.run_delete_benchmark(
+            name3, params3, base_db, EXT_PATH,
+            subset_size=N_TRAIN, dims=DIMS,
+            delete_pcts=[25, 50], k=K_NEIGHBORS, n_queries=N_QUERIES,
+            out_dir=out_dir, seed_val=42,
+        )
+
+        bench_delete.print_report(results3)
+        print(f"  PASS: rescore baseline recall={results3[0]['recall']}")
+
+    print("\n ALL SMOKE TESTS PASSED")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

benchmarks-ann/datasets/cohere10m/Makefile

@@ -0,0 +1,27 @@
+BASE_URL = https://assets.zilliz.com/benchmark/cohere_large_10m
+
+TRAIN_PARQUETS = $(shell printf 'train-%02d-of-10.parquet ' 0 1 2 3 4 5 6 7 8 9)
+OTHER_PARQUETS = test.parquet neighbors.parquet
+PARQUETS = $(TRAIN_PARQUETS) $(OTHER_PARQUETS)
+
+.PHONY: all download clean
+
+all: base.db
+
+# Use: make -j12 download
+download: $(PARQUETS)
+
+train-%-of-10.parquet:
+	curl -L -o $@ $(BASE_URL)/$@
+
+test.parquet:
+	curl -L -o $@ $(BASE_URL)/test.parquet
+
+neighbors.parquet:
+	curl -L -o $@ $(BASE_URL)/neighbors.parquet
+
+base.db: $(PARQUETS) build_base_db.py
+	uv run --with pandas --with pyarrow python build_base_db.py
+
+clean:
+	rm -f base.db

benchmarks-ann/datasets/cohere10m/build_base_db.py

@@ -0,0 +1,134 @@
+#!/usr/bin/env python3
+"""Build base.db from downloaded parquet files (10M dataset, 10 train shards).
+
+Reads train-00-of-10.parquet .. train-09-of-10.parquet, test.parquet,
+neighbors.parquet and creates a SQLite database with tables:
+  train, query_vectors, neighbors.
+
+Usage:
+  uv run --with pandas --with pyarrow python build_base_db.py
+"""
+import json
+import os
+import sqlite3
+import struct
+import sys
+import time
+
+import pandas as pd
+
+TRAIN_SHARDS = 10
+
+
+def float_list_to_blob(floats):
+    """Pack a list of floats into a little-endian f32 blob."""
+    return struct.pack(f"<{len(floats)}f", *floats)
+
+
+def main():
+    script_dir = os.path.dirname(os.path.abspath(__file__))
+    db_path = os.path.join(script_dir, "base.db")
+
+    train_paths = [
+        os.path.join(script_dir, f"train-{i:02d}-of-{TRAIN_SHARDS}.parquet")
+        for i in range(TRAIN_SHARDS)
+    ]
+    test_path = os.path.join(script_dir, "test.parquet")
+    neighbors_path = os.path.join(script_dir, "neighbors.parquet")
+
+    for p in train_paths + [test_path, neighbors_path]:
+        if not os.path.exists(p):
+            print(f"ERROR: {p} not found. Run 'make download' first.")
+            sys.exit(1)
+
+    if os.path.exists(db_path):
+        os.remove(db_path)
+
+    conn = sqlite3.connect(db_path)
+    conn.execute("PRAGMA journal_mode=WAL")
+    conn.execute("PRAGMA page_size=4096")
+
+    # --- query_vectors (from test.parquet) ---
+    print("Loading test.parquet (query vectors)...")
+    t0 = time.perf_counter()
+    df_test = pd.read_parquet(test_path)
+    conn.execute(
+        "CREATE TABLE query_vectors (id INTEGER PRIMARY KEY, vector BLOB)"
+    )
+    rows = []
+    for _, row in df_test.iterrows():
+        rows.append((int(row["id"]), float_list_to_blob(row["emb"])))
+    conn.executemany("INSERT INTO query_vectors (id, vector) VALUES (?, ?)", rows)
+    conn.commit()
+    print(f"  {len(rows)} query vectors in {time.perf_counter() - t0:.1f}s")
+
+    # --- neighbors (from neighbors.parquet) ---
+    print("Loading neighbors.parquet...")
+    t0 = time.perf_counter()
+    df_neighbors = pd.read_parquet(neighbors_path)
+    conn.execute(
+        "CREATE TABLE neighbors ("
+        "  query_vector_id INTEGER, rank INTEGER, neighbors_id TEXT,"
+        "  UNIQUE(query_vector_id, rank))"
+    )
+    rows = []
+    for _, row in df_neighbors.iterrows():
+        qid = int(row["id"])
+        nids = row["neighbors_id"]
+        if isinstance(nids, str):
+            nids = json.loads(nids)
+        for rank, nid in enumerate(nids):
+            rows.append((qid, rank, str(int(nid))))
+    conn.executemany(
+        "INSERT INTO neighbors (query_vector_id, rank, neighbors_id) VALUES (?, ?, ?)",
+        rows,
+    )
+    conn.commit()
+    print(f"  {len(rows)} neighbor rows in {time.perf_counter() - t0:.1f}s")
+
+    # --- train (from 10 shard parquets) ---
+    print(f"Loading {TRAIN_SHARDS} train shards (10M vectors, this will take a while)...")
+    conn.execute(
+        "CREATE TABLE train (id INTEGER PRIMARY KEY, vector BLOB)"
+    )
+
+    global_t0 = time.perf_counter()
+    total_inserted = 0
+    batch_size = 10000
+
+    for shard_idx, train_path in enumerate(train_paths):
+        print(f"  Shard {shard_idx + 1}/{TRAIN_SHARDS}: {os.path.basename(train_path)}")
+        t0 = time.perf_counter()
+        df = pd.read_parquet(train_path)
+        shard_len = len(df)
+
+        for start in range(0, shard_len, batch_size):
+            chunk = df.iloc[start : start + batch_size]
+            rows = []
+            for _, row in chunk.iterrows():
+                rows.append((int(row["id"]), float_list_to_blob(row["emb"])))
+            conn.executemany("INSERT INTO train (id, vector) VALUES (?, ?)", rows)
+            conn.commit()
+
+            total_inserted += len(rows)
+            if total_inserted % 100000 < batch_size:
+                elapsed = time.perf_counter() - global_t0
+                rate = total_inserted / elapsed if elapsed > 0 else 0
+                print(
+                    f"    {total_inserted:>10}  {elapsed:.0f}s  {rate:.0f} rows/s",
+                    flush=True,
+                )
+
+        shard_elapsed = time.perf_counter() - t0
+        print(f"    shard done: {shard_len} rows in {shard_elapsed:.1f}s")
+
+    elapsed = time.perf_counter() - global_t0
+    print(f"  {total_inserted} train vectors in {elapsed:.1f}s")
+
+    conn.close()
+    size_mb = os.path.getsize(db_path) / (1024 * 1024)
+    print(f"\nDone: {db_path} ({size_mb:.0f} MB)")
+
+
+if __name__ == "__main__":
+    main()

benchmarks-ann/datasets/cohere1m/.gitignore

@@ -0,0 +1,2 @@
+*.parquet
+base.db

benchmarks-ann/datasets/cohere1m/Makefile

@@ -0,0 +1,24 @@
+BASE_URL = https://assets.zilliz.com/benchmark/cohere_medium_1m
+
+PARQUETS = train.parquet test.parquet neighbors.parquet
+
+.PHONY: all download base.db clean
+
+all: base.db
+
+download: $(PARQUETS)
+
+train.parquet:
+	curl -L -o $@ $(BASE_URL)/train.parquet
+
+test.parquet:
+	curl -L -o $@ $(BASE_URL)/test.parquet
+
+neighbors.parquet:
+	curl -L -o $@ $(BASE_URL)/neighbors.parquet
+
+base.db: $(PARQUETS) build_base_db.py
+	uv run --with pandas --with pyarrow python build_base_db.py
+
+clean:
+	rm -f base.db

benchmarks-ann/datasets/cohere1m/build_base_db.py

@@ -0,0 +1,121 @@
+#!/usr/bin/env python3
+"""Build base.db from downloaded parquet files.
+
+Reads train.parquet, test.parquet, neighbors.parquet and creates a SQLite
+database with tables: train, query_vectors, neighbors.
+
+Usage:
+  uv run --with pandas --with pyarrow python build_base_db.py
+"""
+import json
+import os
+import sqlite3
+import struct
+import sys
+import time
+
+import pandas as pd
+
+
+def float_list_to_blob(floats):
+    """Pack a list of floats into a little-endian f32 blob."""
+    return struct.pack(f"<{len(floats)}f", *floats)
+
+
+def main():
+    seed_dir = os.path.dirname(os.path.abspath(__file__))
+    db_path = os.path.join(seed_dir, "base.db")
+
+    train_path = os.path.join(seed_dir, "train.parquet")
+    test_path = os.path.join(seed_dir, "test.parquet")
+    neighbors_path = os.path.join(seed_dir, "neighbors.parquet")
+
+    for p in (train_path, test_path, neighbors_path):
+        if not os.path.exists(p):
+            print(f"ERROR: {p} not found. Run 'make download' first.")
+            sys.exit(1)
+
+    if os.path.exists(db_path):
+        os.remove(db_path)
+
+    conn = sqlite3.connect(db_path)
+    conn.execute("PRAGMA journal_mode=WAL")
+    conn.execute("PRAGMA page_size=4096")
+
+    # --- query_vectors (from test.parquet) ---
+    print("Loading test.parquet (query vectors)...")
+    t0 = time.perf_counter()
+    df_test = pd.read_parquet(test_path)
+    conn.execute(
+        "CREATE TABLE query_vectors (id INTEGER PRIMARY KEY, vector BLOB)"
+    )
+    rows = []
+    for _, row in df_test.iterrows():
+        rows.append((int(row["id"]), float_list_to_blob(row["emb"])))
+    conn.executemany("INSERT INTO query_vectors (id, vector) VALUES (?, ?)", rows)
+    conn.commit()
+    print(f"  {len(rows)} query vectors in {time.perf_counter() - t0:.1f}s")
+
+    # --- neighbors (from neighbors.parquet) ---
+    print("Loading neighbors.parquet...")
+    t0 = time.perf_counter()
+    df_neighbors = pd.read_parquet(neighbors_path)
+    conn.execute(
+        "CREATE TABLE neighbors ("
+        "  query_vector_id INTEGER, rank INTEGER, neighbors_id TEXT,"
+        "  UNIQUE(query_vector_id, rank))"
+    )
+    rows = []
+    for _, row in df_neighbors.iterrows():
+        qid = int(row["id"])
+        # neighbors_id may be a numpy array or JSON string
+        nids = row["neighbors_id"]
+        if isinstance(nids, str):
+            nids = json.loads(nids)
+        for rank, nid in enumerate(nids):
+            rows.append((qid, rank, str(int(nid))))
+    conn.executemany(
+        "INSERT INTO neighbors (query_vector_id, rank, neighbors_id) VALUES (?, ?, ?)",
+        rows,
+    )
+    conn.commit()
+    print(f"  {len(rows)} neighbor rows in {time.perf_counter() - t0:.1f}s")
+
+    # --- train (from train.parquet) ---
+    print("Loading train.parquet (1M vectors, this takes a few minutes)...")
+    t0 = time.perf_counter()
+    conn.execute(
+        "CREATE TABLE train (id INTEGER PRIMARY KEY, vector BLOB)"
+    )
+
+    batch_size = 10000
+    df_iter = pd.read_parquet(train_path)
+    total = len(df_iter)
+
+    for start in range(0, total, batch_size):
+        chunk = df_iter.iloc[start : start + batch_size]
+        rows = []
+        for _, row in chunk.iterrows():
+            rows.append((int(row["id"]), float_list_to_blob(row["emb"])))
+        conn.executemany("INSERT INTO train (id, vector) VALUES (?, ?)", rows)
+        conn.commit()
+
+        done = min(start + batch_size, total)
+        elapsed = time.perf_counter() - t0
+        rate = done / elapsed if elapsed > 0 else 0
+        eta = (total - done) / rate if rate > 0 else 0
+        print(
+            f"    {done:>8}/{total}  {elapsed:.0f}s  {rate:.0f} rows/s  eta {eta:.0f}s",
+            flush=True,
+        )
+
+    elapsed = time.perf_counter() - t0
+    print(f"  {total} train vectors in {elapsed:.1f}s")
+
+    conn.close()
+    size_mb = os.path.getsize(db_path) / (1024 * 1024)
+    print(f"\nDone: {db_path} ({size_mb:.0f} MB)")
+
+
+if __name__ == "__main__":
+    main()

benchmarks-ann/datasets/nyt-1024/Makefile

@@ -0,0 +1,30 @@
+MODEL ?= mixedbread-ai/mxbai-embed-large-v1
+K ?= 100
+BATCH_SIZE ?= 256
+DATA_DIR ?= ../nyt/data
+
+all: base.db
+
+# Reuse data from ../nyt
+$(DATA_DIR):
+	$(MAKE) -C ../nyt data
+
+contents.db: $(DATA_DIR)
+	uv run ../nyt-768/build-contents.py --data-dir $(DATA_DIR) -o $@
+
+base.db: contents.db queries.txt
+	uv run build-base.py \
+		--contents-db contents.db \
+		--model $(MODEL) \
+		--queries-file queries.txt \
+		--batch-size $(BATCH_SIZE) \
+		--k $(K) \
+		-o $@
+
+queries.txt:
+	cp ../nyt/queries.txt $@
+
+clean:
+	rm -f base.db contents.db
+
+.PHONY: all clean

benchmarks-ann/datasets/nyt-1024/build-base.py

@@ -0,0 +1,163 @@
+# /// script
+# requires-python = ">=3.12"
+# dependencies = [
+#     "sentence-transformers",
+#     "torch<=2.7",
+#     "tqdm",
+# ]
+# ///
+
+import argparse
+import sqlite3
+from array import array
+from itertools import batched
+
+from sentence_transformers import SentenceTransformer
+from tqdm import tqdm
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Build base.db with train vectors, query vectors, and brute-force KNN neighbors",
+    )
+    parser.add_argument(
+        "--contents-db", "-c", default=None,
+        help="Path to contents.db (source of headlines and IDs)",
+    )
+    parser.add_argument(
+        "--model", "-m", default="mixedbread-ai/mxbai-embed-large-v1",
+        help="HuggingFace model ID (default: mixedbread-ai/mxbai-embed-large-v1)",
+    )
+    parser.add_argument(
+        "--queries-file", "-q", default="queries.txt",
+        help="Path to the queries file (default: queries.txt)",
+    )
+    parser.add_argument(
+        "--output", "-o", required=True,
+        help="Path to the output base.db",
+    )
+    parser.add_argument(
+        "--batch-size", "-b", type=int, default=256,
+        help="Batch size for embedding (default: 256)",
+    )
+    parser.add_argument(
+        "--k", "-k", type=int, default=100,
+        help="Number of nearest neighbors (default: 100)",
+    )
+    parser.add_argument(
+        "--limit", "-l", type=int, default=0,
+        help="Limit number of headlines to embed (0 = all)",
+    )
+    parser.add_argument(
+        "--vec-path", "-v", default="~/projects/sqlite-vec/dist/vec0",
+        help="Path to sqlite-vec extension (default: ~/projects/sqlite-vec/dist/vec0)",
+    )
+    parser.add_argument(
+        "--skip-neighbors", action="store_true",
+        help="Skip the brute-force KNN neighbor computation",
+    )
+    args = parser.parse_args()
+
+    import os
+    vec_path = os.path.expanduser(args.vec_path)
+
+    print(f"Loading model {args.model}...")
+    model = SentenceTransformer(args.model)
+
+    # Read headlines from contents.db
+    src = sqlite3.connect(args.contents_db)
+    limit_clause = f" LIMIT {args.limit}" if args.limit > 0 else ""
+    headlines = src.execute(
+        f"SELECT id, headline FROM contents ORDER BY id{limit_clause}"
+    ).fetchall()
+    src.close()
+    print(f"Loaded {len(headlines)} headlines from {args.contents_db}")
+
+    # Read queries
+    with open(args.queries_file) as f:
+        queries = [line.strip() for line in f if line.strip()]
+    print(f"Loaded {len(queries)} queries from {args.queries_file}")
+
+    # Create output database
+    db = sqlite3.connect(args.output)
+    db.enable_load_extension(True)
+    db.load_extension(vec_path)
+    db.enable_load_extension(False)
+
+    db.execute("CREATE TABLE IF NOT EXISTS train(id INTEGER PRIMARY KEY, vector BLOB)")
+    db.execute("CREATE TABLE IF NOT EXISTS query_vectors(id INTEGER PRIMARY KEY, vector BLOB)")
+    db.execute(
+        "CREATE TABLE IF NOT EXISTS neighbors("
+        "  query_vector_id INTEGER, rank INTEGER, neighbors_id TEXT,"
+        "  UNIQUE(query_vector_id, rank))"
+    )
+
+    # Step 1: Embed headlines -> train table
+    print("Embedding headlines...")
+    for batch in tqdm(
+        batched(headlines, args.batch_size),
+        total=(len(headlines) + args.batch_size - 1) // args.batch_size,
+    ):
+        ids = [r[0] for r in batch]
+        texts = [r[1] for r in batch]
+        embeddings = model.encode(texts, normalize_embeddings=True)
+
+        params = [
+            (int(rid), array("f", emb.tolist()).tobytes())
+            for rid, emb in zip(ids, embeddings)
+        ]
+        db.executemany("INSERT INTO train VALUES (?, ?)", params)
+        db.commit()
+
+    del headlines
+    n = db.execute("SELECT count(*) FROM train").fetchone()[0]
+    print(f"Embedded {n} headlines")
+
+    # Step 2: Embed queries -> query_vectors table
+    print("Embedding queries...")
+    query_embeddings = model.encode(queries, normalize_embeddings=True)
+    query_params = []
+    for i, emb in enumerate(query_embeddings, 1):
+        blob = array("f", emb.tolist()).tobytes()
+        query_params.append((i, blob))
+    db.executemany("INSERT INTO query_vectors VALUES (?, ?)", query_params)
+    db.commit()
+    print(f"Embedded {len(queries)} queries")
+
+    if args.skip_neighbors:
+        db.close()
+        print(f"Done (skipped neighbors). Wrote {args.output}")
+        return
+
+    # Step 3: Brute-force KNN via sqlite-vec -> neighbors table
+    n_queries = db.execute("SELECT count(*) FROM query_vectors").fetchone()[0]
+    print(f"Computing {args.k}-NN for {n_queries} queries via sqlite-vec...")
+    for query_id, query_blob in tqdm(
+        db.execute("SELECT id, vector FROM query_vectors").fetchall()
+    ):
+        results = db.execute(
+            """
+            SELECT
+                train.id,
+                vec_distance_cosine(train.vector, ?) AS distance
+            FROM train
+            WHERE distance IS NOT NULL
+            ORDER BY distance ASC
+            LIMIT ?
+            """,
+            (query_blob, args.k),
+        ).fetchall()
+
+        params = [
+            (query_id, rank, str(rid))
+            for rank, (rid, _dist) in enumerate(results)
+        ]
+        db.executemany("INSERT INTO neighbors VALUES (?, ?, ?)", params)
+
+    db.commit()
+    db.close()
+    print(f"Done. Wrote {args.output}")
+
+
+if __name__ == "__main__":
+    main()

benchmarks-ann/datasets/nyt-1024/queries.txt

@@ -0,0 +1,100 @@
+latest news on climate change policy
+presidential election results and analysis
+stock market crash causes
+coronavirus vaccine development updates
+artificial intelligence breakthrough in healthcare
+supreme court ruling on abortion rights
+tech companies layoff announcements
+earthquake damages in California
+cybersecurity breach at major corporation
+space exploration mission to Mars
+immigration reform legislation debate
+renewable energy investment trends
+healthcare costs rising across America
+protests against police brutality
+wildfires destroy homes in the West
+Olympic games highlights and records
+celebrity scandal rocks Hollywood
+breakthrough cancer treatment discovered
+housing market bubble concerns
+federal reserve interest rate decision
+school shooting tragedy response
+diplomatic tensions between superpowers
+drone strike kills terrorist leader
+social media platform faces regulation
+archaeological discovery reveals ancient civilization
+unemployment rate hits record low
+autonomous vehicles testing expansion
+streaming service launches original content
+opioid crisis intervention programs
+trade war tariffs impact economy
+infrastructure bill passes Congress
+data privacy concerns grow
+minimum wage increase proposal
+college admissions scandal exposed
+NFL player protest during anthem
+cryptocurrency regulation debate
+pandemic lockdown restrictions eased
+mass shooting gun control debate
+tax reform legislation impact
+ransomware attack cripples pipeline
+climate activists stage demonstration
+sports team wins championship
+banking system collapse fears
+pharmaceutical company fraud charges
+genetic engineering ethical concerns
+border wall funding controversy
+impeachment proceedings begin
+nuclear weapons treaty violation
+artificial meat alternative launch
+student loan debt forgiveness
+venture capital funding decline
+facial recognition ban proposed
+election interference investigation
+pandemic preparedness failures
+police reform measures announced
+wildfire prevention strategies
+ocean pollution crisis worsens
+manufacturing jobs returning
+pension fund shortfall concerns
+antitrust investigation launched
+voting rights protection act
+mental health awareness campaign
+homeless population increasing
+space debris collision risk
+drug cartel violence escalates
+renewable energy jobs growth
+infrastructure deterioration report
+vaccine mandate legal challenge
+cryptocurrency market volatility
+autonomous drone delivery service
+deep fake technology dangers
+Arctic ice melting accelerates
+income inequality gap widens
+election fraud claims disputed
+corporate merger blocked
+medical breakthrough extends life
+transportation strike disrupts city
+racial justice protests spread
+carbon emissions reduction goals
+financial crisis warning signs
+cyberbullying prevention efforts
+asteroid near miss with Earth
+gene therapy approval granted
+labor union organizing drive
+surveillance technology expansion
+education funding cuts proposed
+disaster relief efforts underway
+housing affordability crisis
+clean water access shortage
+artificial intelligence job displacement
+trade agreement negotiations
+prison reform initiative launched
+species extinction accelerates
+political corruption scandal
+terrorism threat level raised
+food safety contamination outbreak
+ai model release
+affordability interest rates
+peanut allergies in newbons
+breaking bad walter white

benchmarks-ann/datasets/nyt-384/Makefile

@@ -0,0 +1,29 @@
+MODEL ?= mixedbread-ai/mxbai-embed-xsmall-v1
+K ?= 100
+BATCH_SIZE ?= 512
+DATA_DIR ?= ../nyt/data
+
+all: base.db
+
+$(DATA_DIR):
+	$(MAKE) -C ../nyt data
+
+contents.db: $(DATA_DIR)
+	uv run ../nyt-768/build-contents.py --data-dir $(DATA_DIR) -o $@
+
+base.db: contents.db queries.txt
+	uv run ../nyt-1024/build-base.py \
+		--contents-db contents.db \
+		--model $(MODEL) \
+		--queries-file queries.txt \
+		--batch-size $(BATCH_SIZE) \
+		--k $(K) \
+		-o $@
+
+queries.txt:
+	cp ../nyt/queries.txt $@
+
+clean:
+	rm -f base.db contents.db
+
+.PHONY: all clean

benchmarks-ann/datasets/nyt-384/queries.txt

@@ -0,0 +1,100 @@
+latest news on climate change policy
+presidential election results and analysis
+stock market crash causes
+coronavirus vaccine development updates
+artificial intelligence breakthrough in healthcare
+supreme court ruling on abortion rights
+tech companies layoff announcements
+earthquake damages in California
+cybersecurity breach at major corporation
+space exploration mission to Mars
+immigration reform legislation debate
+renewable energy investment trends
+healthcare costs rising across America
+protests against police brutality
+wildfires destroy homes in the West
+Olympic games highlights and records
+celebrity scandal rocks Hollywood
+breakthrough cancer treatment discovered
+housing market bubble concerns
+federal reserve interest rate decision
+school shooting tragedy response
+diplomatic tensions between superpowers
+drone strike kills terrorist leader
+social media platform faces regulation
+archaeological discovery reveals ancient civilization
+unemployment rate hits record low
+autonomous vehicles testing expansion
+streaming service launches original content
+opioid crisis intervention programs
+trade war tariffs impact economy
+infrastructure bill passes Congress
+data privacy concerns grow
+minimum wage increase proposal
+college admissions scandal exposed
+NFL player protest during anthem
+cryptocurrency regulation debate
+pandemic lockdown restrictions eased
+mass shooting gun control debate
+tax reform legislation impact
+ransomware attack cripples pipeline
+climate activists stage demonstration
+sports team wins championship
+banking system collapse fears
+pharmaceutical company fraud charges
+genetic engineering ethical concerns
+border wall funding controversy
+impeachment proceedings begin
+nuclear weapons treaty violation
+artificial meat alternative launch
+student loan debt forgiveness
+venture capital funding decline
+facial recognition ban proposed
+election interference investigation
+pandemic preparedness failures
+police reform measures announced
+wildfire prevention strategies
+ocean pollution crisis worsens
+manufacturing jobs returning
+pension fund shortfall concerns
+antitrust investigation launched
+voting rights protection act
+mental health awareness campaign
+homeless population increasing
+space debris collision risk
+drug cartel violence escalates
+renewable energy jobs growth
+infrastructure deterioration report
+vaccine mandate legal challenge
+cryptocurrency market volatility
+autonomous drone delivery service
+deep fake technology dangers
+Arctic ice melting accelerates
+income inequality gap widens
+election fraud claims disputed
+corporate merger blocked
+medical breakthrough extends life
+transportation strike disrupts city
+racial justice protests spread
+carbon emissions reduction goals
+financial crisis warning signs
+cyberbullying prevention efforts
+asteroid near miss with Earth
+gene therapy approval granted
+labor union organizing drive
+surveillance technology expansion
+education funding cuts proposed
+disaster relief efforts underway
+housing affordability crisis
+clean water access shortage
+artificial intelligence job displacement
+trade agreement negotiations
+prison reform initiative launched
+species extinction accelerates
+political corruption scandal
+terrorism threat level raised
+food safety contamination outbreak
+ai model release
+affordability interest rates
+peanut allergies in newbons
+breaking bad walter white

benchmarks-ann/datasets/nyt-768/Makefile

@@ -0,0 +1,37 @@
+MODEL ?= bge-base-en-v1.5-768
+K ?= 100
+BATCH_SIZE ?= 512
+DATA_DIR ?= ../nyt/data
+
+all: base.db
+
+# Reuse data from ../nyt
+$(DATA_DIR):
+	$(MAKE) -C ../nyt data
+
+# Distill model (separate step, may take a while)
+$(MODEL):
+	uv run distill-model.py
+
+contents.db: $(DATA_DIR)
+	uv run build-contents.py --data-dir $(DATA_DIR) -o $@
+
+base.db: contents.db queries.txt $(MODEL)
+	uv run ../nyt/build-base.py \
+		--contents-db contents.db \
+		--model $(MODEL) \
+		--queries-file queries.txt \
+		--batch-size $(BATCH_SIZE) \
+		--k $(K) \
+		-o $@
+
+queries.txt:
+	cp ../nyt/queries.txt $@
+
+clean:
+	rm -f base.db contents.db
+
+clean-all: clean
+	rm -rf $(MODEL)
+
+.PHONY: all clean clean-all

benchmarks-ann/datasets/nyt-768/build-contents.py

@@ -0,0 +1,64 @@
+# /// script
+# requires-python = ">=3.12"
+# dependencies = [
+#     "duckdb",
+# ]
+# ///
+
+import argparse
+import sqlite3
+import duckdb
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Load NYT headline CSVs into a SQLite contents database (most recent 1M, deduplicated)",
+    )
+    parser.add_argument(
+        "--data-dir", "-d", default="../nyt/data",
+        help="Directory containing NYT CSV files (default: ../nyt/data)",
+    )
+    parser.add_argument(
+        "--limit", "-l", type=int, default=1_000_000,
+        help="Maximum number of headlines to keep (default: 1000000)",
+    )
+    parser.add_argument(
+        "--output", "-o", required=True,
+        help="Path to the output SQLite database",
+    )
+    args = parser.parse_args()
+
+    glob_pattern = f"{args.data_dir}/new_york_times_stories_*.csv"
+
+    con = duckdb.connect()
+    rows = con.execute(
+        f"""
+        WITH deduped AS (
+            SELECT
+                headline,
+                max(pub_date) AS pub_date
+            FROM read_csv('{glob_pattern}', auto_detect=true, union_by_name=true)
+            WHERE headline IS NOT NULL AND trim(headline) != ''
+            GROUP BY headline
+        )
+        SELECT
+            row_number() OVER (ORDER BY pub_date DESC) AS id,
+            headline
+        FROM deduped
+        ORDER BY pub_date DESC
+        LIMIT {args.limit}
+        """
+    ).fetchall()
+    con.close()
+
+    db = sqlite3.connect(args.output)
+    db.execute("CREATE TABLE contents(id INTEGER PRIMARY KEY, headline TEXT)")
+    db.executemany("INSERT INTO contents VALUES (?, ?)", rows)
+    db.commit()
+    db.close()
+
+    print(f"Wrote {len(rows)} headlines to {args.output}")
+
+
+if __name__ == "__main__":
+    main()

benchmarks-ann/datasets/nyt-768/distill-model.py

@@ -0,0 +1,13 @@
+# /// script
+# requires-python = ">=3.12"
+# dependencies = [
+#     "model2vec[distill]",
+#     "torch<=2.7",
+# ]
+# ///
+
+from model2vec.distill import distill
+
+model = distill(model_name="BAAI/bge-base-en-v1.5", pca_dims=768)
+model.save_pretrained("bge-base-en-v1.5-768")
+print("Saved distilled model to bge-base-en-v1.5-768/")

benchmarks-ann/datasets/nyt-768/queries.txt

@@ -0,0 +1,100 @@
+latest news on climate change policy
+presidential election results and analysis
+stock market crash causes
+coronavirus vaccine development updates
+artificial intelligence breakthrough in healthcare
+supreme court ruling on abortion rights
+tech companies layoff announcements
+earthquake damages in California
+cybersecurity breach at major corporation
+space exploration mission to Mars
+immigration reform legislation debate
+renewable energy investment trends
+healthcare costs rising across America
+protests against police brutality
+wildfires destroy homes in the West
+Olympic games highlights and records
+celebrity scandal rocks Hollywood
+breakthrough cancer treatment discovered
+housing market bubble concerns
+federal reserve interest rate decision
+school shooting tragedy response
+diplomatic tensions between superpowers
+drone strike kills terrorist leader
+social media platform faces regulation
+archaeological discovery reveals ancient civilization
+unemployment rate hits record low
+autonomous vehicles testing expansion
+streaming service launches original content
+opioid crisis intervention programs
+trade war tariffs impact economy
+infrastructure bill passes Congress
+data privacy concerns grow
+minimum wage increase proposal
+college admissions scandal exposed
+NFL player protest during anthem
+cryptocurrency regulation debate
+pandemic lockdown restrictions eased
+mass shooting gun control debate
+tax reform legislation impact
+ransomware attack cripples pipeline
+climate activists stage demonstration
+sports team wins championship
+banking system collapse fears
+pharmaceutical company fraud charges
+genetic engineering ethical concerns
+border wall funding controversy
+impeachment proceedings begin
+nuclear weapons treaty violation
+artificial meat alternative launch
+student loan debt forgiveness
+venture capital funding decline
+facial recognition ban proposed
+election interference investigation
+pandemic preparedness failures
+police reform measures announced
+wildfire prevention strategies
+ocean pollution crisis worsens
+manufacturing jobs returning
+pension fund shortfall concerns
+antitrust investigation launched
+voting rights protection act
+mental health awareness campaign
+homeless population increasing
+space debris collision risk
+drug cartel violence escalates
+renewable energy jobs growth
+infrastructure deterioration report
+vaccine mandate legal challenge
+cryptocurrency market volatility
+autonomous drone delivery service
+deep fake technology dangers
+Arctic ice melting accelerates
+income inequality gap widens
+election fraud claims disputed
+corporate merger blocked
+medical breakthrough extends life
+transportation strike disrupts city
+racial justice protests spread
+carbon emissions reduction goals
+financial crisis warning signs
+cyberbullying prevention efforts
+asteroid near miss with Earth
+gene therapy approval granted
+labor union organizing drive
+surveillance technology expansion
+education funding cuts proposed
+disaster relief efforts underway
+housing affordability crisis
+clean water access shortage
+artificial intelligence job displacement
+trade agreement negotiations
+prison reform initiative launched
+species extinction accelerates
+political corruption scandal
+terrorism threat level raised
+food safety contamination outbreak
+ai model release
+affordability interest rates
+peanut allergies in newbons
+breaking bad walter white

benchmarks-ann/datasets/nyt/.gitignore

@@ -0,0 +1 @@
+data/

benchmarks-ann/datasets/nyt/Makefile

@@ -0,0 +1,30 @@
+MODEL ?= minishlab/potion-base-8M
+K ?= 100
+BATCH_SIZE ?= 512
+DATA_DIR ?= data
+
+all: base.db contents.db
+
+# Download NYT headlines CSVs from Kaggle (requires `kaggle` CLI + API token)
+$(DATA_DIR):
+	kaggle datasets download -d johnbandy/new-york-times-headlines -p $(DATA_DIR) --unzip
+
+contents.db: $(DATA_DIR)
+	uv run build-contents.py --data-dir $(DATA_DIR) -o $@
+
+base.db: contents.db queries.txt
+	uv run build-base.py \
+		--contents-db contents.db \
+		--model $(MODEL) \
+		--queries-file queries.txt \
+		--batch-size $(BATCH_SIZE) \
+		--k $(K) \
+		-o $@
+
+clean:
+	rm -f base.db contents.db
+
+clean-all: clean
+	rm -rf $(DATA_DIR)
+
+.PHONY: all clean clean-all

benchmarks-ann/datasets/nyt/build-base.py

@@ -0,0 +1,165 @@
+# /// script
+# requires-python = ">=3.12"
+# dependencies = [
+#     "model2vec",
+#     "torch<=2.7",
+#     "tqdm",
+# ]
+# ///
+
+import argparse
+import sqlite3
+from array import array
+from itertools import batched
+
+from model2vec import StaticModel
+from tqdm import tqdm
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Build base.db with train vectors, query vectors, and brute-force KNN neighbors",
+    )
+    parser.add_argument(
+        "--contents-db", "-c", default=None,
+        help="Path to contents.db (source of headlines and IDs)",
+    )
+    parser.add_argument(
+        "--model", "-m", default="minishlab/potion-base-8M",
+        help="HuggingFace model ID or local path (default: minishlab/potion-base-8M)",
+    )
+    parser.add_argument(
+        "--queries-file", "-q", default="queries.txt",
+        help="Path to the queries file (default: queries.txt)",
+    )
+    parser.add_argument(
+        "--output", "-o", required=True,
+        help="Path to the output base.db",
+    )
+    parser.add_argument(
+        "--batch-size", "-b", type=int, default=512,
+        help="Batch size for embedding (default: 512)",
+    )
+    parser.add_argument(
+        "--k", "-k", type=int, default=100,
+        help="Number of nearest neighbors (default: 100)",
+    )
+    parser.add_argument(
+        "--vec-path", "-v", default="~/projects/sqlite-vec/dist/vec0",
+        help="Path to sqlite-vec extension (default: ~/projects/sqlite-vec/dist/vec0)",
+    )
+    parser.add_argument(
+        "--rebuild-neighbors", action="store_true",
+        help="Only rebuild the neighbors table (skip embedding steps)",
+    )
+    args = parser.parse_args()
+
+    import os
+    vec_path = os.path.expanduser(args.vec_path)
+
+    if args.rebuild_neighbors:
+        # Skip embedding, just open existing DB and rebuild neighbors
+        db = sqlite3.connect(args.output)
+        db.enable_load_extension(True)
+        db.load_extension(vec_path)
+        db.enable_load_extension(False)
+        db.execute("DROP TABLE IF EXISTS neighbors")
+        db.execute(
+            "CREATE TABLE neighbors("
+            "  query_vector_id INTEGER, rank INTEGER, neighbors_id TEXT,"
+            "  UNIQUE(query_vector_id, rank))"
+        )
+        print(f"Rebuilding neighbors in {args.output}...")
+    else:
+        print(f"Loading model {args.model}...")
+        model = StaticModel.from_pretrained(args.model)
+
+        # Read headlines from contents.db
+        src = sqlite3.connect(args.contents_db)
+        headlines = src.execute("SELECT id, headline FROM contents ORDER BY id").fetchall()
+        src.close()
+        print(f"Loaded {len(headlines)} headlines from {args.contents_db}")
+
+        # Read queries
+        with open(args.queries_file) as f:
+            queries = [line.strip() for line in f if line.strip()]
+        print(f"Loaded {len(queries)} queries from {args.queries_file}")
+
+        # Create output database
+        db = sqlite3.connect(args.output)
+        db.enable_load_extension(True)
+        db.load_extension(vec_path)
+        db.enable_load_extension(False)
+
+        db.execute("CREATE TABLE train(id INTEGER PRIMARY KEY, vector BLOB)")
+        db.execute("CREATE TABLE query_vectors(id INTEGER PRIMARY KEY, vector BLOB)")
+        db.execute(
+            "CREATE TABLE neighbors("
+            "  query_vector_id INTEGER, rank INTEGER, neighbors_id TEXT,"
+            "  UNIQUE(query_vector_id, rank))"
+        )
+
+        # Step 1: Embed headlines -> train table
+        print("Embedding headlines...")
+        for batch in tqdm(
+            batched(headlines, args.batch_size),
+            total=(len(headlines) + args.batch_size - 1) // args.batch_size,
+        ):
+            ids = [r[0] for r in batch]
+            texts = [r[1] for r in batch]
+            embeddings = model.encode(texts)
+
+            params = [
+                (int(rid), array("f", emb.tolist()).tobytes())
+                for rid, emb in zip(ids, embeddings)
+            ]
+            db.executemany("INSERT INTO train VALUES (?, ?)", params)
+            db.commit()
+
+        del headlines
+        n = db.execute("SELECT count(*) FROM train").fetchone()[0]
+        print(f"Embedded {n} headlines")
+
+        # Step 2: Embed queries -> query_vectors table
+        print("Embedding queries...")
+        query_embeddings = model.encode(queries)
+        query_params = []
+        for i, emb in enumerate(query_embeddings, 1):
+            blob = array("f", emb.tolist()).tobytes()
+            query_params.append((i, blob))
+        db.executemany("INSERT INTO query_vectors VALUES (?, ?)", query_params)
+        db.commit()
+        print(f"Embedded {len(queries)} queries")
+
+    # Step 3: Brute-force KNN via sqlite-vec -> neighbors table
+    n_queries = db.execute("SELECT count(*) FROM query_vectors").fetchone()[0]
+    print(f"Computing {args.k}-NN for {n_queries} queries via sqlite-vec...")
+    for query_id, query_blob in tqdm(
+        db.execute("SELECT id, vector FROM query_vectors").fetchall()
+    ):
+        results = db.execute(
+            """
+            SELECT
+                train.id,
+                vec_distance_cosine(train.vector, ?) AS distance
+            FROM train
+            WHERE distance IS NOT NULL
+            ORDER BY distance ASC
+            LIMIT ?
+            """,
+            (query_blob, args.k),
+        ).fetchall()
+
+        params = [
+            (query_id, rank, str(rid))
+            for rank, (rid, _dist) in enumerate(results)
+        ]
+        db.executemany("INSERT INTO neighbors VALUES (?, ?, ?)", params)
+
+    db.commit()
+    db.close()
+    print(f"Done. Wrote {args.output}")
+
+
+if __name__ == "__main__":
+    main()

benchmarks-ann/datasets/nyt/build-contents.py

@@ -0,0 +1,52 @@
+# /// script
+# requires-python = ">=3.12"
+# dependencies = [
+#     "duckdb",
+# ]
+# ///
+
+import argparse
+import os
+import sqlite3
+import duckdb
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Load NYT headline CSVs into a SQLite contents database via DuckDB",
+    )
+    parser.add_argument(
+        "--data-dir", "-d", default="data",
+        help="Directory containing NYT CSV files (default: data)",
+    )
+    parser.add_argument(
+        "--output", "-o", required=True,
+        help="Path to the output SQLite database",
+    )
+    args = parser.parse_args()
+
+    glob_pattern = os.path.join(args.data_dir, "new_york_times_stories_*.csv")
+
+    con = duckdb.connect()
+    rows = con.execute(
+        f"""
+        SELECT
+            row_number() OVER () AS id,
+            headline
+        FROM read_csv('{glob_pattern}', auto_detect=true, union_by_name=true)
+        WHERE headline IS NOT NULL AND headline != ''
+        """
+    ).fetchall()
+    con.close()
+
+    db = sqlite3.connect(args.output)
+    db.execute("CREATE TABLE contents(id INTEGER PRIMARY KEY, headline TEXT)")
+    db.executemany("INSERT INTO contents VALUES (?, ?)", rows)
+    db.commit()
+    db.close()
+
+    print(f"Wrote {len(rows)} headlines to {args.output}")
+
+
+if __name__ == "__main__":
+    main()

benchmarks-ann/datasets/nyt/queries.txt

@@ -0,0 +1,100 @@
+latest news on climate change policy
+presidential election results and analysis
+stock market crash causes
+coronavirus vaccine development updates
+artificial intelligence breakthrough in healthcare
+supreme court ruling on abortion rights
+tech companies layoff announcements
+earthquake damages in California
+cybersecurity breach at major corporation
+space exploration mission to Mars
+immigration reform legislation debate
+renewable energy investment trends
+healthcare costs rising across America
+protests against police brutality
+wildfires destroy homes in the West
+Olympic games highlights and records
+celebrity scandal rocks Hollywood
+breakthrough cancer treatment discovered
+housing market bubble concerns
+federal reserve interest rate decision
+school shooting tragedy response
+diplomatic tensions between superpowers
+drone strike kills terrorist leader
+social media platform faces regulation
+archaeological discovery reveals ancient civilization
+unemployment rate hits record low
+autonomous vehicles testing expansion
+streaming service launches original content
+opioid crisis intervention programs
+trade war tariffs impact economy
+infrastructure bill passes Congress
+data privacy concerns grow
+minimum wage increase proposal
+college admissions scandal exposed
+NFL player protest during anthem
+cryptocurrency regulation debate
+pandemic lockdown restrictions eased
+mass shooting gun control debate
+tax reform legislation impact
+ransomware attack cripples pipeline
+climate activists stage demonstration
+sports team wins championship
+banking system collapse fears
+pharmaceutical company fraud charges
+genetic engineering ethical concerns
+border wall funding controversy
+impeachment proceedings begin
+nuclear weapons treaty violation
+artificial meat alternative launch
+student loan debt forgiveness
+venture capital funding decline
+facial recognition ban proposed
+election interference investigation
+pandemic preparedness failures
+police reform measures announced
+wildfire prevention strategies
+ocean pollution crisis worsens
+manufacturing jobs returning
+pension fund shortfall concerns
+antitrust investigation launched
+voting rights protection act
+mental health awareness campaign
+homeless population increasing
+space debris collision risk
+drug cartel violence escalates
+renewable energy jobs growth
+infrastructure deterioration report
+vaccine mandate legal challenge
+cryptocurrency market volatility
+autonomous drone delivery service
+deep fake technology dangers
+Arctic ice melting accelerates
+income inequality gap widens
+election fraud claims disputed
+corporate merger blocked
+medical breakthrough extends life
+transportation strike disrupts city
+racial justice protests spread
+carbon emissions reduction goals
+financial crisis warning signs
+cyberbullying prevention efforts
+asteroid near miss with Earth
+gene therapy approval granted
+labor union organizing drive
+surveillance technology expansion
+education funding cuts proposed
+disaster relief efforts underway
+housing affordability crisis
+clean water access shortage
+artificial intelligence job displacement
+trade agreement negotiations
+prison reform initiative launched
+species extinction accelerates
+political corruption scandal
+terrorism threat level raised
+food safety contamination outbreak
+ai model release
+affordability interest rates
+peanut allergies in newbons
+breaking bad walter white

benchmarks-ann/faiss_kmeans.py

@@ -0,0 +1,101 @@
+#!/usr/bin/env python3
+"""Compute k-means centroids using FAISS and save to a centroids DB.
+
+Reads the first N vectors from a base.db, runs FAISS k-means, and writes
+the centroids to an output SQLite DB as float32 blobs.
+
+Usage:
+  python faiss_kmeans.py --base-db datasets/cohere10m/base.db --ntrain 100000 \
+    --nclusters 8192 -o centroids.db
+
+Output schema:
+  CREATE TABLE centroids (
+    centroid_id INTEGER PRIMARY KEY,
+    centroid BLOB NOT NULL          -- float32[D]
+  );
+  CREATE TABLE meta (key TEXT PRIMARY KEY, value TEXT);
+    -- ntrain, nclusters, dimensions, elapsed_s
+"""
+import argparse
+import os
+import sqlite3
+import struct
+import time
+
+import faiss
+import numpy as np
+
+
+def main():
+    parser = argparse.ArgumentParser(description="FAISS k-means centroid computation")
+    parser.add_argument("--base-db", required=True, help="path to base.db with train table")
+    parser.add_argument("--ntrain", type=int, required=True, help="number of vectors to train on")
+    parser.add_argument("--nclusters", type=int, required=True, help="number of clusters (nlist)")
+    parser.add_argument("--niter", type=int, default=20, help="k-means iterations (default 20)")
+    parser.add_argument("--seed", type=int, default=42, help="random seed")
+    parser.add_argument("-o", "--output", required=True, help="output centroids DB path")
+    args = parser.parse_args()
+
+    # Load vectors
+    print(f"Loading {args.ntrain} vectors from {args.base_db}...")
+    conn = sqlite3.connect(args.base_db)
+    rows = conn.execute(
+        "SELECT vector FROM train ORDER BY id LIMIT ?", (args.ntrain,)
+    ).fetchall()
+    conn.close()
+
+    # Parse float32 blobs to numpy
+    first_blob = rows[0][0]
+    D = len(first_blob) // 4  # float32
+    print(f"  Dimensions: {D}, loaded {len(rows)} vectors")
+
+    vectors = np.zeros((len(rows), D), dtype=np.float32)
+    for i, (blob,) in enumerate(rows):
+        vectors[i] = np.frombuffer(blob, dtype=np.float32)
+
+    # Normalize for cosine distance (FAISS k-means on L2 of unit vectors ≈ cosine)
+    norms = np.linalg.norm(vectors, axis=1, keepdims=True)
+    norms[norms == 0] = 1
+    vectors /= norms
+
+    # Run FAISS k-means
+    print(f"Running k-means: {args.nclusters} clusters, {args.niter} iterations...")
+    t0 = time.perf_counter()
+    kmeans = faiss.Kmeans(
+        D, args.nclusters,
+        niter=args.niter,
+        seed=args.seed,
+        verbose=True,
+        gpu=False,
+    )
+    kmeans.train(vectors)
+    elapsed = time.perf_counter() - t0
+    print(f"  Done in {elapsed:.1f}s")
+
+    centroids = kmeans.centroids  # (nclusters, D) float32
+
+    # Write output DB
+    if os.path.exists(args.output):
+        os.remove(args.output)
+    out = sqlite3.connect(args.output)
+    out.execute("CREATE TABLE centroids (centroid_id INTEGER PRIMARY KEY, centroid BLOB NOT NULL)")
+    out.execute("CREATE TABLE meta (key TEXT PRIMARY KEY, value TEXT)")
+
+    for i in range(args.nclusters):
+        blob = centroids[i].tobytes()
+        out.execute("INSERT INTO centroids (centroid_id, centroid) VALUES (?, ?)", (i, blob))
+
+    out.execute("INSERT INTO meta VALUES ('ntrain', ?)", (str(args.ntrain),))
+    out.execute("INSERT INTO meta VALUES ('nclusters', ?)", (str(args.nclusters),))
+    out.execute("INSERT INTO meta VALUES ('dimensions', ?)", (str(D),))
+    out.execute("INSERT INTO meta VALUES ('niter', ?)", (str(args.niter),))
+    out.execute("INSERT INTO meta VALUES ('elapsed_s', ?)", (str(round(elapsed, 3)),))
+    out.execute("INSERT INTO meta VALUES ('seed', ?)", (str(args.seed),))
+    out.commit()
+    out.close()
+
+    print(f"Wrote {args.nclusters} centroids to {args.output}")
+
+
+if __name__ == "__main__":
+    main()

benchmarks-ann/ground_truth.py

@@ -0,0 +1,168 @@
+#!/usr/bin/env python3
+"""Compute per-subset ground truth for ANN benchmarks.
+
+For subset sizes < 1M, builds a temporary vec0 float table with the first N
+vectors and runs brute-force KNN to get correct ground truth per subset.
+
+For 1M (the full dataset), converts the existing `neighbors` table.
+
+Output: ground_truth.{subset_size}.db with table:
+  ground_truth(query_vector_id, rank, neighbor_id, distance)
+
+Usage:
+  python ground_truth.py --subset-size 50000
+  python ground_truth.py --subset-size 1000000
+"""
+import argparse
+import os
+import sqlite3
+import time
+
+_SCRIPT_DIR = os.path.dirname(os.path.abspath(__file__))
+EXT_PATH = os.path.join(_SCRIPT_DIR, "..", "dist", "vec0")
+BASE_DB = os.path.join(_SCRIPT_DIR, "seed", "base.db")
+FULL_DATASET_SIZE = 1_000_000
+
+
+def gen_ground_truth_subset(base_db, ext_path, subset_size, n_queries, k, out_path):
+    """Build ground truth by brute-force KNN over the first `subset_size` vectors."""
+    if os.path.exists(out_path):
+        os.remove(out_path)
+
+    conn = sqlite3.connect(out_path)
+    conn.enable_load_extension(True)
+    conn.load_extension(ext_path)
+
+    conn.execute(
+        "CREATE TABLE ground_truth ("
+        "  query_vector_id INTEGER NOT NULL,"
+        "  rank INTEGER NOT NULL,"
+        "  neighbor_id INTEGER NOT NULL,"
+        "  distance REAL NOT NULL,"
+        "  PRIMARY KEY (query_vector_id, rank)"
+        ")"
+    )
+
+    conn.execute(f"ATTACH DATABASE '{base_db}' AS base")
+
+    print(f"  Building temp vec0 table with {subset_size} vectors...")
+    conn.execute(
+        "CREATE VIRTUAL TABLE tmp_vec USING vec0("
+        "  id integer primary key,"
+        "  embedding float[768] distance_metric=cosine"
+        ")"
+    )
+
+    t0 = time.perf_counter()
+    conn.execute(
+        "INSERT INTO tmp_vec(id, embedding) "
+        "SELECT id, vector FROM base.train WHERE id < :n",
+        {"n": subset_size},
+    )
+    conn.commit()
+    build_time = time.perf_counter() - t0
+    print(f"  Temp table built in {build_time:.1f}s")
+
+    query_vectors = conn.execute(
+        "SELECT id, vector FROM base.query_vectors ORDER BY id LIMIT :n",
+        {"n": n_queries},
+    ).fetchall()
+
+    print(f"  Running brute-force KNN for {len(query_vectors)} queries, k={k}...")
+    t0 = time.perf_counter()
+
+    for i, (qid, qvec) in enumerate(query_vectors):
+        results = conn.execute(
+            "SELECT id, distance FROM tmp_vec "
+            "WHERE embedding MATCH :query AND k = :k",
+            {"query": qvec, "k": k},
+        ).fetchall()
+
+        for rank, (nid, dist) in enumerate(results):
+            conn.execute(
+                "INSERT INTO ground_truth(query_vector_id, rank, neighbor_id, distance) "
+                "VALUES (?, ?, ?, ?)",
+                (qid, rank, nid, dist),
+            )
+
+        if (i + 1) % 10 == 0 or i == 0:
+            elapsed = time.perf_counter() - t0
+            eta = (elapsed / (i + 1)) * (len(query_vectors) - i - 1)
+            print(
+                f"    {i+1}/{len(query_vectors)} queries  "
+                f"elapsed={elapsed:.1f}s  eta={eta:.1f}s",
+                flush=True,
+            )
+
+    conn.commit()
+    conn.execute("DROP TABLE tmp_vec")
+    conn.execute("DETACH DATABASE base")
+    conn.commit()
+
+    elapsed = time.perf_counter() - t0
+    total_rows = conn.execute("SELECT count(*) FROM ground_truth").fetchone()[0]
+    conn.close()
+    print(f"  Ground truth: {total_rows} rows in {elapsed:.1f}s -> {out_path}")
+
+
+def gen_ground_truth_full(base_db, n_queries, k, out_path):
+    """Convert the existing neighbors table for the full 1M dataset."""
+    if os.path.exists(out_path):
+        os.remove(out_path)
+
+    conn = sqlite3.connect(out_path)
+    conn.execute(f"ATTACH DATABASE '{base_db}' AS base")
+
+    conn.execute(
+        "CREATE TABLE ground_truth ("
+        "  query_vector_id INTEGER NOT NULL,"
+        "  rank INTEGER NOT NULL,"
+        "  neighbor_id INTEGER NOT NULL,"
+        "  distance REAL,"
+        "  PRIMARY KEY (query_vector_id, rank)"
+        ")"
+    )
+
+    conn.execute(
+        "INSERT INTO ground_truth(query_vector_id, rank, neighbor_id) "
+        "SELECT query_vector_id, rank, CAST(neighbors_id AS INTEGER) "
+        "FROM base.neighbors "
+        "WHERE query_vector_id < :n AND rank < :k",
+        {"n": n_queries, "k": k},
+    )
+    conn.commit()
+
+    total_rows = conn.execute("SELECT count(*) FROM ground_truth").fetchone()[0]
+    conn.execute("DETACH DATABASE base")
+    conn.close()
+    print(f"  Ground truth (full): {total_rows} rows -> {out_path}")
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Generate per-subset ground truth")
+    parser.add_argument(
+        "--subset-size", type=int, required=True, help="number of vectors in subset"
+    )
+    parser.add_argument("-n", type=int, default=100, help="number of query vectors")
+    parser.add_argument("-k", type=int, default=100, help="max k for ground truth")
+    parser.add_argument("--base-db", default=BASE_DB)
+    parser.add_argument("--ext", default=EXT_PATH)
+    parser.add_argument(
+        "-o", "--out-dir", default=os.path.join(_SCRIPT_DIR, "seed"),
+        help="output directory for ground_truth.{N}.db",
+    )
+    args = parser.parse_args()
+
+    os.makedirs(args.out_dir, exist_ok=True)
+    out_path = os.path.join(args.out_dir, f"ground_truth.{args.subset_size}.db")
+
+    if args.subset_size >= FULL_DATASET_SIZE:
+        gen_ground_truth_full(args.base_db, args.n, args.k, out_path)
+    else:
+        gen_ground_truth_subset(
+            args.base_db, args.ext, args.subset_size, args.n, args.k, out_path
+        )
+
+
+if __name__ == "__main__":
+    main()

benchmarks-ann/profile.py

@@ -0,0 +1,440 @@
+#!/usr/bin/env python3
+"""CPU profiling for sqlite-vec KNN configurations using macOS `sample` tool.
+
+Builds dist/sqlite3 (with -g3), generates a SQL workload (inserts + repeated
+KNN queries) for each config, profiles the sqlite3 process with `sample`, and
+prints the top-N hottest functions by self (exclusive) CPU samples.
+
+Usage:
+  cd benchmarks-ann
+  uv run profile.py --subset-size 50000 -n 50 \\
+    "baseline-int8:type=baseline,variant=int8,oversample=8" \\
+    "rescore-int8:type=rescore,quantizer=int8,oversample=8"
+"""
+
+import argparse
+import os
+import re
+import shutil
+import subprocess
+import sys
+import tempfile
+
+_SCRIPT_DIR = os.path.dirname(os.path.abspath(__file__))
+_PROJECT_ROOT = os.path.join(_SCRIPT_DIR, "..")
+
+sys.path.insert(0, _SCRIPT_DIR)
+from bench import (
+    BASE_DB,
+    DEFAULT_INSERT_SQL,
+    INDEX_REGISTRY,
+    INSERT_BATCH_SIZE,
+    parse_config,
+)
+
+SQLITE3_PATH = os.path.join(_PROJECT_ROOT, "dist", "sqlite3")
+EXT_PATH = os.path.join(_PROJECT_ROOT, "dist", "vec0")
+
+
+# ============================================================================
+# SQL generation
+# ============================================================================
+
+
+def _query_sql_for_config(params, query_id, k):
+    """Return a SQL query string for a single KNN query by query_vector id."""
+    index_type = params["index_type"]
+    qvec = f"(SELECT vector FROM base.query_vectors WHERE id = {query_id})"
+
+    if index_type == "baseline":
+        variant = params.get("variant", "float")
+        oversample = params.get("oversample", 8)
+        oversample_k = k * oversample
+
+        if variant == "int8":
+            return (
+                f"WITH coarse AS ("
+                f"  SELECT id, embedding FROM vec_items"
+                f"  WHERE embedding_int8 MATCH vec_quantize_int8({qvec}, 'unit')"
+                f"  LIMIT {oversample_k}"
+                f") "
+                f"SELECT id, vec_distance_cosine(embedding, {qvec}) as distance "
+                f"FROM coarse ORDER BY 2 LIMIT {k};"
+            )
+        elif variant == "bit":
+            return (
+                f"WITH coarse AS ("
+                f"  SELECT id, embedding FROM vec_items"
+                f"  WHERE embedding_bq MATCH vec_quantize_binary({qvec})"
+                f"  LIMIT {oversample_k}"
+                f") "
+                f"SELECT id, vec_distance_cosine(embedding, {qvec}) as distance "
+                f"FROM coarse ORDER BY 2 LIMIT {k};"
+            )
+
+    # Default MATCH query (baseline-float, rescore, and others)
+    return (
+        f"SELECT id, distance FROM vec_items"
+        f" WHERE embedding MATCH {qvec} AND k = {k};"
+    )
+
+
+def generate_sql(db_path, params, subset_size, n_queries, k, repeats):
+    """Generate a complete SQL workload: load ext, create table, insert, query."""
+    lines = []
+    lines.append(".bail on")
+    lines.append(f".load {EXT_PATH}")
+    lines.append(f"ATTACH DATABASE '{os.path.abspath(BASE_DB)}' AS base;")
+    lines.append("PRAGMA page_size=8192;")
+
+    # Create table
+    reg = INDEX_REGISTRY[params["index_type"]]
+    lines.append(reg["create_table_sql"](params) + ";")
+
+    # Inserts
+    sql_fn = reg.get("insert_sql")
+    insert_sql = sql_fn(params) if sql_fn else None
+    if insert_sql is None:
+        insert_sql = DEFAULT_INSERT_SQL
+    for lo in range(0, subset_size, INSERT_BATCH_SIZE):
+        hi = min(lo + INSERT_BATCH_SIZE, subset_size)
+        stmt = insert_sql.replace(":lo", str(lo)).replace(":hi", str(hi))
+        lines.append(stmt + ";")
+        if hi % 10000 == 0 or hi == subset_size:
+            lines.append("-- progress: inserted %d/%d" % (hi, subset_size))
+
+    # Queries (repeated)
+    lines.append("-- BEGIN QUERIES")
+    for _rep in range(repeats):
+        for qid in range(n_queries):
+            lines.append(_query_sql_for_config(params, qid, k))
+
+    return "\n".join(lines)
+
+
+# ============================================================================
+# Profiling with macOS `sample`
+# ============================================================================
+
+
+def run_profile(sqlite3_path, db_path, sql_file, sample_output, duration=120):
+    """Run sqlite3 under macOS `sample` profiler.
+
+    Starts sqlite3 directly with stdin from the SQL file, then immediately
+    attaches `sample` to its PID with -mayDie (tolerates process exit).
+    The workload must be long enough for sample to attach and capture useful data.
+    """
+    sql_fd = open(sql_file, "r")
+    proc = subprocess.Popen(
+        [sqlite3_path, db_path],
+        stdin=sql_fd,
+        stdout=subprocess.DEVNULL,
+        stderr=subprocess.PIPE,
+    )
+
+    pid = proc.pid
+    print(f"    sqlite3 PID: {pid}")
+
+    # Attach sample immediately (1ms interval, -mayDie tolerates process exit)
+    sample_proc = subprocess.Popen(
+        ["sample", str(pid), str(duration), "1", "-mayDie", "-file", sample_output],
+        stdout=subprocess.DEVNULL,
+        stderr=subprocess.PIPE,
+    )
+
+    # Wait for sqlite3 to finish
+    _, stderr = proc.communicate()
+    sql_fd.close()
+    rc = proc.returncode
+    if rc != 0:
+        print(f"    sqlite3 failed (rc={rc}):", file=sys.stderr)
+        print(f"    {stderr.decode().strip()}", file=sys.stderr)
+        sample_proc.kill()
+        return False
+
+    # Wait for sample to finish
+    sample_proc.wait()
+    return True
+
+
+# ============================================================================
+# Parse `sample` output
+# ============================================================================
+
+# Tree-drawing characters used by macOS `sample` to represent hierarchy.
+# We replace them with spaces so indentation depth reflects tree depth.
+_TREE_CHARS_RE = re.compile(r"[+!:|]")
+
+# After tree chars are replaced with spaces, each call-graph line looks like:
+#   "          800 rescore_knn  (in vec0.dylib) + 3808,3640,...  [0x1a,0x2b,...]  file.c:123"
+# We extract just (indent, count, symbol, module) — everything after "(in ...)"
+# is decoration we don't need.
+_LEADING_RE = re.compile(r"^(\s+)(\d+)\s+(.+)")
+
+
+def _extract_symbol_and_module(rest):
+    """Given the text after 'count ', extract (symbol, module).
+
+    Handles patterns like:
+      'rescore_knn  (in vec0.dylib) + 3808,3640,...  [0x...]'
+      'pread  (in libsystem_kernel.dylib) + 8  [0x...]'
+      '???  (in <unknown binary>)  [0x...]'
+      'start  (in dyld) + 2840  [0x198650274]'
+      'Thread_26759239   DispatchQueue_1: ...'
+    """
+    # Try to find "(in ...)" to split symbol from module
+    m = re.match(r"^(.+?)\s+\(in\s+(.+?)\)", rest)
+    if m:
+        return m.group(1).strip(), m.group(2).strip()
+    # No module — return whole thing as symbol, strip trailing junk
+    sym = re.sub(r"\s+\[0x[0-9a-f].*", "", rest).strip()
+    return sym, ""
+
+
+def _parse_call_graph_lines(text):
+    """Parse call-graph section into list of (depth, count, symbol, module)."""
+    entries = []
+    for raw_line in text.split("\n"):
+        # Strip tree-drawing characters, replace with spaces to preserve depth
+        line = _TREE_CHARS_RE.sub(" ", raw_line)
+        m = _LEADING_RE.match(line)
+        if not m:
+            continue
+        depth = len(m.group(1))
+        count = int(m.group(2))
+        rest = m.group(3)
+        symbol, module = _extract_symbol_and_module(rest)
+        entries.append((depth, count, symbol, module))
+    return entries
+
+
+def parse_sample_output(filepath):
+    """Parse `sample` call-graph output, compute exclusive (self) samples per function.
+
+    Returns dict of {display_name: self_sample_count}.
+    """
+    with open(filepath, "r") as f:
+        text = f.read()
+
+    # Find "Call graph:" section
+    cg_start = text.find("Call graph:")
+    if cg_start == -1:
+        print("    Warning: no 'Call graph:' section found in sample output")
+        return {}
+
+    # End at "Total number in stack" or EOF
+    cg_end = text.find("\nTotal number in stack", cg_start)
+    if cg_end == -1:
+        cg_end = len(text)
+
+    entries = _parse_call_graph_lines(text[cg_start:cg_end])
+
+    if not entries:
+        print("    Warning: no call graph entries parsed")
+        return {}
+
+    # Compute self (exclusive) samples per function:
+    #   self = count - sum(direct_children_counts)
+    self_samples = {}
+    for i, (depth, count, sym, mod) in enumerate(entries):
+        children_sum = 0
+        child_depth = None
+        for j in range(i + 1, len(entries)):
+            j_depth = entries[j][0]
+            if j_depth <= depth:
+                break
+            if child_depth is None:
+                child_depth = j_depth
+            if j_depth == child_depth:
+                children_sum += entries[j][1]
+
+        self_count = count - children_sum
+        if self_count > 0:
+            key = f"{sym}  ({mod})" if mod else sym
+            self_samples[key] = self_samples.get(key, 0) + self_count
+
+    return self_samples
+
+
+# ============================================================================
+# Display
+# ============================================================================
+
+
+def print_profile(title, self_samples, top_n=20):
+    total = sum(self_samples.values())
+    if total == 0:
+        print(f"\n=== {title} (no samples) ===")
+        return
+
+    sorted_syms = sorted(self_samples.items(), key=lambda x: -x[1])
+
+    print(f"\n=== {title} (top {top_n}, {total} total self-samples) ===")
+    for sym, count in sorted_syms[:top_n]:
+        pct = 100.0 * count / total
+        print(f"  {pct:5.1f}%  {count:>6}  {sym}")
+
+
+# ============================================================================
+# Main
+# ============================================================================
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="CPU profiling for sqlite-vec KNN configurations",
+        formatter_class=argparse.RawDescriptionHelpFormatter,
+        epilog=__doc__,
+    )
+    parser.add_argument(
+        "configs", nargs="+", help="config specs (name:type=X,key=val,...)"
+    )
+    parser.add_argument("--subset-size", type=int, required=True)
+    parser.add_argument("-k", type=int, default=10, help="KNN k (default 10)")
+    parser.add_argument(
+        "-n", type=int, default=50, help="number of distinct queries (default 50)"
+    )
+    parser.add_argument(
+        "--repeats",
+        type=int,
+        default=10,
+        help="repeat query set N times for more samples (default 10)",
+    )
+    parser.add_argument(
+        "--top", type=int, default=20, help="show top N functions (default 20)"
+    )
+    parser.add_argument("--base-db", default=BASE_DB)
+    parser.add_argument("--sqlite3", default=SQLITE3_PATH)
+    parser.add_argument(
+        "--keep-temp",
+        action="store_true",
+        help="keep temp directory with DBs, SQL, and sample output",
+    )
+    args = parser.parse_args()
+
+    # Check prerequisites
+    if not os.path.exists(args.base_db):
+        print(f"Error: base DB not found at {args.base_db}", file=sys.stderr)
+        print("Run 'make seed' in benchmarks-ann/ first.", file=sys.stderr)
+        sys.exit(1)
+
+    if not shutil.which("sample"):
+        print("Error: macOS 'sample' tool not found.", file=sys.stderr)
+        sys.exit(1)
+
+    # Build CLI
+    print("Building dist/sqlite3...")
+    result = subprocess.run(
+        ["make", "cli"], cwd=_PROJECT_ROOT, capture_output=True, text=True
+    )
+    if result.returncode != 0:
+        print(f"Error: make cli failed:\n{result.stderr}", file=sys.stderr)
+        sys.exit(1)
+    print("  done.")
+
+    if not os.path.exists(args.sqlite3):
+        print(f"Error: sqlite3 not found at {args.sqlite3}", file=sys.stderr)
+        sys.exit(1)
+
+    configs = [parse_config(c) for c in args.configs]
+
+    tmpdir = tempfile.mkdtemp(prefix="sqlite-vec-profile-")
+    print(f"Working directory: {tmpdir}")
+
+    all_profiles = []
+
+    for i, (name, params) in enumerate(configs, 1):
+        reg = INDEX_REGISTRY[params["index_type"]]
+        desc = reg["describe"](params)
+        print(f"\n[{i}/{len(configs)}] {name}  ({desc})")
+
+        # Generate SQL workload
+        db_path = os.path.join(tmpdir, f"{name}.db")
+        sql_text = generate_sql(
+            db_path, params, args.subset_size, args.n, args.k, args.repeats
+        )
+        sql_file = os.path.join(tmpdir, f"{name}.sql")
+        with open(sql_file, "w") as f:
+            f.write(sql_text)
+
+        total_queries = args.n * args.repeats
+        print(
+            f"  SQL workload: {args.subset_size} inserts + "
+            f"{total_queries} queries ({args.n} x {args.repeats} repeats)"
+        )
+
+        # Profile
+        sample_file = os.path.join(tmpdir, f"{name}.sample.txt")
+        print(f"  Profiling...")
+        ok = run_profile(args.sqlite3, db_path, sql_file, sample_file)
+        if not ok:
+            print(f"  FAILED — skipping {name}")
+            all_profiles.append((name, desc, {}))
+            continue
+
+        if not os.path.exists(sample_file):
+            print(f"  Warning: sample output not created")
+            all_profiles.append((name, desc, {}))
+            continue
+
+        # Parse
+        self_samples = parse_sample_output(sample_file)
+        all_profiles.append((name, desc, self_samples))
+
+        # Show individual profile
+        print_profile(f"{name} ({desc})", self_samples, args.top)
+
+    # Side-by-side comparison if multiple configs
+    if len(all_profiles) > 1:
+        print("\n" + "=" * 80)
+        print("COMPARISON")
+        print("=" * 80)
+
+        # Collect all symbols that appear in top-N of any config
+        all_syms = set()
+        for _name, _desc, prof in all_profiles:
+            sorted_syms = sorted(prof.items(), key=lambda x: -x[1])
+            for sym, _count in sorted_syms[: args.top]:
+                all_syms.add(sym)
+
+        # Build comparison table
+        rows = []
+        for sym in all_syms:
+            row = [sym]
+            for _name, _desc, prof in all_profiles:
+                total = sum(prof.values())
+                count = prof.get(sym, 0)
+                pct = 100.0 * count / total if total > 0 else 0.0
+                row.append((pct, count))
+            max_pct = max(r[0] for r in row[1:])
+            rows.append((max_pct, row))
+
+        rows.sort(key=lambda x: -x[0])
+
+        # Header
+        header = f"{'function':>40}"
+        for name, desc, _ in all_profiles:
+            header += f"  {name:>14}"
+        print(header)
+        print("-" * len(header))
+
+        for _sort_key, row in rows[: args.top * 2]:
+            sym = row[0]
+            display_sym = sym if len(sym) <= 40 else sym[:37] + "..."
+            line = f"{display_sym:>40}"
+            for pct, count in row[1:]:
+                if count > 0:
+                    line += f"  {pct:>13.1f}%"
+                else:
+                    line += f"  {'-':>14}"
+            print(line)
+
+    if args.keep_temp:
+        print(f"\nTemp files kept at: {tmpdir}")
+    else:
+        shutil.rmtree(tmpdir)
+        print(f"\nTemp files cleaned up. Use --keep-temp to preserve.")
+
+
+if __name__ == "__main__":
+    main()

benchmarks-ann/results_schema.sql

@@ -0,0 +1,76 @@
+-- Comprehensive results schema for vec0 KNN benchmark runs.
+-- Created in WAL mode: PRAGMA journal_mode=WAL
+
+CREATE TABLE IF NOT EXISTS runs (
+  run_id          INTEGER PRIMARY KEY AUTOINCREMENT,
+  config_name     TEXT    NOT NULL,
+  index_type      TEXT    NOT NULL,
+  params          TEXT    NOT NULL,   -- JSON: {"R":48,"L":128,"quantizer":"binary"}
+  dataset         TEXT    NOT NULL,   -- "cohere1m"
+  subset_size     INTEGER NOT NULL,
+  k               INTEGER NOT NULL,
+  n_queries       INTEGER NOT NULL,
+  phase           TEXT    NOT NULL DEFAULT 'both',
+    -- 'build', 'query', or 'both'
+  status          TEXT    NOT NULL DEFAULT 'pending',
+    -- pending → inserting → training → querying → done | built | error
+  created_at_ns   INTEGER NOT NULL    -- time.time_ns()
+);
+
+CREATE TABLE IF NOT EXISTS run_results (
+  run_id              INTEGER PRIMARY KEY REFERENCES runs(run_id),
+  insert_started_ns   INTEGER,
+  insert_ended_ns     INTEGER,
+  insert_duration_ns  INTEGER,
+  train_started_ns    INTEGER,       -- NULL if no training
+  train_ended_ns      INTEGER,
+  train_duration_ns   INTEGER,
+  build_duration_ns   INTEGER,       -- insert + train
+  db_file_size_bytes  INTEGER,
+  db_file_path        TEXT,
+  create_sql          TEXT,           -- CREATE VIRTUAL TABLE ...
+  insert_sql          TEXT,           -- INSERT INTO vec_items ...
+  train_sql           TEXT,           -- NULL if no training step
+  query_sql           TEXT,           -- SELECT ... WHERE embedding MATCH ...
+  k                   INTEGER,       -- denormalized from runs for easy filtering
+  query_mean_ms       REAL,          -- denormalized aggregates
+  query_median_ms     REAL,
+  query_p99_ms        REAL,
+  query_total_ms      REAL,
+  qps                 REAL,
+  recall              REAL
+);
+
+CREATE TABLE IF NOT EXISTS insert_batches (
+  batch_id        INTEGER PRIMARY KEY AUTOINCREMENT,
+  run_id          INTEGER NOT NULL REFERENCES runs(run_id),
+  batch_lo        INTEGER NOT NULL,   -- start index (inclusive)
+  batch_hi        INTEGER NOT NULL,   -- end index (exclusive)
+  rows_in_batch   INTEGER NOT NULL,
+  started_ns      INTEGER NOT NULL,
+  ended_ns        INTEGER NOT NULL,
+  duration_ns     INTEGER NOT NULL,
+  cumulative_rows INTEGER NOT NULL,   -- total rows inserted so far
+  rate_rows_per_s REAL    NOT NULL    -- cumulative rate
+);
+
+CREATE TABLE IF NOT EXISTS queries (
+  query_id          INTEGER PRIMARY KEY AUTOINCREMENT,
+  run_id            INTEGER NOT NULL REFERENCES runs(run_id),
+  k                 INTEGER NOT NULL,
+  query_vector_id   INTEGER NOT NULL,
+  started_ns        INTEGER NOT NULL,
+  ended_ns          INTEGER NOT NULL,
+  duration_ms       REAL    NOT NULL,
+  result_ids        TEXT    NOT NULL,  -- JSON array
+  result_distances  TEXT    NOT NULL,  -- JSON array
+  ground_truth_ids  TEXT    NOT NULL,  -- JSON array
+  recall            REAL    NOT NULL,
+  UNIQUE(run_id, k, query_vector_id)
+);
+
+CREATE INDEX IF NOT EXISTS idx_runs_config  ON runs(config_name);
+CREATE INDEX IF NOT EXISTS idx_runs_type    ON runs(index_type);
+CREATE INDEX IF NOT EXISTS idx_runs_status  ON runs(status);
+CREATE INDEX IF NOT EXISTS idx_batches_run  ON insert_batches(run_id);
+CREATE INDEX IF NOT EXISTS idx_queries_run  ON queries(run_id);

benchmarks-ann/schema.sql

@@ -0,0 +1,60 @@
+-- Canonical results schema for vec0 KNN benchmark comparisons.
+-- The index_type column is a free-form TEXT field. Baseline configs use
+-- "baseline"; index-specific branches add their own types (registered
+-- via INDEX_REGISTRY in bench.py).
+
+CREATE TABLE IF NOT EXISTS runs (
+  run_id       INTEGER PRIMARY KEY AUTOINCREMENT,
+  config_name  TEXT NOT NULL,
+  index_type   TEXT NOT NULL,
+  subset_size  INTEGER NOT NULL,
+  phase        TEXT NOT NULL DEFAULT 'both',  -- 'build', 'query', or 'both'
+  status       TEXT NOT NULL DEFAULT 'pending',
+  k            INTEGER,
+  n            INTEGER,
+  db_path      TEXT,
+  insert_time_s REAL,
+  train_time_s REAL,
+  total_build_time_s REAL,
+  rows         INTEGER,
+  file_size_mb REAL,
+  mean_ms      REAL,
+  median_ms    REAL,
+  p99_ms       REAL,
+  total_query_ms REAL,
+  qps          REAL,
+  recall       REAL,
+  created_at   TEXT NOT NULL DEFAULT (datetime('now')),
+  finished_at  TEXT
+);
+
+CREATE TABLE IF NOT EXISTS build_results (
+  config_name  TEXT NOT NULL,
+  index_type   TEXT NOT NULL,
+  subset_size  INTEGER NOT NULL,
+  db_path      TEXT NOT NULL,
+  insert_time_s REAL NOT NULL,
+  train_time_s REAL,            -- NULL when no training/build step is needed
+  total_time_s REAL NOT NULL,
+  rows         INTEGER NOT NULL,
+  file_size_mb REAL NOT NULL,
+  created_at   TEXT NOT NULL DEFAULT (datetime('now')),
+  PRIMARY KEY (config_name, subset_size)
+);
+
+CREATE TABLE IF NOT EXISTS bench_results (
+  config_name  TEXT NOT NULL,
+  index_type   TEXT NOT NULL,
+  subset_size  INTEGER NOT NULL,
+  k            INTEGER NOT NULL,
+  n            INTEGER NOT NULL,
+  mean_ms      REAL NOT NULL,
+  median_ms    REAL NOT NULL,
+  p99_ms       REAL NOT NULL,
+  total_ms     REAL NOT NULL,
+  qps          REAL NOT NULL,
+  recall       REAL NOT NULL,
+  db_path      TEXT NOT NULL,
+  created_at   TEXT NOT NULL DEFAULT (datetime('now')),
+  PRIMARY KEY (config_name, subset_size, k)
+);

benchmarks/exhaustive-memory/bench.py

@@ -248,59 +248,6 @@ def bench_libsql(base, query, page_size, k) -> BenchResult:
     return BenchResult(f"libsql ({page_size})", build_time, times)
 
 
-def register_np(db, array, name):
-    ptr = array.__array_interface__["data"][0]
-    nvectors, dimensions = array.__array_interface__["shape"]
-    element_type = array.__array_interface__["typestr"]
-
-    assert element_type == "<f4"
-
-    name_escaped = db.execute("select printf('%w', ?)", [name]).fetchone()[0]
-
-    db.execute(
-        "insert into temp.vec_static_blobs(name, data) select ?, vec_static_blob_from_raw(?, ?, ?, ?)",
-        [name, ptr, element_type, dimensions, nvectors],
-    )
-
-    db.execute(
-        f'create virtual table "{name_escaped}" using vec_static_blob_entries({name_escaped})'
-    )
-
-def bench_sqlite_vec_static(base, query, k) -> BenchResult:
-    print(f"sqlite-vec static...")
-
-    db = sqlite3.connect(":memory:")
-    db.enable_load_extension(True)
-    db.load_extension("../../dist/vec0")
-
-
-
-    t = time.time()
-    register_np(db, base, "base")
-    build_time = time.time() - t
-
-    times = []
-    results = []
-    for (
-        idx,
-        q,
-    ) in enumerate(query):
-        t0 = time.time()
-        result = db.execute(
-            """
-              select
-                rowid
-              from base
-              where vector match ?
-                and k = ?
-              order by distance
-            """,
-            [q.tobytes(), k],
-        ).fetchall()
-        assert len(result) == k
-        times.append(time.time() - t0)
-    return BenchResult(f"sqlite-vec static", build_time, times)
-
 def bench_faiss(base, query, k) -> BenchResult:
     import faiss
     dimensions = base.shape[1]
@@ -438,8 +385,6 @@ def suite(name, base, query, k, benchmarks):
     for b in benchmarks:
         if b == "faiss":
             results.append(bench_faiss(base, query, k=k))
-        elif b == "vec-static":
-          results.append(bench_sqlite_vec_static(base, query, k=k))
         elif b.startswith("vec-scalar"):
             _, page_size = b.split('.')
             results.append(bench_sqlite_vec_scalar(base, query, page_size, k=k))
@@ -541,7 +486,7 @@ def parse_args():
         help="Number of queries to use. Defaults all",
     )
     parser.add_argument(
-        "-x", help="type of runs to make", default="faiss,vec-scalar.4096,vec-static,vec-vec0.4096.16,usearch,duckdb,hnswlib,numpy"
+        "-x", help="type of runs to make", default="faiss,vec-scalar.4096,vec-vec0.4096.16,usearch,duckdb,hnswlib,numpy"
     )
 
     args = parser.parse_args()

benchmarks/profiling/build-from-npy.sql

@@ -8,10 +8,3 @@ create virtual table vec_items using vec0(
   embedding float[1536]
 );
 
--- 65s (limit 1e5), ~615MB on disk
-insert into vec_items
-  select
-    rowid,
-    vector
-  from vec_npy_each(vec_npy_file('examples/dbpedia-openai/data/vectors.npy'))
-  limit 1e5;

benchmarks/self-params/build.py

@@ -6,7 +6,6 @@ def connect(path):
     db = sqlite3.connect(path)
     db.enable_load_extension(True)
     db.load_extension("../dist/vec0")
-    db.execute("select load_extension('../dist/vec0', 'sqlite3_vec_fs_read_init')")
     db.enable_load_extension(False)
     return db
 
@@ -18,8 +17,6 @@ page_sizes = [  # 4096, 8192,
 chunk_sizes = [128, 256, 1024, 2048]
 types = ["f32", "int8", "bit"]
 
-SRC = "../examples/dbpedia-openai/data/vectors.npy"
-
 for page_size in page_sizes:
     for chunk_size in chunk_sizes:
         for t in types:
@@ -42,15 +39,8 @@ for page_size in page_sizes:
                     func = "vec_quantize_i8(vector, 'unit')"
                 if t == "bit":
                     func = "vec_quantize_binary(vector)"
-                db.execute(
+                # TODO: replace with non-npy data loading
-                    f"""
+                pass
-                      insert into vec_items
-                      select rowid, {func}
-                      from vec_npy_each(vec_npy_file(?))
-                      limit 100000
-                    """,
-                    [SRC],
-                )
             elapsed = time.time() - t0
             print(elapsed)
 

bindings/go/ncruces/go-sqlite3.patch

@@ -6,7 +6,6 @@ index ed2aaec..4cc0b0e 100755
  	-Wl,--initial-memory=327680 \
  	-D_HAVE_SQLITE_CONFIG_H \
  	-DSQLITE_CUSTOM_INCLUDE=sqlite_opt.h \
-+  -DSQLITE_VEC_OMIT_FS=1 \
  	$(awk '{print "-Wl,--export="$0}' exports.txt)
 
  "$BINARYEN/wasm-ctor-eval" -g -c _initialize sqlite3.wasm -o sqlite3.tmp

bindings/python/extra_init.py

@@ -1,6 +1,5 @@
 from typing import List
 from struct import pack
-from sqlite3 import Connection
 
 
 def serialize_float32(vector: List[float]) -> bytes:
@@ -13,33 +12,3 @@ def serialize_int8(vector: List[int]) -> bytes:
     return pack("%sb" % len(vector), *vector)
 
 
-try:
-    import numpy.typing as npt
-
-    def register_numpy(db: Connection, name: str, array: npt.NDArray):
-        """ayoo"""
-
-        ptr = array.__array_interface__["data"][0]
-        nvectors, dimensions = array.__array_interface__["shape"]
-        element_type = array.__array_interface__["typestr"]
-
-        assert element_type == "<f4"
-
-        name_escaped = db.execute("select printf('%w', ?)", [name]).fetchone()[0]
-
-        db.execute(
-            """
-              insert into temp.vec_static_blobs(name, data)
-              select ?, vec_static_blob_from_raw(?, ?, ?, ?)
-            """,
-            [name, ptr, element_type, dimensions, nvectors],
-        )
-
-        db.execute(
-            f'create virtual table "{name_escaped}" using vec_static_blob_entries({name_escaped})'
-        )
-
-except ImportError:
-
-    def register_numpy(db: Connection, name: str, array):
-        raise Exception("numpy package is required for register_numpy")

scripts/amalgamate.py

@@ -0,0 +1,119 @@
+#!/usr/bin/env python3
+"""
+Amalgamate sqlite-vec into a single distributable .c file.
+
+Reads the dev sqlite-vec.c and inlines any #include "sqlite-vec-*.c" files,
+stripping LSP-support blocks and per-file include guards.
+
+Usage:
+    python3 scripts/amalgamate.py sqlite-vec.c > dist/sqlite-vec.c
+"""
+
+import re
+import sys
+import os
+
+
+def strip_lsp_block(content):
+    """Remove the LSP-support pattern:
+        #ifndef SQLITE_VEC_H
+        #include "sqlite-vec.c" // ...
+        #endif
+    """
+    pattern = re.compile(
+        r'^\s*#ifndef\s+SQLITE_VEC_H\s*\n'
+        r'\s*#include\s+"sqlite-vec\.c"[^\n]*\n'
+        r'\s*#endif[^\n]*\n',
+        re.MULTILINE,
+    )
+    return pattern.sub('', content)
+
+
+def strip_include_guard(content, guard_macro):
+    """Remove the include guard pair:
+        #ifndef GUARD_MACRO
+        #define GUARD_MACRO
+        ...content...
+        (trailing #endif removed)
+    """
+    # Strip the #ifndef / #define pair at the top
+    header_pattern = re.compile(
+        r'^\s*#ifndef\s+' + re.escape(guard_macro) + r'\s*\n'
+        r'\s*#define\s+' + re.escape(guard_macro) + r'\s*\n',
+        re.MULTILINE,
+    )
+    content = header_pattern.sub('', content, count=1)
+
+    # Strip the trailing #endif (last one in file that closes the guard)
+    # Find the last #endif and remove it
+    lines = content.rstrip('\n').split('\n')
+    for i in range(len(lines) - 1, -1, -1):
+        if re.match(r'^\s*#endif', lines[i]):
+            lines.pop(i)
+            break
+
+    return '\n'.join(lines) + '\n'
+
+
+def detect_include_guard(content):
+    """Detect an include guard macro like SQLITE_VEC_IVF_C."""
+    m = re.match(
+        r'\s*(?:/\*[\s\S]*?\*/\s*)?'  # optional block comment
+        r'#ifndef\s+(SQLITE_VEC_\w+_C)\s*\n'
+        r'#define\s+\1',
+        content,
+    )
+    return m.group(1) if m else None
+
+
+def inline_include(match, base_dir):
+    """Replace an #include "sqlite-vec-*.c" with the file's contents."""
+    filename = match.group(1)
+    filepath = os.path.join(base_dir, filename)
+
+    if not os.path.exists(filepath):
+        print(f"Warning: {filepath} not found, leaving #include in place", file=sys.stderr)
+        return match.group(0)
+
+    with open(filepath, 'r') as f:
+        content = f.read()
+
+    # Strip LSP-support block
+    content = strip_lsp_block(content)
+
+    # Strip include guard if present
+    guard = detect_include_guard(content)
+    if guard:
+        content = strip_include_guard(content, guard)
+
+    separator = '/' * 78
+    header = f'\n{separator}\n// Begin inlined: {filename}\n{separator}\n\n'
+    footer = f'\n{separator}\n// End inlined: {filename}\n{separator}\n'
+
+    return header + content.strip('\n') + footer
+
+
+def amalgamate(input_path):
+    base_dir = os.path.dirname(os.path.abspath(input_path))
+
+    with open(input_path, 'r') as f:
+        content = f.read()
+
+    # Replace #include "sqlite-vec-*.c" with inlined contents
+    include_pattern = re.compile(r'^#include\s+"(sqlite-vec-[^"]+\.c)"\s*$', re.MULTILINE)
+    content = include_pattern.sub(lambda m: inline_include(m, base_dir), content)
+
+    return content
+
+
+def main():
+    if len(sys.argv) != 2:
+        print(f"Usage: {sys.argv[0]} <input-file>", file=sys.stderr)
+        sys.exit(1)
+
+    result = amalgamate(sys.argv[1])
+    sys.stdout.write(result)
+
+
+if __name__ == '__main__':
+    main()

scripts/vendor.sh

@@ -1,7 +1,6 @@
 #!/bin/bash
 mkdir -p vendor
 curl -o sqlite-amalgamation.zip https://www.sqlite.org/2024/sqlite-amalgamation-3450300.zip
-unzip -d
 unzip sqlite-amalgamation.zip
 mv sqlite-amalgamation-3450300/* vendor/
 rmdir sqlite-amalgamation-3450300

site/api-reference.md

@@ -568,65 +568,6 @@ select 'todo';
 -- 'todo'
 
 
-```
-
-## NumPy Utilities {#numpy} 
-
-Functions to read data from or work with [NumPy arrays](https://numpy.org/doc/stable/reference/generated/numpy.array.html).
-
-### `vec_npy_each(vector)` {#vec_npy_each}
-
-xxx
-
-
-```sql
--- db.execute('select quote(?)', [to_npy(np.array([[1.0], [2.0], [3.0]], dtype=np.float32))]).fetchone()
-select
-  rowid,
-  vector,
-  vec_type(vector),
-  vec_to_json(vector)
-from vec_npy_each(
-  X'934E554D5059010076007B276465736372273A20273C6634272C2027666F727472616E5F6F72646572273A2046616C73652C20277368617065273A2028332C2031292C207D202020202020202020202020202020202020202020202020202020202020202020202020202020202020202020202020202020202020202020200A0000803F0000004000004040'
-)
-/*
-┌───────┬─────────────┬──────────────────┬─────────────────────┐
-│ rowid │ vector      │ vec_type(vector) │ vec_to_json(vector) │
-├───────┼─────────────┼──────────────────┼─────────────────────┤
-│ 0     │ X'0000803F' │ 'float32'        │ '[1.000000]'        │
-├───────┼─────────────┼──────────────────┼─────────────────────┤
-│ 1     │ X'00000040' │ 'float32'        │ '[2.000000]'        │
-├───────┼─────────────┼──────────────────┼─────────────────────┤
-│ 2     │ X'00004040' │ 'float32'        │ '[3.000000]'        │
-└───────┴─────────────┴──────────────────┴─────────────────────┘
-
-*/
-
-
--- db.execute('select quote(?)', [to_npy(np.array([[1.0], [2.0], [3.0]], dtype=np.float32))]).fetchone()
-select
-  rowid,
-  vector,
-  vec_type(vector),
-  vec_to_json(vector)
-from vec_npy_each(
-  X'934E554D5059010076007B276465736372273A20273C6634272C2027666F727472616E5F6F72646572273A2046616C73652C20277368617065273A2028332C2031292C207D202020202020202020202020202020202020202020202020202020202020202020202020202020202020202020202020202020202020202020200A0000803F0000004000004040'
-)
-/*
-┌───────┬─────────────┬──────────────────┬─────────────────────┐
-│ rowid │ vector      │ vec_type(vector) │ vec_to_json(vector) │
-├───────┼─────────────┼──────────────────┼─────────────────────┤
-│ 0     │ X'0000803F' │ 'float32'        │ '[1.000000]'        │
-├───────┼─────────────┼──────────────────┼─────────────────────┤
-│ 1     │ X'00000040' │ 'float32'        │ '[2.000000]'        │
-├───────┼─────────────┼──────────────────┼─────────────────────┤
-│ 2     │ X'00004040' │ 'float32'        │ '[3.000000]'        │
-└───────┴─────────────┴──────────────────┴─────────────────────┘
-
-*/
-
-
-
 ```
 
 ## Meta {#meta} 

site/compiling.md

@@ -59,5 +59,4 @@ The current compile-time flags are:
 
 - `SQLITE_VEC_ENABLE_AVX`, enables AVX CPU instructions for some vector search operations
 - `SQLITE_VEC_ENABLE_NEON`, enables NEON CPU instructions for some vector search operations
-- `SQLITE_VEC_OMIT_FS`, removes some obsure SQL functions and features that use the filesystem, meant for some WASM builds where there's no available filesystem
 - `SQLITE_VEC_STATIC`, meant for statically linking `sqlite-vec` 

sqlite-dist.toml

@@ -1,6 +1,6 @@
 [package]
 name = "sqlite-vec"
-license = "MIT OR Apache"
+license = "MIT OR Apache-2.0"
 homepage = "https://alexgarcia.xyz/sqlite-vec"
 repo = "https://github.com/asg017/sqlite-vec"
 description = "A vector search SQLite extension."

sqlite-vec-ivf-kmeans.c

@@ -0,0 +1,214 @@
+/**
+ * sqlite-vec-ivf-kmeans.c — Pure k-means clustering algorithm.
+ *
+ * No SQLite dependency. Operates on float arrays in memory.
+ * #include'd into sqlite-vec.c after struct definitions.
+ */
+
+#ifndef SQLITE_VEC_IVF_KMEANS_C
+#define SQLITE_VEC_IVF_KMEANS_C
+
+// When opened standalone in an editor, pull in types so the LSP is happy.
+// When #include'd from sqlite-vec.c, SQLITE_VEC_H is already defined.
+#ifndef SQLITE_VEC_H
+#include "sqlite-vec.c" // IWYU pragma: keep
+#endif
+
+#include <float.h>
+#include <string.h>
+
+#define VEC0_IVF_KMEANS_MAX_ITER     25
+#define VEC0_IVF_KMEANS_DEFAULT_SEED  0
+
+// Simple xorshift32 PRNG
+static uint32_t ivf_xorshift32(uint32_t *state) {
+  uint32_t x = *state;
+  x ^= x << 13;
+  x ^= x >> 17;
+  x ^= x << 5;
+  *state = x;
+  return x;
+}
+
+// L2 squared distance between two float vectors
+static float ivf_l2_dist(const float *a, const float *b, int D) {
+  float sum = 0.0f;
+  for (int d = 0; d < D; d++) {
+    float diff = a[d] - b[d];
+    sum += diff * diff;
+  }
+  return sum;
+}
+
+// Find nearest centroid for a single vector. Returns centroid index.
+static int ivf_nearest_centroid(const float *vec, const float *centroids,
+                                int D, int k) {
+  float min_dist = FLT_MAX;
+  int best = 0;
+  for (int c = 0; c < k; c++) {
+    float dist = ivf_l2_dist(vec, &centroids[c * D], D);
+    if (dist < min_dist) {
+      min_dist = dist;
+      best = c;
+    }
+  }
+  return best;
+}
+
+/**
+ * K-means++ initialization.
+ * Picks k initial centroids from the data with probability proportional
+ * to squared distance from nearest existing centroid.
+ */
+static int ivf_kmeans_init_plusplus(const float *vectors, int N, int D,
+                                    int k, uint32_t seed, float *centroids) {
+  if (N <= 0 || k <= 0 || D <= 0)
+    return -1;
+  if (seed == 0)
+    seed = 42;
+
+  // Pick first centroid randomly
+  int first = ivf_xorshift32(&seed) % N;
+  memcpy(centroids, &vectors[first * D], D * sizeof(float));
+
+  if (k == 1)
+    return 0;
+
+  // Allocate distance array
+  float *dists = sqlite3_malloc64((i64)N * sizeof(float));
+  if (!dists)
+    return -1;
+
+  for (int c = 1; c < k; c++) {
+    // Compute D(x) = distance to nearest existing centroid
+    double total = 0.0;
+    for (int i = 0; i < N; i++) {
+      float d = ivf_l2_dist(&vectors[i * D], &centroids[(c - 1) * D], D);
+      if (c == 1 || d < dists[i]) {
+        dists[i] = d;
+      }
+      total += dists[i];
+    }
+
+    // Weighted random selection
+    if (total <= 0.0) {
+      // All distances zero — pick randomly
+      int pick = ivf_xorshift32(&seed) % N;
+      memcpy(&centroids[c * D], &vectors[pick * D], D * sizeof(float));
+    } else {
+      double threshold = ((double)ivf_xorshift32(&seed) / (double)0xFFFFFFFF) * total;
+      double cumulative = 0.0;
+      int pick = N - 1;
+      for (int i = 0; i < N; i++) {
+        cumulative += dists[i];
+        if (cumulative >= threshold) {
+          pick = i;
+          break;
+        }
+      }
+      memcpy(&centroids[c * D], &vectors[pick * D], D * sizeof(float));
+    }
+  }
+
+  sqlite3_free(dists);
+  return 0;
+}
+
+/**
+ * Lloyd's k-means algorithm.
+ *
+ * @param vectors   N*D float array (row-major)
+ * @param N         number of vectors
+ * @param D         dimensionality
+ * @param k         number of clusters
+ * @param max_iter  maximum iterations
+ * @param seed      PRNG seed for initialization
+ * @param out_centroids  output: k*D float array (caller-allocated)
+ * @return 0 on success, -1 on error
+ */
+static int ivf_kmeans(const float *vectors, int N, int D, int k,
+                       int max_iter, uint32_t seed, float *out_centroids) {
+  if (N <= 0 || D <= 0 || k <= 0)
+    return -1;
+
+  // Clamp k to N
+  if (k > N)
+    k = N;
+
+  // Allocate working memory
+  int *assignments = sqlite3_malloc64((i64)N * sizeof(int));
+  float *new_centroids = sqlite3_malloc64((i64)k * D * sizeof(float));
+  int *counts = sqlite3_malloc64((i64)k * sizeof(int));
+
+  if (!assignments || !new_centroids || !counts) {
+    sqlite3_free(assignments);
+    sqlite3_free(new_centroids);
+    sqlite3_free(counts);
+    return -1;
+  }
+
+  memset(assignments, -1, N * sizeof(int));
+
+  // Initialize centroids via k-means++
+  if (ivf_kmeans_init_plusplus(vectors, N, D, k, seed, out_centroids) != 0) {
+    sqlite3_free(assignments);
+    sqlite3_free(new_centroids);
+    sqlite3_free(counts);
+    return -1;
+  }
+
+  for (int iter = 0; iter < max_iter; iter++) {
+    // Assignment step
+    int changed = 0;
+    for (int i = 0; i < N; i++) {
+      int nearest = ivf_nearest_centroid(&vectors[i * D], out_centroids, D, k);
+      if (nearest != assignments[i]) {
+        assignments[i] = nearest;
+        changed++;
+      }
+    }
+    if (changed == 0)
+      break;
+
+    // Update step
+    memset(new_centroids, 0, (size_t)k * D * sizeof(float));
+    memset(counts, 0, k * sizeof(int));
+
+    for (int i = 0; i < N; i++) {
+      int c = assignments[i];
+      counts[c]++;
+      for (int d = 0; d < D; d++) {
+        new_centroids[c * D + d] += vectors[i * D + d];
+      }
+    }
+
+    for (int c = 0; c < k; c++) {
+      if (counts[c] == 0) {
+        // Empty cluster: reassign to farthest point from its nearest centroid
+        float max_dist = -1.0f;
+        int farthest = 0;
+        for (int i = 0; i < N; i++) {
+          float d = ivf_l2_dist(&vectors[i * D],
+                                &out_centroids[assignments[i] * D], D);
+          if (d > max_dist) {
+            max_dist = d;
+            farthest = i;
+          }
+        }
+        memcpy(&out_centroids[c * D], &vectors[farthest * D],
+               D * sizeof(float));
+      } else {
+        for (int d = 0; d < D; d++) {
+          out_centroids[c * D + d] = new_centroids[c * D + d] / counts[c];
+        }
+      }
+    }
+  }
+
+  sqlite3_free(assignments);
+  sqlite3_free(new_centroids);
+  sqlite3_free(counts);
+  return 0;
+}
+
+#endif /* SQLITE_VEC_IVF_KMEANS_C */

sqlite-vec-rescore.c

@@ -0,0 +1,687 @@
+/**
+ * sqlite-vec-rescore.c — Rescore index logic for sqlite-vec.
+ *
+ * This file is #included into sqlite-vec.c after the vec0_vtab definition.
+ * All functions receive a vec0_vtab *p and access p->vector_columns[i].rescore.
+ *
+ * Shadow tables per rescore-enabled vector column:
+ *   _rescore_chunks{NN}  — quantized vectors in chunk layout (for coarse scan)
+ *   _rescore_vectors{NN} — float vectors keyed by rowid (for fast rescore lookup)
+ */
+
+// ============================================================================
+// Shadow table lifecycle
+// ============================================================================
+
+static int rescore_create_tables(vec0_vtab *p, sqlite3 *db, char **pzErr) {
+  for (int i = 0; i < p->numVectorColumns; i++) {
+    if (p->vector_columns[i].index_type != VEC0_INDEX_TYPE_RESCORE)
+      continue;
+
+    // Quantized chunk table (same layout as _vector_chunks)
+    char *zSql = sqlite3_mprintf(
+        "CREATE TABLE \"%w\".\"%w_rescore_chunks%02d\""
+        "(rowid PRIMARY KEY, vectors BLOB NOT NULL)",
+        p->schemaName, p->tableName, i);
+    if (!zSql)
+      return SQLITE_NOMEM;
+    sqlite3_stmt *stmt;
+    int rc = sqlite3_prepare_v2(db, zSql, -1, &stmt, 0);
+    sqlite3_free(zSql);
+    if ((rc != SQLITE_OK) || (sqlite3_step(stmt) != SQLITE_DONE)) {
+      *pzErr = sqlite3_mprintf(
+          "Could not create '_rescore_chunks%02d' shadow table: %s", i,
+          sqlite3_errmsg(db));
+      sqlite3_finalize(stmt);
+      return SQLITE_ERROR;
+    }
+    sqlite3_finalize(stmt);
+
+    // Float vector table (rowid-keyed for fast random access)
+    zSql = sqlite3_mprintf(
+        "CREATE TABLE \"%w\".\"%w_rescore_vectors%02d\""
+        "(rowid INTEGER PRIMARY KEY, vector BLOB NOT NULL)",
+        p->schemaName, p->tableName, i);
+    if (!zSql)
+      return SQLITE_NOMEM;
+    rc = sqlite3_prepare_v2(db, zSql, -1, &stmt, 0);
+    sqlite3_free(zSql);
+    if ((rc != SQLITE_OK) || (sqlite3_step(stmt) != SQLITE_DONE)) {
+      *pzErr = sqlite3_mprintf(
+          "Could not create '_rescore_vectors%02d' shadow table: %s", i,
+          sqlite3_errmsg(db));
+      sqlite3_finalize(stmt);
+      return SQLITE_ERROR;
+    }
+    sqlite3_finalize(stmt);
+  }
+  return SQLITE_OK;
+}
+
+static int rescore_drop_tables(vec0_vtab *p) {
+  for (int i = 0; i < p->numVectorColumns; i++) {
+    sqlite3_stmt *stmt;
+    int rc;
+    char *zSql;
+
+    if (p->shadowRescoreChunksNames[i]) {
+      zSql = sqlite3_mprintf("DROP TABLE IF EXISTS \"%w\".\"%w\"",
+                              p->schemaName, p->shadowRescoreChunksNames[i]);
+      if (!zSql)
+        return SQLITE_NOMEM;
+      rc = sqlite3_prepare_v2(p->db, zSql, -1, &stmt, 0);
+      sqlite3_free(zSql);
+      if ((rc != SQLITE_OK) || (sqlite3_step(stmt) != SQLITE_DONE)) {
+        sqlite3_finalize(stmt);
+        return SQLITE_ERROR;
+      }
+      sqlite3_finalize(stmt);
+    }
+
+    if (p->shadowRescoreVectorsNames[i]) {
+      zSql = sqlite3_mprintf("DROP TABLE IF EXISTS \"%w\".\"%w\"",
+                              p->schemaName, p->shadowRescoreVectorsNames[i]);
+      if (!zSql)
+        return SQLITE_NOMEM;
+      rc = sqlite3_prepare_v2(p->db, zSql, -1, &stmt, 0);
+      sqlite3_free(zSql);
+      if ((rc != SQLITE_OK) || (sqlite3_step(stmt) != SQLITE_DONE)) {
+        sqlite3_finalize(stmt);
+        return SQLITE_ERROR;
+      }
+      sqlite3_finalize(stmt);
+    }
+  }
+  return SQLITE_OK;
+}
+
+static size_t rescore_quantized_byte_size(struct VectorColumnDefinition *col) {
+  switch (col->rescore.quantizer_type) {
+  case VEC0_RESCORE_QUANTIZER_BIT:
+    return col->dimensions / CHAR_BIT;
+  case VEC0_RESCORE_QUANTIZER_INT8:
+    return col->dimensions;
+  default:
+    return 0;
+  }
+}
+
+/**
+ * Insert a new chunk row into each _rescore_chunks{NN} table with a zeroblob.
+ */
+static int rescore_new_chunk(vec0_vtab *p, i64 chunk_rowid) {
+  for (int i = 0; i < p->numVectorColumns; i++) {
+    if (p->vector_columns[i].index_type != VEC0_INDEX_TYPE_RESCORE)
+      continue;
+    size_t quantized_size =
+        rescore_quantized_byte_size(&p->vector_columns[i]);
+    i64 blob_size = (i64)p->chunk_size * (i64)quantized_size;
+
+    char *zSql = sqlite3_mprintf(
+        "INSERT INTO \"%w\".\"%w\"(_rowid_, rowid, vectors) VALUES (?, ?, ?)",
+        p->schemaName, p->shadowRescoreChunksNames[i]);
+    if (!zSql)
+      return SQLITE_NOMEM;
+    sqlite3_stmt *stmt;
+    int rc = sqlite3_prepare_v2(p->db, zSql, -1, &stmt, NULL);
+    sqlite3_free(zSql);
+    if (rc != SQLITE_OK) {
+      sqlite3_finalize(stmt);
+      return rc;
+    }
+    sqlite3_bind_int64(stmt, 1, chunk_rowid);
+    sqlite3_bind_int64(stmt, 2, chunk_rowid);
+    sqlite3_bind_zeroblob64(stmt, 3, blob_size);
+    rc = sqlite3_step(stmt);
+    sqlite3_finalize(stmt);
+    if (rc != SQLITE_DONE)
+      return rc;
+  }
+  return SQLITE_OK;
+}
+
+// ============================================================================
+// Quantization
+// ============================================================================
+
+static void rescore_quantize_float_to_bit(const float *src, uint8_t *dst,
+                                          size_t dimensions) {
+  memset(dst, 0, dimensions / CHAR_BIT);
+  for (size_t i = 0; i < dimensions; i++) {
+    if (src[i] >= 0.0f) {
+      dst[i / CHAR_BIT] |= (1 << (i % CHAR_BIT));
+    }
+  }
+}
+
+static void rescore_quantize_float_to_int8(const float *src, int8_t *dst,
+                                           size_t dimensions) {
+  float step = 2.0f / 255.0f;
+  for (size_t i = 0; i < dimensions; i++) {
+    float v = (src[i] - (-1.0f)) / step - 128.0f;
+    if (!(v <= 127.0f)) v = 127.0f;
+    if (!(v >= -128.0f)) v = -128.0f;
+    dst[i] = (int8_t)v;
+  }
+}
+
+// ============================================================================
+// Insert path
+// ============================================================================
+
+/**
+ * Quantize float vector to _rescore_chunks and store in _rescore_vectors.
+ */
+static int rescore_on_insert(vec0_vtab *p, i64 chunk_rowid, i64 chunk_offset,
+                             i64 rowid, void *vectorDatas[]) {
+  for (int i = 0; i < p->numVectorColumns; i++) {
+    if (p->vector_columns[i].index_type != VEC0_INDEX_TYPE_RESCORE)
+      continue;
+
+    struct VectorColumnDefinition *col = &p->vector_columns[i];
+    size_t qsize = rescore_quantized_byte_size(col);
+    size_t fsize = vector_column_byte_size(*col);
+    int rc;
+
+    // 1. Write quantized vector to _rescore_chunks blob
+    {
+      void *qbuf = sqlite3_malloc(qsize);
+      if (!qbuf)
+        return SQLITE_NOMEM;
+
+      switch (col->rescore.quantizer_type) {
+      case VEC0_RESCORE_QUANTIZER_BIT:
+        rescore_quantize_float_to_bit((const float *)vectorDatas[i],
+                                      (uint8_t *)qbuf, col->dimensions);
+        break;
+      case VEC0_RESCORE_QUANTIZER_INT8:
+        rescore_quantize_float_to_int8((const float *)vectorDatas[i],
+                                       (int8_t *)qbuf, col->dimensions);
+        break;
+      }
+
+      sqlite3_blob *blob = NULL;
+      rc = sqlite3_blob_open(p->db, p->schemaName,
+                             p->shadowRescoreChunksNames[i], "vectors",
+                             chunk_rowid, 1, &blob);
+      if (rc != SQLITE_OK) {
+        sqlite3_free(qbuf);
+        return rc;
+      }
+      rc = sqlite3_blob_write(blob, qbuf, qsize, chunk_offset * qsize);
+      sqlite3_free(qbuf);
+      int brc = sqlite3_blob_close(blob);
+      if (rc != SQLITE_OK)
+        return rc;
+      if (brc != SQLITE_OK)
+        return brc;
+    }
+
+    // 2. Insert float vector into _rescore_vectors (rowid-keyed)
+    {
+      char *zSql = sqlite3_mprintf(
+          "INSERT INTO \"%w\".\"%w\"(rowid, vector) VALUES (?, ?)",
+          p->schemaName, p->shadowRescoreVectorsNames[i]);
+      if (!zSql)
+        return SQLITE_NOMEM;
+      sqlite3_stmt *stmt;
+      rc = sqlite3_prepare_v2(p->db, zSql, -1, &stmt, NULL);
+      sqlite3_free(zSql);
+      if (rc != SQLITE_OK) {
+        sqlite3_finalize(stmt);
+        return rc;
+      }
+      sqlite3_bind_int64(stmt, 1, rowid);
+      sqlite3_bind_blob(stmt, 2, vectorDatas[i], fsize, SQLITE_TRANSIENT);
+      rc = sqlite3_step(stmt);
+      sqlite3_finalize(stmt);
+      if (rc != SQLITE_DONE)
+        return SQLITE_ERROR;
+    }
+  }
+  return SQLITE_OK;
+}
+
+// ============================================================================
+// Delete path
+// ============================================================================
+
+/**
+ * Zero out quantized vector in _rescore_chunks and delete from _rescore_vectors.
+ */
+static int rescore_on_delete(vec0_vtab *p, i64 chunk_id, u64 chunk_offset,
+                             i64 rowid) {
+  for (int i = 0; i < p->numVectorColumns; i++) {
+    if (p->vector_columns[i].index_type != VEC0_INDEX_TYPE_RESCORE)
+      continue;
+    int rc;
+
+    // 1. Zero out quantized data in _rescore_chunks
+    {
+      size_t qsize = rescore_quantized_byte_size(&p->vector_columns[i]);
+      void *zeroBuf = sqlite3_malloc(qsize);
+      if (!zeroBuf)
+        return SQLITE_NOMEM;
+      memset(zeroBuf, 0, qsize);
+
+      sqlite3_blob *blob = NULL;
+      rc = sqlite3_blob_open(p->db, p->schemaName,
+                             p->shadowRescoreChunksNames[i], "vectors",
+                             chunk_id, 1, &blob);
+      if (rc != SQLITE_OK) {
+        sqlite3_free(zeroBuf);
+        return rc;
+      }
+      rc = sqlite3_blob_write(blob, zeroBuf, qsize, chunk_offset * qsize);
+      sqlite3_free(zeroBuf);
+      int brc = sqlite3_blob_close(blob);
+      if (rc != SQLITE_OK)
+        return rc;
+      if (brc != SQLITE_OK)
+        return brc;
+    }
+
+    // 2. Delete from _rescore_vectors
+    {
+      char *zSql = sqlite3_mprintf(
+          "DELETE FROM \"%w\".\"%w\" WHERE rowid = ?",
+          p->schemaName, p->shadowRescoreVectorsNames[i]);
+      if (!zSql)
+        return SQLITE_NOMEM;
+      sqlite3_stmt *stmt;
+      rc = sqlite3_prepare_v2(p->db, zSql, -1, &stmt, NULL);
+      sqlite3_free(zSql);
+      if (rc != SQLITE_OK)
+        return rc;
+      sqlite3_bind_int64(stmt, 1, rowid);
+      rc = sqlite3_step(stmt);
+      sqlite3_finalize(stmt);
+      if (rc != SQLITE_DONE)
+        return SQLITE_ERROR;
+    }
+  }
+  return SQLITE_OK;
+}
+
+/**
+ * Delete a chunk row from _rescore_chunks{NN} tables.
+ * (_rescore_vectors rows were already deleted per-row in rescore_on_delete)
+ */
+static int rescore_delete_chunk(vec0_vtab *p, i64 chunk_id) {
+  for (int i = 0; i < p->numVectorColumns; i++) {
+    if (!p->shadowRescoreChunksNames[i])
+      continue;
+    char *zSql = sqlite3_mprintf(
+        "DELETE FROM \"%w\".\"%w\" WHERE rowid = ?",
+        p->schemaName, p->shadowRescoreChunksNames[i]);
+    if (!zSql)
+      return SQLITE_NOMEM;
+    sqlite3_stmt *stmt;
+    int rc = sqlite3_prepare_v2(p->db, zSql, -1, &stmt, NULL);
+    sqlite3_free(zSql);
+    if (rc != SQLITE_OK)
+      return rc;
+    sqlite3_bind_int64(stmt, 1, chunk_id);
+    rc = sqlite3_step(stmt);
+    sqlite3_finalize(stmt);
+    if (rc != SQLITE_DONE)
+      return SQLITE_ERROR;
+  }
+  return SQLITE_OK;
+}
+
+// ============================================================================
+// KNN rescore query
+// ============================================================================
+
+/**
+ * Phase 1: Coarse scan of quantized chunks → top k*oversample candidates (rowids).
+ * Phase 2: For each candidate, blob_open _rescore_vectors by rowid, read float
+ *          vector, compute float distance. Sort, return top k.
+ *
+ * Phase 2 is fast because _rescore_vectors has INTEGER PRIMARY KEY, so
+ * sqlite3_blob_open/reopen addresses rows directly by rowid — no index lookup.
+ */
+static int rescore_knn(vec0_vtab *p, vec0_cursor *pCur,
+                       struct VectorColumnDefinition *vector_column,
+                       int vectorColumnIdx, struct Array *arrayRowidsIn,
+                       struct Array *aMetadataIn, const char *idxStr, int argc,
+                       sqlite3_value **argv, void *queryVector, i64 k,
+                       struct vec0_query_knn_data *knn_data) {
+  (void)pCur;
+  (void)aMetadataIn;
+  int rc = SQLITE_OK;
+  int oversample = vector_column->rescore.oversample_search > 0
+      ? vector_column->rescore.oversample_search
+      : vector_column->rescore.oversample;
+  i64 k_oversample = k * oversample;
+  if (k_oversample > 4096)
+    k_oversample = 4096;
+
+  size_t qdim = vector_column->dimensions;
+  size_t qsize = rescore_quantized_byte_size(vector_column);
+  size_t fsize = vector_column_byte_size(*vector_column);
+
+  // Quantize the query vector
+  void *quantizedQuery = sqlite3_malloc(qsize);
+  if (!quantizedQuery)
+    return SQLITE_NOMEM;
+
+  switch (vector_column->rescore.quantizer_type) {
+  case VEC0_RESCORE_QUANTIZER_BIT:
+    rescore_quantize_float_to_bit((const float *)queryVector,
+                                  (uint8_t *)quantizedQuery, qdim);
+    break;
+  case VEC0_RESCORE_QUANTIZER_INT8:
+    rescore_quantize_float_to_int8((const float *)queryVector,
+                                   (int8_t *)quantizedQuery, qdim);
+    break;
+  }
+
+  // Phase 1: Scan quantized chunks for k*oversample candidates
+  sqlite3_stmt *stmtChunks = NULL;
+  rc = vec0_chunks_iter(p, idxStr, argc, argv, &stmtChunks);
+  if (rc != SQLITE_OK) {
+    sqlite3_free(quantizedQuery);
+    return rc;
+  }
+
+  i64 *cand_rowids = sqlite3_malloc(k_oversample * sizeof(i64));
+  f32 *cand_distances = sqlite3_malloc(k_oversample * sizeof(f32));
+  i64 *tmp_rowids = sqlite3_malloc(k_oversample * sizeof(i64));
+  f32 *tmp_distances = sqlite3_malloc(k_oversample * sizeof(f32));
+  f32 *chunk_distances = sqlite3_malloc(p->chunk_size * sizeof(f32));
+  i32 *chunk_topk_idxs = sqlite3_malloc(k_oversample * sizeof(i32));
+  u8 *b = sqlite3_malloc(p->chunk_size / CHAR_BIT);
+  u8 *bTaken = sqlite3_malloc(p->chunk_size / CHAR_BIT);
+  u8 *bmRowids = NULL;
+  void *baseVectors = sqlite3_malloc((i64)p->chunk_size * (i64)qsize);
+
+  if (!cand_rowids || !cand_distances || !tmp_rowids || !tmp_distances ||
+      !chunk_distances || !chunk_topk_idxs || !b || !bTaken || !baseVectors) {
+    rc = SQLITE_NOMEM;
+    goto cleanup;
+  }
+  memset(cand_rowids, 0, k_oversample * sizeof(i64));
+  memset(cand_distances, 0, k_oversample * sizeof(f32));
+
+  if (arrayRowidsIn) {
+    bmRowids = sqlite3_malloc(p->chunk_size / CHAR_BIT);
+    if (!bmRowids) {
+      rc = SQLITE_NOMEM;
+      goto cleanup;
+    }
+  }
+
+  i64 cand_used = 0;
+
+  while (1) {
+    rc = sqlite3_step(stmtChunks);
+    if (rc == SQLITE_DONE)
+      break;
+    if (rc != SQLITE_ROW) {
+      rc = SQLITE_ERROR;
+      goto cleanup;
+    }
+
+    i64 chunk_id = sqlite3_column_int64(stmtChunks, 0);
+    unsigned char *chunkValidity =
+        (unsigned char *)sqlite3_column_blob(stmtChunks, 1);
+    i64 *chunkRowids = (i64 *)sqlite3_column_blob(stmtChunks, 2);
+    int validityBytes = sqlite3_column_bytes(stmtChunks, 1);
+    int rowidsBytes = sqlite3_column_bytes(stmtChunks, 2);
+    if (!chunkValidity || !chunkRowids) {
+      rc = SQLITE_ERROR;
+      goto cleanup;
+    }
+    // Validate blob sizes match chunk_size expectations
+    if (validityBytes < (p->chunk_size + 7) / 8 ||
+        rowidsBytes < p->chunk_size * (int)sizeof(i64)) {
+      rc = SQLITE_ERROR;
+      goto cleanup;
+    }
+
+    memset(chunk_distances, 0, p->chunk_size * sizeof(f32));
+    memset(chunk_topk_idxs, 0, k_oversample * sizeof(i32));
+    bitmap_copy(b, chunkValidity, p->chunk_size);
+
+    if (arrayRowidsIn) {
+      bitmap_clear(bmRowids, p->chunk_size);
+      for (int j = 0; j < p->chunk_size; j++) {
+        if (!bitmap_get(chunkValidity, j))
+          continue;
+        i64 rid = chunkRowids[j];
+        void *found = bsearch(&rid, arrayRowidsIn->z, arrayRowidsIn->length,
+                               sizeof(i64), _cmp);
+        bitmap_set(bmRowids, j, found ? 1 : 0);
+      }
+      bitmap_and_inplace(b, bmRowids, p->chunk_size);
+    }
+
+    // Read quantized vectors
+    sqlite3_blob *blobQ = NULL;
+    rc = sqlite3_blob_open(p->db, p->schemaName,
+                           p->shadowRescoreChunksNames[vectorColumnIdx],
+                           "vectors", chunk_id, 0, &blobQ);
+    if (rc != SQLITE_OK)
+      goto cleanup;
+    rc = sqlite3_blob_read(blobQ, baseVectors,
+                           (i64)p->chunk_size * (i64)qsize, 0);
+    sqlite3_blob_close(blobQ);
+    if (rc != SQLITE_OK)
+      goto cleanup;
+
+    // Compute quantized distances
+    for (int j = 0; j < p->chunk_size; j++) {
+      if (!bitmap_get(b, j))
+        continue;
+      f32 dist = FLT_MAX;
+      switch (vector_column->rescore.quantizer_type) {
+      case VEC0_RESCORE_QUANTIZER_BIT: {
+        const u8 *base_j = ((u8 *)baseVectors) + (j * (qdim / CHAR_BIT));
+        dist = distance_hamming(base_j, (u8 *)quantizedQuery, &qdim);
+        break;
+      }
+      case VEC0_RESCORE_QUANTIZER_INT8: {
+        const i8 *base_j = ((i8 *)baseVectors) + (j * qdim);
+        switch (vector_column->distance_metric) {
+        case VEC0_DISTANCE_METRIC_L2:
+          dist = distance_l2_sqr_int8(base_j, (i8 *)quantizedQuery, &qdim);
+          break;
+        case VEC0_DISTANCE_METRIC_COSINE:
+          dist = distance_cosine_int8(base_j, (i8 *)quantizedQuery, &qdim);
+          break;
+        case VEC0_DISTANCE_METRIC_L1:
+          dist = (f32)distance_l1_int8(base_j, (i8 *)quantizedQuery, &qdim);
+          break;
+        }
+        break;
+      }
+      }
+      chunk_distances[j] = dist;
+    }
+
+    int used1;
+    min_idx(chunk_distances, p->chunk_size, b, chunk_topk_idxs,
+            min(k_oversample, p->chunk_size), bTaken, &used1);
+
+    i64 merged_used;
+    merge_sorted_lists(cand_distances, cand_rowids, cand_used, chunk_distances,
+                       chunkRowids, chunk_topk_idxs,
+                       min(min(k_oversample, p->chunk_size), used1),
+                       tmp_distances, tmp_rowids, k_oversample, &merged_used);
+
+    for (i64 j = 0; j < merged_used; j++) {
+      cand_rowids[j] = tmp_rowids[j];
+      cand_distances[j] = tmp_distances[j];
+    }
+    cand_used = merged_used;
+  }
+  rc = SQLITE_OK;
+
+  // Phase 2: Rescore candidates using _rescore_vectors (rowid-keyed)
+  if (cand_used == 0) {
+    knn_data->current_idx = 0;
+    knn_data->k = 0;
+    knn_data->rowids = NULL;
+    knn_data->distances = NULL;
+    knn_data->k_used = 0;
+    goto cleanup;
+  }
+  {
+    f32 *float_distances = sqlite3_malloc(cand_used * sizeof(f32));
+    void *fBuf = sqlite3_malloc(fsize);
+    if (!float_distances || !fBuf) {
+      sqlite3_free(float_distances);
+      sqlite3_free(fBuf);
+      rc = SQLITE_NOMEM;
+      goto cleanup;
+    }
+
+    // Open blob on _rescore_vectors, then reopen for each candidate rowid.
+    // blob_reopen is O(1) for INTEGER PRIMARY KEY tables.
+    sqlite3_blob *blobFloat = NULL;
+    rc = sqlite3_blob_open(p->db, p->schemaName,
+                           p->shadowRescoreVectorsNames[vectorColumnIdx],
+                           "vector", cand_rowids[0], 0, &blobFloat);
+    if (rc != SQLITE_OK) {
+      sqlite3_free(float_distances);
+      sqlite3_free(fBuf);
+      goto cleanup;
+    }
+
+    rc = sqlite3_blob_read(blobFloat, fBuf, fsize, 0);
+    if (rc != SQLITE_OK) {
+      sqlite3_blob_close(blobFloat);
+      sqlite3_free(float_distances);
+      sqlite3_free(fBuf);
+      goto cleanup;
+    }
+    float_distances[0] =
+        vec0_distance_full(fBuf, queryVector, vector_column->dimensions,
+                           vector_column->element_type,
+                           vector_column->distance_metric);
+
+    for (i64 j = 1; j < cand_used; j++) {
+      rc = sqlite3_blob_reopen(blobFloat, cand_rowids[j]);
+      if (rc != SQLITE_OK) {
+        sqlite3_blob_close(blobFloat);
+        sqlite3_free(float_distances);
+        sqlite3_free(fBuf);
+        goto cleanup;
+      }
+      rc = sqlite3_blob_read(blobFloat, fBuf, fsize, 0);
+      if (rc != SQLITE_OK) {
+        sqlite3_blob_close(blobFloat);
+        sqlite3_free(float_distances);
+        sqlite3_free(fBuf);
+        goto cleanup;
+      }
+      float_distances[j] =
+          vec0_distance_full(fBuf, queryVector, vector_column->dimensions,
+                             vector_column->element_type,
+                             vector_column->distance_metric);
+    }
+    sqlite3_blob_close(blobFloat);
+    sqlite3_free(fBuf);
+
+    // Sort by float distance
+    for (i64 a = 0; a + 1 < cand_used; a++) {
+      i64 minIdx = a;
+      for (i64 c = a + 1; c < cand_used; c++) {
+        if (float_distances[c] < float_distances[minIdx])
+          minIdx = c;
+      }
+      if (minIdx != a) {
+        f32 td = float_distances[a];
+        float_distances[a] = float_distances[minIdx];
+        float_distances[minIdx] = td;
+        i64 tr = cand_rowids[a];
+        cand_rowids[a] = cand_rowids[minIdx];
+        cand_rowids[minIdx] = tr;
+      }
+    }
+
+    i64 result_k = min(k, cand_used);
+    i64 *out_rowids = sqlite3_malloc(result_k * sizeof(i64));
+    f32 *out_distances = sqlite3_malloc(result_k * sizeof(f32));
+    if (!out_rowids || !out_distances) {
+      sqlite3_free(out_rowids);
+      sqlite3_free(out_distances);
+      sqlite3_free(float_distances);
+      rc = SQLITE_NOMEM;
+      goto cleanup;
+    }
+    for (i64 j = 0; j < result_k; j++) {
+      out_rowids[j] = cand_rowids[j];
+      out_distances[j] = float_distances[j];
+    }
+
+    knn_data->current_idx = 0;
+    knn_data->k = result_k;
+    knn_data->rowids = out_rowids;
+    knn_data->distances = out_distances;
+    knn_data->k_used = result_k;
+
+    sqlite3_free(float_distances);
+  }
+
+cleanup:
+  sqlite3_finalize(stmtChunks);
+  sqlite3_free(quantizedQuery);
+  sqlite3_free(cand_rowids);
+  sqlite3_free(cand_distances);
+  sqlite3_free(tmp_rowids);
+  sqlite3_free(tmp_distances);
+  sqlite3_free(chunk_distances);
+  sqlite3_free(chunk_topk_idxs);
+  sqlite3_free(b);
+  sqlite3_free(bTaken);
+  sqlite3_free(bmRowids);
+  sqlite3_free(baseVectors);
+  return rc;
+}
+
+/**
+ * Handle FTS5-style command dispatch for rescore parameters.
+ * Returns SQLITE_OK if handled, SQLITE_EMPTY if not a rescore command.
+ */
+static int rescore_handle_command(vec0_vtab *p, const char *command) {
+  if (strncmp(command, "oversample=", 11) == 0) {
+    int val = atoi(command + 11);
+    if (val < 1) {
+      vtab_set_error(&p->base, "oversample must be >= 1");
+      return SQLITE_ERROR;
+    }
+    for (int i = 0; i < p->numVectorColumns; i++) {
+      if (p->vector_columns[i].index_type == VEC0_INDEX_TYPE_RESCORE) {
+        p->vector_columns[i].rescore.oversample_search = val;
+      }
+    }
+    return SQLITE_OK;
+  }
+  return SQLITE_EMPTY;
+}
+
+#ifdef SQLITE_VEC_TEST
+void _test_rescore_quantize_float_to_bit(const float *src, uint8_t *dst, size_t dim) {
+  rescore_quantize_float_to_bit(src, dst, dim);
+}
+void _test_rescore_quantize_float_to_int8(const float *src, int8_t *dst, size_t dim) {
+  rescore_quantize_float_to_int8(src, dst, dim);
+}
+size_t _test_rescore_quantized_byte_size_bit(size_t dimensions) {
+  struct VectorColumnDefinition col;
+  memset(&col, 0, sizeof(col));
+  col.dimensions = dimensions;
+  col.rescore.quantizer_type = VEC0_RESCORE_QUANTIZER_BIT;
+  return rescore_quantized_byte_size(&col);
+}
+size_t _test_rescore_quantized_byte_size_int8(size_t dimensions) {
+  struct VectorColumnDefinition col;
+  memset(&col, 0, sizeof(col));
+  col.dimensions = dimensions;
+  col.rescore.quantizer_type = VEC0_RESCORE_QUANTIZER_INT8;
+  return rescore_quantized_byte_size(&col);
+}
+#endif

tests/__snapshots__/test-metadata.ambr

@@ -27,8 +27,8 @@
         OrderedDict({
           'chunk_id': 1,
           'size': 8,
-          'validity': b'\x06',
+          'validity': b'\x02',
-          'rowids': b'\x00\x00\x00\x00\x00\x00\x00\x00\x02\x00\x00\x00\x00\x00\x00\x00\x03\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00',
+          'rowids': b'\x00\x00\x00\x00\x00\x00\x00\x00\x02\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00',
         }),
       ]),
     }),
@@ -37,7 +37,7 @@
       'rows': list([
         OrderedDict({
           'rowid': 1,
-          'data': b'\x06',
+          'data': b'\x02',
         }),
       ]),
     }),
@@ -46,7 +46,7 @@
       'rows': list([
         OrderedDict({
           'rowid': 1,
-          'data': b'\x00\x00\x00\x00\x00\x00\x00\x00\x02\x00\x00\x00\x00\x00\x00\x00\x03\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00',
+          'data': b'\x00\x00\x00\x00\x00\x00\x00\x00\x02\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00',
         }),
       ]),
     }),
@@ -55,7 +55,7 @@
       'rows': list([
         OrderedDict({
           'rowid': 1,
-          'data': b'\x00\x00\x00\x00\x00\x00\x00\x00\x9a\x99\x99\x99\x99\x99\x01@ffffff\n@\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00',
+          'data': b'\x00\x00\x00\x00\x00\x00\x00\x00\x9a\x99\x99\x99\x99\x99\x01@\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00',
         }),
       ]),
     }),
@@ -64,17 +64,13 @@
       'rows': list([
         OrderedDict({
           'rowid': 1,
-          'data': b'\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x05\x00\x00\x00test2\x00\x00\x00\x00\x00\x00\x00\r\x00\x00\x00123456789012\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00',
+          'data': b'\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x05\x00\x00\x00test2\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00',
         }),
       ]),
     }),
     'v_metadatatext03': OrderedDict({
       'sql': 'select * from v_metadatatext03',
       'rows': list([
-        OrderedDict({
-          'rowid': 3,
-          'data': '1234567890123',
-        }),
       ]),
     }),
     'v_rowids': OrderedDict({
@@ -86,12 +82,6 @@
           'chunk_id': 1,
           'chunk_offset': 1,
         }),
-        OrderedDict({
-          'rowid': 3,
-          'id': None,
-          'chunk_id': 1,
-          'chunk_offset': 2,
-        }),
       ]),
     }),
     'v_vector_chunks00': OrderedDict({
@@ -99,7 +89,7 @@
       'rows': list([
         OrderedDict({
           'rowid': 1,
-          'vectors': b'\x00\x00\x00\x00""""3333\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00',
+          'vectors': b'\x00\x00\x00\x00""""\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00',
         }),
       ]),
     }),
@@ -370,14 +360,6 @@
         'f': 2.2,
         't': 'test2',
       }),
-      OrderedDict({
-        'rowid': 3,
-        'vector': b'3333',
-        'b': 1,
-        'n': 3,
-        'f': 3.3,
-        't': '1234567890123',
-      }),
     ]),
   })
 # ---

tests/conftest.py

@@ -1,5 +1,29 @@
 import pytest
 import sqlite3
+import os
+
+
+def _vec_debug():
+    db = sqlite3.connect(":memory:")
+    db.enable_load_extension(True)
+    db.load_extension("dist/vec0")
+    db.enable_load_extension(False)
+    return db.execute("SELECT vec_debug()").fetchone()[0]
+
+
+def _has_build_flag(flag):
+    return flag in _vec_debug().split("Build flags:")[-1]
+
+
+def pytest_collection_modifyitems(config, items):
+    has_ivf = _has_build_flag("ivf")
+    if has_ivf:
+        return
+    skip_ivf = pytest.mark.skip(reason="IVF not enabled (compile with -DSQLITE_VEC_EXPERIMENTAL_IVF_ENABLE=1)")
+    ivf_prefixes = ("test-ivf",)
+    for item in items:
+        if any(item.fspath.basename.startswith(p) for p in ivf_prefixes):
+            item.add_marker(skip_ivf)
 
 
 @pytest.fixture()

tests/correctness/test-correctness.py

@@ -48,7 +48,6 @@ import json
 db = sqlite3.connect(":memory:")
 db.enable_load_extension(True)
 db.load_extension("../../dist/vec0")
-db.execute("select load_extension('../../dist/vec0', 'sqlite3_vec_fs_read_init')")
 db.enable_load_extension(False)
 
 results = db.execute(
@@ -75,17 +74,21 @@ print(b)
 
 db.execute('PRAGMA page_size=16384')
 
-print("Loading into sqlite-vec vec0 table...")
-t0 = time.time()
-db.execute("create virtual table v using vec0(a float[3072], chunk_size=16)")
-db.execute('insert into v select rowid, vector from vec_npy_each(vec_npy_file("dbpedia_openai_3_large_00.npy"))')
-print(time.time() - t0)
-
 print("loading numpy array...")
 t0 = time.time()
 base = np.load('dbpedia_openai_3_large_00.npy')
 print(time.time() - t0)
 
+print("Loading into sqlite-vec vec0 table...")
+t0 = time.time()
+db.execute("create virtual table v using vec0(a float[3072], chunk_size=16)")
+with db:
+    db.executemany(
+        "insert into v(rowid, a) values (?, ?)",
+        [(i, row.tobytes()) for i, row in enumerate(base)],
+    )
+print(time.time() - t0)
+
 np.random.seed(1)
 queries = base[np.random.choice(base.shape[0], 20, replace=False), :]
 

tests/fuzz/.gitignore

@@ -1,2 +1,7 @@
 *.dSYM
 targets/
+corpus/
+crash-*
+leak-*
+timeout-*
+*.log

tests/fuzz/Makefile

@@ -26,7 +26,7 @@ FUZZ_LDFLAGS ?= $(shell \
     echo "-Wl,-ld_classic"; \
   fi)
 
-FUZZ_CFLAGS = $(FUZZ_SANITIZERS) -I ../../ -I ../../vendor -DSQLITE_CORE -g $(FUZZ_LDFLAGS)
+FUZZ_CFLAGS = $(FUZZ_SANITIZERS) -I ../../ -I ../../vendor -DSQLITE_CORE -DSQLITE_VEC_ENABLE_DISKANN=1 -g $(FUZZ_LDFLAGS)
 FUZZ_SRCS = ../../vendor/sqlite3.c ../../sqlite-vec.c
 
 TARGET_DIR = ./targets
@@ -72,10 +72,94 @@ $(TARGET_DIR)/vec_mismatch: vec-mismatch.c $(FUZZ_SRCS) | $(TARGET_DIR)
 $(TARGET_DIR)/vec0_delete_completeness: vec0-delete-completeness.c $(FUZZ_SRCS) | $(TARGET_DIR)
 	$(FUZZ_CC) $(FUZZ_CFLAGS) $(FUZZ_SRCS) $< -o $@
 
+$(TARGET_DIR)/rescore_operations: rescore-operations.c $(FUZZ_SRCS) | $(TARGET_DIR)
+	$(FUZZ_CC) $(FUZZ_CFLAGS) -DSQLITE_VEC_ENABLE_RESCORE $(FUZZ_SRCS) $< -o $@
+
+$(TARGET_DIR)/rescore_create: rescore-create.c $(FUZZ_SRCS) | $(TARGET_DIR)
+	$(FUZZ_CC) $(FUZZ_CFLAGS) -DSQLITE_VEC_ENABLE_RESCORE $(FUZZ_SRCS) $< -o $@
+
+$(TARGET_DIR)/rescore_quantize: rescore-quantize.c $(FUZZ_SRCS) | $(TARGET_DIR)
+	$(FUZZ_CC) $(FUZZ_CFLAGS) -DSQLITE_VEC_ENABLE_RESCORE -DSQLITE_VEC_TEST $(FUZZ_SRCS) $< -o $@
+
+$(TARGET_DIR)/rescore_shadow_corrupt: rescore-shadow-corrupt.c $(FUZZ_SRCS) | $(TARGET_DIR)
+	$(FUZZ_CC) $(FUZZ_CFLAGS) -DSQLITE_VEC_ENABLE_RESCORE $(FUZZ_SRCS) $< -o $@
+
+$(TARGET_DIR)/rescore_knn_deep: rescore-knn-deep.c $(FUZZ_SRCS) | $(TARGET_DIR)
+	$(FUZZ_CC) $(FUZZ_CFLAGS) -DSQLITE_VEC_ENABLE_RESCORE $(FUZZ_SRCS) $< -o $@
+
+$(TARGET_DIR)/rescore_quantize_edge: rescore-quantize-edge.c $(FUZZ_SRCS) | $(TARGET_DIR)
+	$(FUZZ_CC) $(FUZZ_CFLAGS) -DSQLITE_VEC_ENABLE_RESCORE -DSQLITE_VEC_TEST $(FUZZ_SRCS) $< -o $@
+
+$(TARGET_DIR)/rescore_interleave: rescore-interleave.c $(FUZZ_SRCS) | $(TARGET_DIR)
+	$(FUZZ_CC) $(FUZZ_CFLAGS) -DSQLITE_VEC_ENABLE_RESCORE $(FUZZ_SRCS) $< -o $@
+
+$(TARGET_DIR)/ivf_create: ivf-create.c $(FUZZ_SRCS) | $(TARGET_DIR)
+	$(FUZZ_CC) $(FUZZ_CFLAGS) $(FUZZ_SRCS) $< -o $@
+
+$(TARGET_DIR)/ivf_operations: ivf-operations.c $(FUZZ_SRCS) | $(TARGET_DIR)
+	$(FUZZ_CC) $(FUZZ_CFLAGS) $(FUZZ_SRCS) $< -o $@
+
+$(TARGET_DIR)/ivf_quantize: ivf-quantize.c $(FUZZ_SRCS) | $(TARGET_DIR)
+	$(FUZZ_CC) $(FUZZ_CFLAGS) $(FUZZ_SRCS) $< -o $@
+
+$(TARGET_DIR)/ivf_kmeans: ivf-kmeans.c $(FUZZ_SRCS) | $(TARGET_DIR)
+	$(FUZZ_CC) $(FUZZ_CFLAGS) $(FUZZ_SRCS) $< -o $@
+
+$(TARGET_DIR)/ivf_shadow_corrupt: ivf-shadow-corrupt.c $(FUZZ_SRCS) | $(TARGET_DIR)
+	$(FUZZ_CC) $(FUZZ_CFLAGS) $(FUZZ_SRCS) $< -o $@
+
+$(TARGET_DIR)/ivf_knn_deep: ivf-knn-deep.c $(FUZZ_SRCS) | $(TARGET_DIR)
+	$(FUZZ_CC) $(FUZZ_CFLAGS) $(FUZZ_SRCS) $< -o $@
+
+$(TARGET_DIR)/ivf_cell_overflow: ivf-cell-overflow.c $(FUZZ_SRCS) | $(TARGET_DIR)
+	$(FUZZ_CC) $(FUZZ_CFLAGS) $(FUZZ_SRCS) $< -o $@
+
+$(TARGET_DIR)/ivf_rescore: ivf-rescore.c $(FUZZ_SRCS) | $(TARGET_DIR)
+$(TARGET_DIR)/diskann_operations: diskann-operations.c $(FUZZ_SRCS) | $(TARGET_DIR)
+	$(FUZZ_CC) $(FUZZ_CFLAGS) $(FUZZ_SRCS) $< -o $@
+
+$(TARGET_DIR)/diskann_create: diskann-create.c $(FUZZ_SRCS) | $(TARGET_DIR)
+	$(FUZZ_CC) $(FUZZ_CFLAGS) $(FUZZ_SRCS) $< -o $@
+
+$(TARGET_DIR)/diskann_graph_corrupt: diskann-graph-corrupt.c $(FUZZ_SRCS) | $(TARGET_DIR)
+	$(FUZZ_CC) $(FUZZ_CFLAGS) $(FUZZ_SRCS) $< -o $@
+
+$(TARGET_DIR)/diskann_deep_search: diskann-deep-search.c $(FUZZ_SRCS) | $(TARGET_DIR)
+	$(FUZZ_CC) $(FUZZ_CFLAGS) $(FUZZ_SRCS) $< -o $@
+
+$(TARGET_DIR)/diskann_blob_truncate: diskann-blob-truncate.c $(FUZZ_SRCS) | $(TARGET_DIR)
+	$(FUZZ_CC) $(FUZZ_CFLAGS) $(FUZZ_SRCS) $< -o $@
+
+$(TARGET_DIR)/diskann_delete_stress: diskann-delete-stress.c $(FUZZ_SRCS) | $(TARGET_DIR)
+	$(FUZZ_CC) $(FUZZ_CFLAGS) $(FUZZ_SRCS) $< -o $@
+
+$(TARGET_DIR)/diskann_buffer_flush: diskann-buffer-flush.c $(FUZZ_SRCS) | $(TARGET_DIR)
+	$(FUZZ_CC) $(FUZZ_CFLAGS) $(FUZZ_SRCS) $< -o $@
+
+$(TARGET_DIR)/diskann_int8_quant: diskann-int8-quant.c $(FUZZ_SRCS) | $(TARGET_DIR)
+	$(FUZZ_CC) $(FUZZ_CFLAGS) $(FUZZ_SRCS) $< -o $@
+
+$(TARGET_DIR)/diskann_prune_direct: diskann-prune-direct.c $(FUZZ_SRCS) | $(TARGET_DIR)
+	$(FUZZ_CC) $(FUZZ_CFLAGS) $(FUZZ_SRCS) $< -o $@
+
+$(TARGET_DIR)/diskann_command_inject: diskann-command-inject.c $(FUZZ_SRCS) | $(TARGET_DIR)
+	$(FUZZ_CC) $(FUZZ_CFLAGS) $(FUZZ_SRCS) $< -o $@
+
 FUZZ_TARGETS = vec0_create exec json numpy \
                shadow_corrupt vec0_operations scalar_functions \
                vec0_create_full metadata_columns vec_each vec_mismatch \
-               vec0_delete_completeness
+               vec0_delete_completeness \
+               rescore_operations rescore_create rescore_quantize \
+               rescore_shadow_corrupt rescore_knn_deep \
+               rescore_quantize_edge rescore_interleave \
+               ivf_create ivf_operations \
+               ivf_quantize ivf_kmeans ivf_shadow_corrupt \
+               ivf_knn_deep ivf_cell_overflow ivf_rescore
+               diskann_operations diskann_create diskann_graph_corrupt \
+               diskann_deep_search diskann_blob_truncate \
+               diskann_delete_stress diskann_buffer_flush \
+               diskann_int8_quant diskann_prune_direct \
+               diskann_command_inject
 
 all: $(addprefix $(TARGET_DIR)/,$(FUZZ_TARGETS))
 

tests/fuzz/diskann-blob-truncate.c

@@ -0,0 +1,250 @@
+/**
+ * Fuzz target for DiskANN shadow table blob size mismatches.
+ *
+ * The critical vulnerability: diskann_node_read() copies whatever blob size
+ * SQLite returns, but diskann_search/insert/delete index into those blobs
+ * using cfg->n_neighbors * sizeof(i64) etc. If the blob is truncated,
+ * extended, or has wrong size, this causes out-of-bounds reads/writes.
+ *
+ * This fuzzer:
+ *   1. Creates a valid DiskANN graph with several nodes
+ *   2. Uses fuzz data to directly write malformed blobs to shadow tables:
+ *      - Truncated neighbor_ids (fewer bytes than n_neighbors * 8)
+ *      - Truncated validity bitmaps
+ *      - Oversized blobs with garbage trailing data
+ *      - Zero-length blobs
+ *      - Blobs with valid headers but corrupted neighbor rowids
+ *   3. Runs INSERT, DELETE, and KNN operations that traverse the corrupted graph
+ *
+ * Key code paths targeted:
+ *   - diskann_node_read with mismatched blob sizes
+ *   - diskann_validity_get / diskann_neighbor_id_get on truncated blobs
+ *   - diskann_add_reverse_edge reading corrupted neighbor data
+ *   - diskann_repair_reverse_edges traversing corrupted neighbor lists
+ *   - diskann_search iterating neighbors from corrupted blobs
+ */
+#include <stdint.h>
+#include <stddef.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include "sqlite-vec.h"
+#include "sqlite3.h"
+#include <assert.h>
+
+static uint8_t fuzz_byte(const uint8_t **data, size_t *size, uint8_t def) {
+  if (*size == 0) return def;
+  uint8_t b = **data;
+  (*data)++;
+  (*size)--;
+  return b;
+}
+
+int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) {
+  if (size < 32) return 0;
+
+  int rc;
+  sqlite3 *db;
+
+  rc = sqlite3_open(":memory:", &db);
+  assert(rc == SQLITE_OK);
+  rc = sqlite3_vec_init(db, NULL, NULL);
+  assert(rc == SQLITE_OK);
+
+  /* Use binary quantizer, float[16], n_neighbors=8 for predictable blob sizes:
+   *   validity: 8/8 = 1 byte
+   *   neighbor_ids: 8 * 8 = 64 bytes
+   *   qvecs: 8 * (16/8) = 16 bytes  (binary: 2 bytes per qvec)
+   */
+  rc = sqlite3_exec(db,
+    "CREATE VIRTUAL TABLE v USING vec0("
+    "emb float[16] INDEXED BY diskann(neighbor_quantizer=binary, n_neighbors=8))",
+    NULL, NULL, NULL);
+  if (rc != SQLITE_OK) { sqlite3_close(db); return 0; }
+
+  /* Insert 12 vectors to create a valid graph structure */
+  {
+    sqlite3_stmt *stmt;
+    sqlite3_prepare_v2(db,
+      "INSERT INTO v(rowid, emb) VALUES (?, ?)", -1, &stmt, NULL);
+    for (int i = 1; i <= 12; i++) {
+      float vec[16];
+      for (int j = 0; j < 16; j++) {
+        vec[j] = (float)i * 0.1f + (float)j * 0.01f;
+      }
+      sqlite3_reset(stmt);
+      sqlite3_bind_int64(stmt, 1, i);
+      sqlite3_bind_blob(stmt, 2, vec, sizeof(vec), SQLITE_TRANSIENT);
+      sqlite3_step(stmt);
+    }
+    sqlite3_finalize(stmt);
+  }
+
+  /* Now corrupt shadow table blobs using fuzz data */
+  const char *columns[] = {
+    "neighbors_validity",
+    "neighbor_ids",
+    "neighbor_quantized_vectors"
+  };
+
+  /* Expected sizes for n_neighbors=8, dims=16, binary quantizer */
+  int expected_sizes[] = {1, 64, 16};
+
+  while (size >= 4) {
+    int target_row = (fuzz_byte(&data, &size, 0) % 12) + 1;
+    int col_idx = fuzz_byte(&data, &size, 0) % 3;
+    uint8_t corrupt_mode = fuzz_byte(&data, &size, 0) % 6;
+    uint8_t extra = fuzz_byte(&data, &size, 0);
+
+    char sqlbuf[256];
+    snprintf(sqlbuf, sizeof(sqlbuf),
+      "UPDATE v_diskann_nodes00 SET %s = ? WHERE rowid = ?",
+      columns[col_idx]);
+
+    sqlite3_stmt *writeStmt;
+    rc = sqlite3_prepare_v2(db, sqlbuf, -1, &writeStmt, NULL);
+    if (rc != SQLITE_OK) continue;
+
+    int expected = expected_sizes[col_idx];
+    unsigned char *blob = NULL;
+    int blob_size = 0;
+
+    switch (corrupt_mode) {
+      case 0: {
+        /* Truncated blob: 0 to expected-1 bytes */
+        blob_size = extra % expected;
+        if (blob_size == 0) blob_size = 0;  /* zero-length is interesting */
+        blob = sqlite3_malloc(blob_size > 0 ? blob_size : 1);
+        if (!blob) { sqlite3_finalize(writeStmt); continue; }
+        for (int i = 0; i < blob_size; i++) {
+          blob[i] = fuzz_byte(&data, &size, 0);
+        }
+        break;
+      }
+      case 1: {
+        /* Oversized blob: expected + extra bytes */
+        blob_size = expected + (extra % 64);
+        blob = sqlite3_malloc(blob_size);
+        if (!blob) { sqlite3_finalize(writeStmt); continue; }
+        for (int i = 0; i < blob_size; i++) {
+          blob[i] = fuzz_byte(&data, &size, 0xFF);
+        }
+        break;
+      }
+      case 2: {
+        /* Zero-length blob */
+        blob_size = 0;
+        blob = NULL;
+        sqlite3_bind_zeroblob(writeStmt, 1, 0);
+        sqlite3_bind_int64(writeStmt, 2, target_row);
+        sqlite3_step(writeStmt);
+        sqlite3_finalize(writeStmt);
+        continue;
+      }
+      case 3: {
+        /* Correct size but all-ones validity (all slots "valid") with
+         * garbage neighbor IDs -- forces reading non-existent nodes */
+        blob_size = expected;
+        blob = sqlite3_malloc(blob_size);
+        if (!blob) { sqlite3_finalize(writeStmt); continue; }
+        memset(blob, 0xFF, blob_size);
+        break;
+      }
+      case 4: {
+        /* neighbor_ids with very large rowid values (near INT64_MAX) */
+        blob_size = expected;
+        blob = sqlite3_malloc(blob_size);
+        if (!blob) { sqlite3_finalize(writeStmt); continue; }
+        memset(blob, 0x7F, blob_size); /* fills with large positive values */
+        break;
+      }
+      case 5: {
+        /* neighbor_ids with negative rowid values (rowid=0 is sentinel) */
+        blob_size = expected;
+        blob = sqlite3_malloc(blob_size);
+        if (!blob) { sqlite3_finalize(writeStmt); continue; }
+        memset(blob, 0x80, blob_size); /* fills with large negative values */
+        /* Flip some bytes from fuzz data */
+        for (int i = 0; i < blob_size && size > 0; i++) {
+          blob[i] ^= fuzz_byte(&data, &size, 0);
+        }
+        break;
+      }
+    }
+
+    if (blob) {
+      sqlite3_bind_blob(writeStmt, 1, blob, blob_size, SQLITE_TRANSIENT);
+    } else {
+      sqlite3_bind_blob(writeStmt, 1, "", 0, SQLITE_STATIC);
+    }
+    sqlite3_bind_int64(writeStmt, 2, target_row);
+    sqlite3_step(writeStmt);
+    sqlite3_finalize(writeStmt);
+    sqlite3_free(blob);
+  }
+
+  /* Exercise the corrupted graph with various operations */
+
+  /* KNN query */
+  {
+    float qvec[16];
+    for (int j = 0; j < 16; j++) qvec[j] = (float)j * 0.1f;
+    sqlite3_stmt *knnStmt;
+    rc = sqlite3_prepare_v2(db,
+      "SELECT rowid, distance FROM v WHERE emb MATCH ? AND k = 5",
+      -1, &knnStmt, NULL);
+    if (rc == SQLITE_OK) {
+      sqlite3_bind_blob(knnStmt, 1, qvec, sizeof(qvec), SQLITE_STATIC);
+      while (sqlite3_step(knnStmt) == SQLITE_ROW) {}
+      sqlite3_finalize(knnStmt);
+    }
+  }
+
+  /* Insert into corrupted graph (triggers add_reverse_edge on corrupted nodes) */
+  {
+    float vec[16];
+    for (int j = 0; j < 16; j++) vec[j] = 0.5f;
+    sqlite3_stmt *stmt;
+    sqlite3_prepare_v2(db,
+      "INSERT INTO v(rowid, emb) VALUES (?, ?)", -1, &stmt, NULL);
+    if (stmt) {
+      sqlite3_bind_int64(stmt, 1, 100);
+      sqlite3_bind_blob(stmt, 2, vec, sizeof(vec), SQLITE_TRANSIENT);
+      sqlite3_step(stmt);
+      sqlite3_finalize(stmt);
+    }
+  }
+
+  /* Delete from corrupted graph (triggers repair_reverse_edges) */
+  {
+    sqlite3_stmt *stmt;
+    sqlite3_prepare_v2(db,
+      "DELETE FROM v WHERE rowid = ?", -1, &stmt, NULL);
+    if (stmt) {
+      sqlite3_bind_int64(stmt, 1, 5);
+      sqlite3_step(stmt);
+      sqlite3_finalize(stmt);
+    }
+  }
+
+  /* Another KNN to traverse the post-mutation graph */
+  {
+    float qvec[16];
+    for (int j = 0; j < 16; j++) qvec[j] = -0.5f + (float)j * 0.07f;
+    sqlite3_stmt *knnStmt;
+    rc = sqlite3_prepare_v2(db,
+      "SELECT rowid, distance FROM v WHERE emb MATCH ? AND k = 12",
+      -1, &knnStmt, NULL);
+    if (rc == SQLITE_OK) {
+      sqlite3_bind_blob(knnStmt, 1, qvec, sizeof(qvec), SQLITE_STATIC);
+      while (sqlite3_step(knnStmt) == SQLITE_ROW) {}
+      sqlite3_finalize(knnStmt);
+    }
+  }
+
+  /* Full scan */
+  sqlite3_exec(db, "SELECT * FROM v", NULL, NULL, NULL);
+
+  sqlite3_close(db);
+  return 0;
+}

tests/fuzz/diskann-buffer-flush.c

@@ -0,0 +1,164 @@
+/**
+ * Fuzz target for DiskANN buffered insert and flush paths.
+ *
+ * When buffer_threshold > 0, inserts go into a flat buffer table and
+ * are flushed into the graph in batch. This fuzzer exercises:
+ *
+ *   - diskann_buffer_write / diskann_buffer_delete / diskann_buffer_exists
+ *   - diskann_flush_buffer (batch graph insertion)
+ *   - diskann_insert with buffer_threshold (batching logic)
+ *   - Buffer-graph merge in vec0Filter_knn_diskann (unflushed vectors
+ *     must be scanned during KNN and merged with graph results)
+ *   - Delete of a buffered (not yet flushed) vector
+ *   - Delete of a graph vector while buffer has pending inserts
+ *   - Interaction: insert to buffer, query (triggers buffer scan), flush,
+ *     query again (now from graph)
+ *
+ * The buffer merge path in vec0Filter_knn_diskann is particularly
+ * interesting because it does a brute-force scan of buffer vectors and
+ * merges with the top-k from graph search.
+ */
+#include <stdint.h>
+#include <stddef.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include "sqlite-vec.h"
+#include "sqlite3.h"
+#include <assert.h>
+
+static uint8_t fuzz_byte(const uint8_t **data, size_t *size, uint8_t def) {
+  if (*size == 0) return def;
+  uint8_t b = **data;
+  (*data)++;
+  (*size)--;
+  return b;
+}
+
+int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) {
+  if (size < 16) return 0;
+
+  int rc;
+  sqlite3 *db;
+  rc = sqlite3_open(":memory:", &db);
+  assert(rc == SQLITE_OK);
+  rc = sqlite3_vec_init(db, NULL, NULL);
+  assert(rc == SQLITE_OK);
+
+  /* buffer_threshold: small (3-8) to trigger frequent flushes */
+  int buf_threshold = 3 + (fuzz_byte(&data, &size, 0) % 6);
+  int dims = 8;
+
+  char sql[512];
+  snprintf(sql, sizeof(sql),
+    "CREATE VIRTUAL TABLE v USING vec0("
+    "emb float[%d] INDEXED BY diskann("
+    "neighbor_quantizer=binary, n_neighbors=8, "
+    "search_list_size=16, buffer_threshold=%d"
+    "))", dims, buf_threshold);
+
+  rc = sqlite3_exec(db, sql, NULL, NULL, NULL);
+  if (rc != SQLITE_OK) { sqlite3_close(db); return 0; }
+
+  sqlite3_stmt *stmtInsert = NULL, *stmtDelete = NULL, *stmtKnn = NULL;
+  sqlite3_prepare_v2(db,
+    "INSERT INTO v(rowid, emb) VALUES (?, ?)", -1, &stmtInsert, NULL);
+  sqlite3_prepare_v2(db,
+    "DELETE FROM v WHERE rowid = ?", -1, &stmtDelete, NULL);
+  sqlite3_prepare_v2(db,
+    "SELECT rowid, distance FROM v WHERE emb MATCH ? AND k = ?",
+    -1, &stmtKnn, NULL);
+
+  if (!stmtInsert || !stmtDelete || !stmtKnn) goto cleanup;
+
+  float vec[8];
+  int next_rowid = 1;
+
+  while (size >= 2) {
+    uint8_t op = fuzz_byte(&data, &size, 0) % 6;
+    uint8_t param = fuzz_byte(&data, &size, 0);
+
+    switch (op) {
+      case 0: { /* Insert: accumulates in buffer until threshold */
+        int64_t rowid = next_rowid++;
+        if (next_rowid > 64) next_rowid = 1; /* wrap around for reuse */
+        for (int j = 0; j < dims; j++) {
+          vec[j] = (float)((int8_t)fuzz_byte(&data, &size, 0)) / 10.0f;
+        }
+        sqlite3_reset(stmtInsert);
+        sqlite3_bind_int64(stmtInsert, 1, rowid);
+        sqlite3_bind_blob(stmtInsert, 2, vec, sizeof(vec), SQLITE_TRANSIENT);
+        sqlite3_step(stmtInsert);
+        break;
+      }
+      case 1: { /* KNN query while buffer may have unflushed vectors */
+        for (int j = 0; j < dims; j++) {
+          vec[j] = (float)((int8_t)fuzz_byte(&data, &size, 0)) / 10.0f;
+        }
+        int k = (param % 10) + 1;
+        sqlite3_reset(stmtKnn);
+        sqlite3_bind_blob(stmtKnn, 1, vec, sizeof(vec), SQLITE_TRANSIENT);
+        sqlite3_bind_int(stmtKnn, 2, k);
+        while (sqlite3_step(stmtKnn) == SQLITE_ROW) {}
+        break;
+      }
+      case 2: { /* Delete a potentially-buffered vector */
+        int64_t rowid = (int64_t)(param % 64) + 1;
+        sqlite3_reset(stmtDelete);
+        sqlite3_bind_int64(stmtDelete, 1, rowid);
+        sqlite3_step(stmtDelete);
+        break;
+      }
+      case 3: { /* Insert several at once to trigger flush mid-batch */
+        for (int i = 0; i < buf_threshold + 1 && size >= 2; i++) {
+          int64_t rowid = (int64_t)(fuzz_byte(&data, &size, 0) % 64) + 1;
+          for (int j = 0; j < dims; j++) {
+            vec[j] = (float)((int8_t)fuzz_byte(&data, &size, 0)) / 10.0f;
+          }
+          sqlite3_reset(stmtInsert);
+          sqlite3_bind_int64(stmtInsert, 1, rowid);
+          sqlite3_bind_blob(stmtInsert, 2, vec, sizeof(vec), SQLITE_TRANSIENT);
+          sqlite3_step(stmtInsert);
+        }
+        break;
+      }
+      case 4: { /* Insert then immediately delete (still in buffer) */
+        int64_t rowid = (int64_t)(param % 64) + 1;
+        for (int j = 0; j < dims; j++) vec[j] = 0.1f * param;
+        sqlite3_reset(stmtInsert);
+        sqlite3_bind_int64(stmtInsert, 1, rowid);
+        sqlite3_bind_blob(stmtInsert, 2, vec, sizeof(vec), SQLITE_TRANSIENT);
+        sqlite3_step(stmtInsert);
+
+        sqlite3_reset(stmtDelete);
+        sqlite3_bind_int64(stmtDelete, 1, rowid);
+        sqlite3_step(stmtDelete);
+        break;
+      }
+      case 5: { /* Query with k=0 and k=1 (boundary) */
+        for (int j = 0; j < dims; j++) vec[j] = 0.0f;
+        sqlite3_reset(stmtKnn);
+        sqlite3_bind_blob(stmtKnn, 1, vec, sizeof(vec), SQLITE_TRANSIENT);
+        sqlite3_bind_int(stmtKnn, 2, param % 2); /* k=0 or k=1 */
+        while (sqlite3_step(stmtKnn) == SQLITE_ROW) {}
+        break;
+      }
+    }
+  }
+
+  /* Final query to exercise post-operation state */
+  {
+    float qvec[8] = {1.0f, -1.0f, 0.5f, -0.5f, 0.0f, 0.0f, 0.0f, 0.0f};
+    sqlite3_reset(stmtKnn);
+    sqlite3_bind_blob(stmtKnn, 1, qvec, sizeof(qvec), SQLITE_TRANSIENT);
+    sqlite3_bind_int(stmtKnn, 2, 20);
+    while (sqlite3_step(stmtKnn) == SQLITE_ROW) {}
+  }
+
+cleanup:
+  sqlite3_finalize(stmtInsert);
+  sqlite3_finalize(stmtDelete);
+  sqlite3_finalize(stmtKnn);
+  sqlite3_close(db);
+  return 0;
+}

tests/fuzz/diskann-command-inject.c

@@ -0,0 +1,158 @@
+/**
+ * Fuzz target for DiskANN runtime command dispatch (diskann_handle_command).
+ *
+ * The command handler parses strings like "search_list_size_search=42" and
+ * modifies live DiskANN config. This fuzzer exercises:
+ *
+ *   - atoi on fuzz-controlled strings (integer overflow, negative, non-numeric)
+ *   - strncmp boundary with fuzz data (near-matches to valid commands)
+ *   - Changing search_list_size mid-operation (affects subsequent queries)
+ *   - Setting search_list_size to 1 (minimum - single-candidate beam search)
+ *   - Setting search_list_size very large (memory pressure)
+ *   - Interleaving command changes with inserts and queries
+ *
+ * Also tests the UPDATE v SET command = ? path through the vtable.
+ */
+#include <stdint.h>
+#include <stddef.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include "sqlite-vec.h"
+#include "sqlite3.h"
+#include <assert.h>
+
+static uint8_t fuzz_byte(const uint8_t **data, size_t *size, uint8_t def) {
+  if (*size == 0) return def;
+  uint8_t b = **data;
+  (*data)++;
+  (*size)--;
+  return b;
+}
+
+int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) {
+  if (size < 20) return 0;
+
+  int rc;
+  sqlite3 *db;
+  rc = sqlite3_open(":memory:", &db);
+  assert(rc == SQLITE_OK);
+  rc = sqlite3_vec_init(db, NULL, NULL);
+  assert(rc == SQLITE_OK);
+
+  rc = sqlite3_exec(db,
+    "CREATE VIRTUAL TABLE v USING vec0("
+    "emb float[8] INDEXED BY diskann(neighbor_quantizer=binary, n_neighbors=8))",
+    NULL, NULL, NULL);
+  if (rc != SQLITE_OK) { sqlite3_close(db); return 0; }
+
+  /* Insert some vectors first */
+  {
+    sqlite3_stmt *stmt;
+    sqlite3_prepare_v2(db,
+      "INSERT INTO v(v, emb) VALUES (?, ?)", -1, &stmt, NULL);
+    for (int i = 1; i <= 8; i++) {
+      float vec[8];
+      for (int j = 0; j < 8; j++) vec[j] = (float)i * 0.1f + (float)j * 0.01f;
+      sqlite3_reset(stmt);
+      sqlite3_bind_int64(stmt, 1, i);
+      sqlite3_bind_blob(stmt, 2, vec, sizeof(vec), SQLITE_TRANSIENT);
+      sqlite3_step(stmt);
+    }
+    sqlite3_finalize(stmt);
+  }
+
+  sqlite3_stmt *stmtCmd = NULL;
+  sqlite3_stmt *stmtInsert = NULL;
+  sqlite3_stmt *stmtKnn = NULL;
+
+  /* Commands are dispatched via INSERT INTO t(t) VALUES ('cmd_string') */
+  sqlite3_prepare_v2(db,
+    "INSERT INTO v(v) VALUES (?)", -1, &stmtCmd, NULL);
+  sqlite3_prepare_v2(db,
+    "INSERT INTO v(v, emb) VALUES (?, ?)", -1, &stmtInsert, NULL);
+  sqlite3_prepare_v2(db,
+    "SELECT rowid, distance FROM v WHERE emb MATCH ? AND k = ?",
+    -1, &stmtKnn, NULL);
+
+  if (!stmtCmd || !stmtInsert || !stmtKnn) goto cleanup;
+
+  /* Fuzz-driven command + operation interleaving */
+  while (size >= 2) {
+    uint8_t op = fuzz_byte(&data, &size, 0) % 5;
+
+    switch (op) {
+      case 0: { /* Send fuzz command string */
+        int cmd_len = fuzz_byte(&data, &size, 0) % 64;
+        char cmd[65];
+        for (int i = 0; i < cmd_len && size > 0; i++) {
+          cmd[i] = (char)fuzz_byte(&data, &size, 0);
+        }
+        cmd[cmd_len] = '\0';
+        sqlite3_reset(stmtCmd);
+        sqlite3_bind_text(stmtCmd, 1, cmd, -1, SQLITE_TRANSIENT);
+        sqlite3_step(stmtCmd); /* May fail -- that's expected */
+        break;
+      }
+      case 1: { /* Send valid-looking command with fuzz value */
+        const char *prefixes[] = {
+          "search_list_size=",
+          "search_list_size_search=",
+          "search_list_size_insert=",
+        };
+        int prefix_idx = fuzz_byte(&data, &size, 0) % 3;
+        int val = (int)(int8_t)fuzz_byte(&data, &size, 0);
+
+        char cmd[128];
+        snprintf(cmd, sizeof(cmd), "%s%d", prefixes[prefix_idx], val);
+        sqlite3_reset(stmtCmd);
+        sqlite3_bind_text(stmtCmd, 1, cmd, -1, SQLITE_TRANSIENT);
+        sqlite3_step(stmtCmd);
+        break;
+      }
+      case 2: { /* KNN query (uses whatever search_list_size is set) */
+        float qvec[8] = {1.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f};
+        qvec[0] = (float)((int8_t)fuzz_byte(&data, &size, 127)) / 10.0f;
+        int k = fuzz_byte(&data, &size, 3) % 10 + 1;
+        sqlite3_reset(stmtKnn);
+        sqlite3_bind_blob(stmtKnn, 1, qvec, sizeof(qvec), SQLITE_TRANSIENT);
+        sqlite3_bind_int(stmtKnn, 2, k);
+        while (sqlite3_step(stmtKnn) == SQLITE_ROW) {}
+        break;
+      }
+      case 3: { /* Insert (uses whatever search_list_size_insert is set) */
+        int64_t rowid = (int64_t)(fuzz_byte(&data, &size, 0) % 32) + 1;
+        float vec[8];
+        for (int j = 0; j < 8; j++) {
+          vec[j] = (float)((int8_t)fuzz_byte(&data, &size, 0)) / 10.0f;
+        }
+        sqlite3_reset(stmtInsert);
+        sqlite3_bind_int64(stmtInsert, 1, rowid);
+        sqlite3_bind_blob(stmtInsert, 2, vec, sizeof(vec), SQLITE_TRANSIENT);
+        sqlite3_step(stmtInsert);
+        break;
+      }
+      case 4: { /* Set search_list_size to extreme values */
+        const char *extreme_cmds[] = {
+          "search_list_size=1",
+          "search_list_size=2",
+          "search_list_size=1000",
+          "search_list_size_search=1",
+          "search_list_size_insert=1",
+        };
+        int idx = fuzz_byte(&data, &size, 0) % 5;
+        sqlite3_reset(stmtCmd);
+        sqlite3_bind_text(stmtCmd, 1, extreme_cmds[idx], -1, SQLITE_STATIC);
+        sqlite3_step(stmtCmd);
+        break;
+      }
+    }
+  }
+
+cleanup:
+  sqlite3_finalize(stmtCmd);
+  sqlite3_finalize(stmtInsert);
+  sqlite3_finalize(stmtKnn);
+  sqlite3_close(db);
+  return 0;
+}

tests/fuzz/diskann-create.c

@@ -0,0 +1,44 @@
+/**
+ * Fuzz target for DiskANN CREATE TABLE config parsing.
+ * Feeds fuzz data as the INDEXED BY diskann(...) option string.
+ */
+#include <stdint.h>
+#include <stddef.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include "sqlite-vec.h"
+#include "sqlite3.h"
+#include <assert.h>
+
+int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) {
+  if (size > 4096) return 0;  /* Limit input size */
+
+  int rc;
+  sqlite3 *db;
+  sqlite3_stmt *stmt;
+
+  rc = sqlite3_open(":memory:", &db);
+  assert(rc == SQLITE_OK);
+  rc = sqlite3_vec_init(db, NULL, NULL);
+  assert(rc == SQLITE_OK);
+
+  sqlite3_str *s = sqlite3_str_new(NULL);
+  assert(s);
+  sqlite3_str_appendall(s,
+    "CREATE VIRTUAL TABLE v USING vec0("
+    "emb float[64] INDEXED BY diskann(");
+  sqlite3_str_appendf(s, "%.*s", (int)size, data);
+  sqlite3_str_appendall(s, "))");
+  const char *zSql = sqlite3_str_finish(s);
+  assert(zSql);
+
+  rc = sqlite3_prepare_v2(db, zSql, -1, &stmt, NULL);
+  sqlite3_free((char *)zSql);
+  if (rc == SQLITE_OK) {
+    sqlite3_step(stmt);
+  }
+  sqlite3_finalize(stmt);
+  sqlite3_close(db);
+  return 0;
+}

tests/fuzz/diskann-deep-search.c

@@ -0,0 +1,187 @@
+/**
+ * Fuzz target for DiskANN greedy beam search deep paths.
+ *
+ * Builds a graph with enough nodes to force multi-hop traversal, then
+ * uses fuzz data to control: query vector values, k, search_list_size
+ * overrides, and interleaved insert/delete/query sequences that stress
+ * the candidate list growth, visited set hash collisions, and the
+ * re-ranking logic.
+ *
+ * Key code paths targeted:
+ *   - diskann_candidate_list_insert (sorted insert, dedup, eviction)
+ *   - diskann_visited_set (hash collisions, capacity)
+ *   - diskann_search (full beam search loop, re-ranking with exact dist)
+ *   - diskann_distance_quantized_precomputed (both binary and int8)
+ *   - Buffer merge in vec0Filter_knn_diskann
+ */
+#include <stdint.h>
+#include <stddef.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <math.h>
+#include "sqlite-vec.h"
+#include "sqlite3.h"
+#include <assert.h>
+
+/* Consume one byte from fuzz input, or return default. */
+static uint8_t fuzz_byte(const uint8_t **data, size_t *size, uint8_t def) {
+  if (*size == 0) return def;
+  uint8_t b = **data;
+  (*data)++;
+  (*size)--;
+  return b;
+}
+
+static uint16_t fuzz_u16(const uint8_t **data, size_t *size) {
+  uint8_t lo = fuzz_byte(data, size, 0);
+  uint8_t hi = fuzz_byte(data, size, 0);
+  return (uint16_t)hi << 8 | lo;
+}
+
+static float fuzz_float(const uint8_t **data, size_t *size) {
+  return (float)((int8_t)fuzz_byte(data, size, 0)) / 10.0f;
+}
+
+int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) {
+  if (size < 32) return 0;
+
+  /* Use first bytes to pick quantizer type and dimensions */
+  uint8_t quantizer_choice = fuzz_byte(&data, &size, 0) % 2;
+  const char *quantizer = quantizer_choice ? "int8" : "binary";
+
+  /* Dimensions must be divisible by 8. Pick from {8, 16, 32} */
+  int dim_choices[] = {8, 16, 32};
+  int dims = dim_choices[fuzz_byte(&data, &size, 0) % 3];
+
+  /* n_neighbors: 8 or 16 -- small to force full-neighbor scenarios quickly */
+  int n_neighbors = (fuzz_byte(&data, &size, 0) % 2) ? 16 : 8;
+
+  /* search_list_size: small so beam search terminates quickly but still exercises loops */
+  int search_list_size = 8 + (fuzz_byte(&data, &size, 0) % 24);
+
+  /* alpha: vary to test RobustPrune pruning logic */
+  float alpha_choices[] = {1.0f, 1.2f, 1.5f, 2.0f};
+  float alpha = alpha_choices[fuzz_byte(&data, &size, 0) % 4];
+
+  int rc;
+  sqlite3 *db;
+  rc = sqlite3_open(":memory:", &db);
+  assert(rc == SQLITE_OK);
+  rc = sqlite3_vec_init(db, NULL, NULL);
+  assert(rc == SQLITE_OK);
+
+  char sql[512];
+  snprintf(sql, sizeof(sql),
+    "CREATE VIRTUAL TABLE v USING vec0("
+    "emb float[%d] INDEXED BY diskann("
+    "neighbor_quantizer=%s, n_neighbors=%d, "
+    "search_list_size=%d"
+    "))", dims, quantizer, n_neighbors, search_list_size);
+
+  rc = sqlite3_exec(db, sql, NULL, NULL, NULL);
+  if (rc != SQLITE_OK) { sqlite3_close(db); return 0; }
+
+  sqlite3_stmt *stmtInsert = NULL, *stmtDelete = NULL, *stmtKnn = NULL;
+  sqlite3_prepare_v2(db,
+    "INSERT INTO v(rowid, emb) VALUES (?, ?)", -1, &stmtInsert, NULL);
+  sqlite3_prepare_v2(db,
+    "DELETE FROM v WHERE rowid = ?", -1, &stmtDelete, NULL);
+
+  char knn_sql[256];
+  snprintf(knn_sql, sizeof(knn_sql),
+    "SELECT rowid, distance FROM v WHERE emb MATCH ? AND k = ?");
+  sqlite3_prepare_v2(db, knn_sql, -1, &stmtKnn, NULL);
+
+  if (!stmtInsert || !stmtDelete || !stmtKnn) goto cleanup;
+
+  /* Phase 1: Seed the graph with enough nodes to create multi-hop structure.
+   * Insert 2*n_neighbors nodes so the graph is dense enough for search
+   * to actually traverse multiple hops. */
+  int seed_count = n_neighbors * 2;
+  if (seed_count > 64) seed_count = 64; /* Bound for performance */
+  {
+    float *vec = malloc(dims * sizeof(float));
+    if (!vec) goto cleanup;
+    for (int i = 1; i <= seed_count; i++) {
+      for (int j = 0; j < dims; j++) {
+        vec[j] = fuzz_float(&data, &size);
+      }
+      sqlite3_reset(stmtInsert);
+      sqlite3_bind_int64(stmtInsert, 1, i);
+      sqlite3_bind_blob(stmtInsert, 2, vec, dims * sizeof(float), SQLITE_TRANSIENT);
+      sqlite3_step(stmtInsert);
+    }
+    free(vec);
+  }
+
+  /* Phase 2: Fuzz-driven operations on the seeded graph */
+  float *vec = malloc(dims * sizeof(float));
+  if (!vec) goto cleanup;
+
+  while (size >= 2) {
+    uint8_t op = fuzz_byte(&data, &size, 0) % 5;
+    uint8_t param = fuzz_byte(&data, &size, 0);
+
+    switch (op) {
+      case 0: { /* INSERT with fuzz-controlled vector and rowid */
+        int64_t rowid = (int64_t)(param % 128) + 1;
+        for (int j = 0; j < dims; j++) {
+          vec[j] = fuzz_float(&data, &size);
+        }
+        sqlite3_reset(stmtInsert);
+        sqlite3_bind_int64(stmtInsert, 1, rowid);
+        sqlite3_bind_blob(stmtInsert, 2, vec, dims * sizeof(float), SQLITE_TRANSIENT);
+        sqlite3_step(stmtInsert);
+        break;
+      }
+      case 1: { /* DELETE */
+        int64_t rowid = (int64_t)(param % 128) + 1;
+        sqlite3_reset(stmtDelete);
+        sqlite3_bind_int64(stmtDelete, 1, rowid);
+        sqlite3_step(stmtDelete);
+        break;
+      }
+      case 2: { /* KNN with fuzz query vector and variable k */
+        for (int j = 0; j < dims; j++) {
+          vec[j] = fuzz_float(&data, &size);
+        }
+        int k = (param % 20) + 1;
+        sqlite3_reset(stmtKnn);
+        sqlite3_bind_blob(stmtKnn, 1, vec, dims * sizeof(float), SQLITE_TRANSIENT);
+        sqlite3_bind_int(stmtKnn, 2, k);
+        while (sqlite3_step(stmtKnn) == SQLITE_ROW) {}
+        break;
+      }
+      case 3: { /* KNN with k > number of nodes (boundary) */
+        for (int j = 0; j < dims; j++) {
+          vec[j] = fuzz_float(&data, &size);
+        }
+        sqlite3_reset(stmtKnn);
+        sqlite3_bind_blob(stmtKnn, 1, vec, dims * sizeof(float), SQLITE_TRANSIENT);
+        sqlite3_bind_int(stmtKnn, 2, 1000); /* k >> graph size */
+        while (sqlite3_step(stmtKnn) == SQLITE_ROW) {}
+        break;
+      }
+      case 4: { /* INSERT duplicate rowid (triggers OR REPLACE path) */
+        int64_t rowid = (int64_t)(param % 32) + 1;
+        for (int j = 0; j < dims; j++) {
+          vec[j] = (float)(param + j) / 50.0f;
+        }
+        sqlite3_reset(stmtInsert);
+        sqlite3_bind_int64(stmtInsert, 1, rowid);
+        sqlite3_bind_blob(stmtInsert, 2, vec, dims * sizeof(float), SQLITE_TRANSIENT);
+        sqlite3_step(stmtInsert);
+        break;
+      }
+    }
+  }
+  free(vec);
+
+cleanup:
+  sqlite3_finalize(stmtInsert);
+  sqlite3_finalize(stmtDelete);
+  sqlite3_finalize(stmtKnn);
+  sqlite3_close(db);
+  return 0;
+}

tests/fuzz/diskann-delete-stress.c

@@ -0,0 +1,175 @@
+/**
+ * Fuzz target for DiskANN delete path and graph connectivity maintenance.
+ *
+ * The delete path is the most complex graph mutation:
+ *   1. Read deleted node's neighbor list
+ *   2. For each neighbor, remove deleted node from their list
+ *   3. Try to fill the gap with one of deleted node's other neighbors
+ *   4. Handle medoid deletion (pick new medoid)
+ *
+ * Edge cases this targets:
+ *   - Delete the medoid (entry point) -- forces medoid reassignment
+ *   - Delete all nodes except one -- graph degenerates
+ *   - Delete nodes in a chain -- cascading dangling edges
+ *   - Re-insert at deleted rowids -- stale graph edges to old data
+ *   - Delete nonexistent rowids -- should be no-op
+ *   - Insert-delete-insert same rowid rapidly
+ *   - Delete when graph has exactly n_neighbors entries (full nodes)
+ *
+ * Key code paths:
+ *   - diskann_delete -> diskann_repair_reverse_edges
+ *   - diskann_medoid_handle_delete
+ *   - diskann_node_clear_neighbor
+ *   - Interaction between delete and concurrent search
+ */
+#include <stdint.h>
+#include <stddef.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include "sqlite-vec.h"
+#include "sqlite3.h"
+#include <assert.h>
+
+static uint8_t fuzz_byte(const uint8_t **data, size_t *size, uint8_t def) {
+  if (*size == 0) return def;
+  uint8_t b = **data;
+  (*data)++;
+  (*size)--;
+  return b;
+}
+
+int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) {
+  if (size < 20) return 0;
+
+  int rc;
+  sqlite3 *db;
+  rc = sqlite3_open(":memory:", &db);
+  assert(rc == SQLITE_OK);
+  rc = sqlite3_vec_init(db, NULL, NULL);
+  assert(rc == SQLITE_OK);
+
+  /* int8 quantizer to exercise that distance code path */
+  uint8_t quant = fuzz_byte(&data, &size, 0) % 2;
+  const char *qname = quant ? "int8" : "binary";
+
+  char sql[256];
+  snprintf(sql, sizeof(sql),
+    "CREATE VIRTUAL TABLE v USING vec0("
+    "emb float[8] INDEXED BY diskann(neighbor_quantizer=%s, n_neighbors=8))",
+    qname);
+  rc = sqlite3_exec(db, sql, NULL, NULL, NULL);
+  if (rc != SQLITE_OK) { sqlite3_close(db); return 0; }
+
+  sqlite3_stmt *stmtInsert = NULL, *stmtDelete = NULL, *stmtKnn = NULL;
+  sqlite3_prepare_v2(db,
+    "INSERT INTO v(rowid, emb) VALUES (?, ?)", -1, &stmtInsert, NULL);
+  sqlite3_prepare_v2(db,
+    "DELETE FROM v WHERE rowid = ?", -1, &stmtDelete, NULL);
+  sqlite3_prepare_v2(db,
+    "SELECT rowid, distance FROM v WHERE emb MATCH ? AND k = ?",
+    -1, &stmtKnn, NULL);
+
+  if (!stmtInsert || !stmtDelete || !stmtKnn) goto cleanup;
+
+  /* Phase 1: Build a graph of exactly n_neighbors+2 = 10 nodes.
+   * This makes every node nearly full, maximizing the chance that
+   * inserts trigger the "full node" path in add_reverse_edge. */
+  for (int i = 1; i <= 10; i++) {
+    float vec[8];
+    for (int j = 0; j < 8; j++) {
+      vec[j] = (float)((int8_t)fuzz_byte(&data, &size, (uint8_t)(i*13+j*7))) / 20.0f;
+    }
+    sqlite3_reset(stmtInsert);
+    sqlite3_bind_int64(stmtInsert, 1, i);
+    sqlite3_bind_blob(stmtInsert, 2, vec, sizeof(vec), SQLITE_TRANSIENT);
+    sqlite3_step(stmtInsert);
+  }
+
+  /* Phase 2: Fuzz-driven delete-heavy workload */
+  while (size >= 2) {
+    uint8_t op = fuzz_byte(&data, &size, 0);
+    uint8_t param = fuzz_byte(&data, &size, 0);
+
+    switch (op % 6) {
+      case 0: /* Delete existing node */
+      case 1: { /* (weighted toward deletes) */
+        int64_t rowid = (int64_t)(param % 16) + 1;
+        sqlite3_reset(stmtDelete);
+        sqlite3_bind_int64(stmtDelete, 1, rowid);
+        sqlite3_step(stmtDelete);
+        break;
+      }
+      case 2: { /* Delete then immediately re-insert same rowid */
+        int64_t rowid = (int64_t)(param % 10) + 1;
+        sqlite3_reset(stmtDelete);
+        sqlite3_bind_int64(stmtDelete, 1, rowid);
+        sqlite3_step(stmtDelete);
+
+        float vec[8];
+        for (int j = 0; j < 8; j++) {
+          vec[j] = (float)((int8_t)fuzz_byte(&data, &size, (uint8_t)(rowid+j))) / 15.0f;
+        }
+        sqlite3_reset(stmtInsert);
+        sqlite3_bind_int64(stmtInsert, 1, rowid);
+        sqlite3_bind_blob(stmtInsert, 2, vec, sizeof(vec), SQLITE_TRANSIENT);
+        sqlite3_step(stmtInsert);
+        break;
+      }
+      case 3: { /* KNN query on potentially sparse/empty graph */
+        float qvec[8];
+        for (int j = 0; j < 8; j++) {
+          qvec[j] = (float)((int8_t)fuzz_byte(&data, &size, 0)) / 10.0f;
+        }
+        int k = (param % 15) + 1;
+        sqlite3_reset(stmtKnn);
+        sqlite3_bind_blob(stmtKnn, 1, qvec, sizeof(qvec), SQLITE_TRANSIENT);
+        sqlite3_bind_int(stmtKnn, 2, k);
+        while (sqlite3_step(stmtKnn) == SQLITE_ROW) {}
+        break;
+      }
+      case 4: { /* Insert new node */
+        int64_t rowid = (int64_t)(param % 32) + 1;
+        float vec[8];
+        for (int j = 0; j < 8; j++) {
+          vec[j] = (float)((int8_t)fuzz_byte(&data, &size, 0)) / 10.0f;
+        }
+        sqlite3_reset(stmtInsert);
+        sqlite3_bind_int64(stmtInsert, 1, rowid);
+        sqlite3_bind_blob(stmtInsert, 2, vec, sizeof(vec), SQLITE_TRANSIENT);
+        sqlite3_step(stmtInsert);
+        break;
+      }
+      case 5: { /* Delete ALL remaining nodes, then insert fresh */
+        for (int i = 1; i <= 32; i++) {
+          sqlite3_reset(stmtDelete);
+          sqlite3_bind_int64(stmtDelete, 1, i);
+          sqlite3_step(stmtDelete);
+        }
+        /* Now insert one node into empty graph */
+        float vec[8] = {1.0f, 0, 0, 0, 0, 0, 0, 0};
+        sqlite3_reset(stmtInsert);
+        sqlite3_bind_int64(stmtInsert, 1, 1);
+        sqlite3_bind_blob(stmtInsert, 2, vec, sizeof(vec), SQLITE_TRANSIENT);
+        sqlite3_step(stmtInsert);
+        break;
+      }
+    }
+  }
+
+  /* Final KNN on whatever state the graph is in */
+  {
+    float qvec[8] = {0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f};
+    sqlite3_reset(stmtKnn);
+    sqlite3_bind_blob(stmtKnn, 1, qvec, sizeof(qvec), SQLITE_TRANSIENT);
+    sqlite3_bind_int(stmtKnn, 2, 10);
+    while (sqlite3_step(stmtKnn) == SQLITE_ROW) {}
+  }
+
+cleanup:
+  sqlite3_finalize(stmtInsert);
+  sqlite3_finalize(stmtDelete);
+  sqlite3_finalize(stmtKnn);
+  sqlite3_close(db);
+  return 0;
+}

tests/fuzz/diskann-graph-corrupt.c

@@ -0,0 +1,123 @@
+/**
+ * Fuzz target for DiskANN shadow table corruption resilience.
+ * Creates and populates a DiskANN table, then corrupts shadow table blobs
+ * using fuzz data and runs queries.
+ */
+#include <stdint.h>
+#include <stddef.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include "sqlite-vec.h"
+#include "sqlite3.h"
+#include <assert.h>
+
+int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) {
+  if (size < 16) return 0;
+
+  int rc;
+  sqlite3 *db;
+
+  rc = sqlite3_open(":memory:", &db);
+  assert(rc == SQLITE_OK);
+  rc = sqlite3_vec_init(db, NULL, NULL);
+  assert(rc == SQLITE_OK);
+
+  rc = sqlite3_exec(db,
+    "CREATE VIRTUAL TABLE v USING vec0("
+    "emb float[8] INDEXED BY diskann(neighbor_quantizer=binary, n_neighbors=8))",
+    NULL, NULL, NULL);
+  if (rc != SQLITE_OK) { sqlite3_close(db); return 0; }
+
+  /* Insert a few vectors to create graph structure */
+  {
+    sqlite3_stmt *stmt;
+    sqlite3_prepare_v2(db,
+      "INSERT INTO v(rowid, emb) VALUES (?, ?)", -1, &stmt, NULL);
+    for (int i = 1; i <= 10; i++) {
+      float vec[8];
+      for (int j = 0; j < 8; j++) {
+        vec[j] = (float)i * 0.1f + (float)j * 0.01f;
+      }
+      sqlite3_reset(stmt);
+      sqlite3_bind_int64(stmt, 1, i);
+      sqlite3_bind_blob(stmt, 2, vec, sizeof(vec), SQLITE_TRANSIENT);
+      sqlite3_step(stmt);
+    }
+    sqlite3_finalize(stmt);
+  }
+
+  /* Corrupt shadow table data using fuzz bytes */
+  size_t offset = 0;
+
+  /* Determine which row and column to corrupt */
+  int target_row = (data[offset++] % 10) + 1;
+  int corrupt_type = data[offset++] % 3;  /* 0=validity, 1=neighbor_ids, 2=qvecs */
+
+  const char *column_name;
+  switch (corrupt_type) {
+    case 0: column_name = "neighbors_validity"; break;
+    case 1: column_name = "neighbor_ids"; break;
+    default: column_name = "neighbor_quantized_vectors"; break;
+  }
+
+  /* Read the blob, corrupt it, write it back */
+  {
+    sqlite3_stmt *readStmt;
+    char sqlbuf[256];
+    snprintf(sqlbuf, sizeof(sqlbuf),
+      "SELECT %s FROM v_diskann_nodes00 WHERE rowid = ?", column_name);
+    rc = sqlite3_prepare_v2(db, sqlbuf, -1, &readStmt, NULL);
+    if (rc == SQLITE_OK) {
+      sqlite3_bind_int64(readStmt, 1, target_row);
+      if (sqlite3_step(readStmt) == SQLITE_ROW) {
+        const void *blob = sqlite3_column_blob(readStmt, 0);
+        int blobSize = sqlite3_column_bytes(readStmt, 0);
+        if (blob && blobSize > 0) {
+          unsigned char *corrupt = sqlite3_malloc(blobSize);
+          if (corrupt) {
+            memcpy(corrupt, blob, blobSize);
+            /* Apply fuzz bytes as XOR corruption */
+            size_t remaining = size - offset;
+            for (size_t i = 0; i < remaining && i < (size_t)blobSize; i++) {
+              corrupt[i % blobSize] ^= data[offset + i];
+            }
+            /* Write back */
+            sqlite3_stmt *writeStmt;
+            snprintf(sqlbuf, sizeof(sqlbuf),
+              "UPDATE v_diskann_nodes00 SET %s = ? WHERE rowid = ?", column_name);
+            rc = sqlite3_prepare_v2(db, sqlbuf, -1, &writeStmt, NULL);
+            if (rc == SQLITE_OK) {
+              sqlite3_bind_blob(writeStmt, 1, corrupt, blobSize, SQLITE_TRANSIENT);
+              sqlite3_bind_int64(writeStmt, 2, target_row);
+              sqlite3_step(writeStmt);
+              sqlite3_finalize(writeStmt);
+            }
+            sqlite3_free(corrupt);
+          }
+        }
+      }
+      sqlite3_finalize(readStmt);
+    }
+  }
+
+  /* Run queries on corrupted graph -- should not crash */
+  {
+    float qvec[8] = {1.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f};
+    sqlite3_stmt *knnStmt;
+    rc = sqlite3_prepare_v2(db,
+      "SELECT rowid, distance FROM v WHERE emb MATCH ? AND k = 5",
+      -1, &knnStmt, NULL);
+    if (rc == SQLITE_OK) {
+      sqlite3_bind_blob(knnStmt, 1, qvec, sizeof(qvec), SQLITE_STATIC);
+      while (sqlite3_step(knnStmt) == SQLITE_ROW) {}
+      sqlite3_finalize(knnStmt);
+    }
+  }
+
+  /* Full scan */
+  sqlite3_exec(db, "SELECT * FROM v", NULL, NULL, NULL);
+
+  sqlite3_close(db);
+  return 0;
+}

tests/fuzz/diskann-int8-quant.c

@@ -0,0 +1,164 @@
+/**
+ * Fuzz target for DiskANN int8 quantizer edge cases.
+ *
+ * The binary quantizer is simple (sign bit), but the int8 quantizer has
+ * interesting arithmetic:
+ *   i8_val = (i8)(((src - (-1.0f)) / step) - 128.0f)
+ * where step = 2.0f / 255.0f
+ *
+ * Edge cases in this formula:
+ *   - src values outside [-1, 1] cause clamping issues (no explicit clamp!)
+ *   - src = NaN, +Inf, -Inf  (from corrupted vectors or div-by-zero)
+ *   - src very close to boundaries (-1.0, 1.0) -- rounding
+ *   - The cast to i8 can overflow for extreme src values
+ *
+ * Also exercises int8 distance functions:
+ *   - distance_l2_sqr_int8: accumulates squared differences, possible overflow
+ *   - distance_cosine_int8: dot product with normalization
+ *   - distance_l1_int8: absolute differences
+ *
+ * This fuzzer also tests the cosine distance metric path which the
+ * other fuzzers (using L2 default) don't cover.
+ */
+#include <stdint.h>
+#include <stddef.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <math.h>
+#include "sqlite-vec.h"
+#include "sqlite3.h"
+#include <assert.h>
+
+static uint8_t fuzz_byte(const uint8_t **data, size_t *size, uint8_t def) {
+  if (*size == 0) return def;
+  uint8_t b = **data;
+  (*data)++;
+  (*size)--;
+  return b;
+}
+
+static float fuzz_extreme_float(const uint8_t **data, size_t *size) {
+  uint8_t mode = fuzz_byte(data, size, 0) % 8;
+  uint8_t raw = fuzz_byte(data, size, 0);
+  switch (mode) {
+    case 0: return (float)((int8_t)raw) / 10.0f;   /* Normal range */
+    case 1: return (float)((int8_t)raw) * 100.0f;   /* Large values */
+    case 2: return (float)((int8_t)raw) / 1000.0f;  /* Tiny values near 0 */
+    case 3: return -1.0f;                             /* Exact boundary */
+    case 4: return 1.0f;                              /* Exact boundary */
+    case 5: return 0.0f;                              /* Zero */
+    case 6: return (float)raw / 255.0f;              /* [0, 1] range */
+    case 7: return -(float)raw / 255.0f;             /* [-1, 0] range */
+  }
+  return 0.0f;
+}
+
+int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) {
+  if (size < 40) return 0;
+
+  int rc;
+  sqlite3 *db;
+  rc = sqlite3_open(":memory:", &db);
+  assert(rc == SQLITE_OK);
+  rc = sqlite3_vec_init(db, NULL, NULL);
+  assert(rc == SQLITE_OK);
+
+  /* Test both distance metrics with int8 quantizer */
+  uint8_t metric_choice = fuzz_byte(&data, &size, 0) % 2;
+  const char *metric = metric_choice ? "cosine" : "L2";
+
+  int dims = 8 + (fuzz_byte(&data, &size, 0) % 3) * 8; /* 8, 16, or 24 */
+
+  char sql[512];
+  snprintf(sql, sizeof(sql),
+    "CREATE VIRTUAL TABLE v USING vec0("
+    "emb float[%d] distance_metric=%s "
+    "INDEXED BY diskann(neighbor_quantizer=int8, n_neighbors=8, search_list_size=16))",
+    dims, metric);
+
+  rc = sqlite3_exec(db, sql, NULL, NULL, NULL);
+  if (rc != SQLITE_OK) { sqlite3_close(db); return 0; }
+
+  sqlite3_stmt *stmtInsert = NULL, *stmtKnn = NULL, *stmtDelete = NULL;
+  sqlite3_prepare_v2(db,
+    "INSERT INTO v(rowid, emb) VALUES (?, ?)", -1, &stmtInsert, NULL);
+  sqlite3_prepare_v2(db,
+    "SELECT rowid, distance FROM v WHERE emb MATCH ? AND k = ?",
+    -1, &stmtKnn, NULL);
+  sqlite3_prepare_v2(db,
+    "DELETE FROM v WHERE rowid = ?", -1, &stmtDelete, NULL);
+
+  if (!stmtInsert || !stmtKnn || !stmtDelete) goto cleanup;
+
+  /* Insert vectors with extreme float values to stress quantization */
+  float *vec = malloc(dims * sizeof(float));
+  if (!vec) goto cleanup;
+
+  for (int i = 1; i <= 16; i++) {
+    for (int j = 0; j < dims; j++) {
+      vec[j] = fuzz_extreme_float(&data, &size);
+    }
+    sqlite3_reset(stmtInsert);
+    sqlite3_bind_int64(stmtInsert, 1, i);
+    sqlite3_bind_blob(stmtInsert, 2, vec, dims * sizeof(float), SQLITE_TRANSIENT);
+    sqlite3_step(stmtInsert);
+  }
+
+  /* Fuzz-driven operations */
+  while (size >= 2) {
+    uint8_t op = fuzz_byte(&data, &size, 0) % 4;
+    uint8_t param = fuzz_byte(&data, &size, 0);
+
+    switch (op) {
+      case 0: { /* KNN with extreme query values */
+        for (int j = 0; j < dims; j++) {
+          vec[j] = fuzz_extreme_float(&data, &size);
+        }
+        int k = (param % 10) + 1;
+        sqlite3_reset(stmtKnn);
+        sqlite3_bind_blob(stmtKnn, 1, vec, dims * sizeof(float), SQLITE_TRANSIENT);
+        sqlite3_bind_int(stmtKnn, 2, k);
+        while (sqlite3_step(stmtKnn) == SQLITE_ROW) {}
+        break;
+      }
+      case 1: { /* Insert with extreme values */
+        int64_t rowid = (int64_t)(param % 32) + 1;
+        for (int j = 0; j < dims; j++) {
+          vec[j] = fuzz_extreme_float(&data, &size);
+        }
+        sqlite3_reset(stmtInsert);
+        sqlite3_bind_int64(stmtInsert, 1, rowid);
+        sqlite3_bind_blob(stmtInsert, 2, vec, dims * sizeof(float), SQLITE_TRANSIENT);
+        sqlite3_step(stmtInsert);
+        break;
+      }
+      case 2: { /* Delete */
+        int64_t rowid = (int64_t)(param % 32) + 1;
+        sqlite3_reset(stmtDelete);
+        sqlite3_bind_int64(stmtDelete, 1, rowid);
+        sqlite3_step(stmtDelete);
+        break;
+      }
+      case 3: { /* KNN with all-zero or all-same-value query */
+        float val = (param % 3 == 0) ? 0.0f :
+                    (param % 3 == 1) ? 1.0f : -1.0f;
+        for (int j = 0; j < dims; j++) vec[j] = val;
+        sqlite3_reset(stmtKnn);
+        sqlite3_bind_blob(stmtKnn, 1, vec, dims * sizeof(float), SQLITE_TRANSIENT);
+        sqlite3_bind_int(stmtKnn, 2, 5);
+        while (sqlite3_step(stmtKnn) == SQLITE_ROW) {}
+        break;
+      }
+    }
+  }
+
+  free(vec);
+
+cleanup:
+  sqlite3_finalize(stmtInsert);
+  sqlite3_finalize(stmtKnn);
+  sqlite3_finalize(stmtDelete);
+  sqlite3_close(db);
+  return 0;
+}

tests/fuzz/diskann-operations.c

@@ -0,0 +1,100 @@
+/**
+ * Fuzz target for DiskANN insert/delete/query operation sequences.
+ * Uses fuzz bytes to drive random operations on a DiskANN-indexed table.
+ */
+#include <stdint.h>
+#include <stddef.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include "sqlite-vec.h"
+#include "sqlite3.h"
+#include <assert.h>
+
+int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) {
+  if (size < 6) return 0;
+
+  int rc;
+  sqlite3 *db;
+  sqlite3_stmt *stmtInsert = NULL;
+  sqlite3_stmt *stmtDelete = NULL;
+  sqlite3_stmt *stmtKnn = NULL;
+  sqlite3_stmt *stmtScan = NULL;
+
+  rc = sqlite3_open(":memory:", &db);
+  assert(rc == SQLITE_OK);
+  rc = sqlite3_vec_init(db, NULL, NULL);
+  assert(rc == SQLITE_OK);
+
+  rc = sqlite3_exec(db,
+    "CREATE VIRTUAL TABLE v USING vec0("
+    "emb float[8] INDEXED BY diskann(neighbor_quantizer=binary, n_neighbors=8))",
+    NULL, NULL, NULL);
+  if (rc != SQLITE_OK) { sqlite3_close(db); return 0; }
+
+  sqlite3_prepare_v2(db,
+    "INSERT INTO v(rowid, emb) VALUES (?, ?)", -1, &stmtInsert, NULL);
+  sqlite3_prepare_v2(db,
+    "DELETE FROM v WHERE rowid = ?", -1, &stmtDelete, NULL);
+  sqlite3_prepare_v2(db,
+    "SELECT rowid, distance FROM v WHERE emb MATCH ? AND k = 3",
+    -1, &stmtKnn, NULL);
+  sqlite3_prepare_v2(db,
+    "SELECT rowid FROM v", -1, &stmtScan, NULL);
+
+  if (!stmtInsert || !stmtDelete || !stmtKnn || !stmtScan) goto cleanup;
+
+  size_t i = 0;
+  while (i + 2 <= size) {
+    uint8_t op = data[i++] % 4;
+    uint8_t rowid_byte = data[i++];
+    int64_t rowid = (int64_t)(rowid_byte % 32) + 1;
+
+    switch (op) {
+      case 0: {
+        /* INSERT: consume 32 bytes for 8 floats, or use what's left */
+        float vec[8] = {0};
+        for (int j = 0; j < 8 && i < size; j++, i++) {
+          vec[j] = (float)((int8_t)data[i]) / 10.0f;
+        }
+        sqlite3_reset(stmtInsert);
+        sqlite3_bind_int64(stmtInsert, 1, rowid);
+        sqlite3_bind_blob(stmtInsert, 2, vec, sizeof(vec), SQLITE_TRANSIENT);
+        sqlite3_step(stmtInsert);
+        break;
+      }
+      case 1: {
+        /* DELETE */
+        sqlite3_reset(stmtDelete);
+        sqlite3_bind_int64(stmtDelete, 1, rowid);
+        sqlite3_step(stmtDelete);
+        break;
+      }
+      case 2: {
+        /* KNN query */
+        float qvec[8] = {1.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f};
+        sqlite3_reset(stmtKnn);
+        sqlite3_bind_blob(stmtKnn, 1, qvec, sizeof(qvec), SQLITE_STATIC);
+        while (sqlite3_step(stmtKnn) == SQLITE_ROW) {}
+        break;
+      }
+      case 3: {
+        /* Full scan */
+        sqlite3_reset(stmtScan);
+        while (sqlite3_step(stmtScan) == SQLITE_ROW) {}
+        break;
+      }
+    }
+  }
+
+  /* Final operations -- must not crash regardless of prior state */
+  sqlite3_exec(db, "SELECT * FROM v", NULL, NULL, NULL);
+
+cleanup:
+  sqlite3_finalize(stmtInsert);
+  sqlite3_finalize(stmtDelete);
+  sqlite3_finalize(stmtKnn);
+  sqlite3_finalize(stmtScan);
+  sqlite3_close(db);
+  return 0;
+}

tests/fuzz/diskann-prune-direct.c

@@ -0,0 +1,131 @@
+/**
+ * Fuzz target for DiskANN RobustPrune algorithm (diskann_prune_select).
+ *
+ * diskann_prune_select is exposed for testing and takes:
+ *   - inter_distances: flattened NxN matrix of inter-candidate distances
+ *   - p_distances: N distances from node p to each candidate
+ *   - num_candidates, alpha, max_neighbors
+ *
+ * This is a pure function that doesn't need a database, so we can
+ * call it directly with fuzz-controlled inputs. This gives the fuzzer
+ * maximum speed (no SQLite overhead) to explore:
+ *
+ *   - alpha boundary: alpha=0 (prunes nothing), alpha=very large (prunes all)
+ *   - max_neighbors = 0, 1, num_candidates, > num_candidates
+ *   - num_candidates = 0, 1, large
+ *   - Distance matrices with: all zeros, all same, negative values, NaN, Inf
+ *   - Non-symmetric distance matrices (should still work)
+ *   - Memory: large num_candidates to stress malloc
+ *
+ * Key code paths:
+ *   - diskann_prune_select alpha-pruning loop
+ *   - Boundary: selectedCount reaches max_neighbors exactly
+ *   - All candidates pruned before max_neighbors reached
+ */
+#include <stdint.h>
+#include <stddef.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <math.h>
+#include "sqlite-vec.h"
+#include "sqlite3.h"
+#include <assert.h>
+
+/* Declare the test-exposed function.
+ * diskann_prune_select is not static -- it's a public symbol. */
+extern int diskann_prune_select(
+    const float *inter_distances, const float *p_distances,
+    int num_candidates, float alpha, int max_neighbors,
+    int *outSelected, int *outCount);
+
+static uint8_t fuzz_byte(const uint8_t **data, size_t *size, uint8_t def) {
+  if (*size == 0) return def;
+  uint8_t b = **data;
+  (*data)++;
+  (*size)--;
+  return b;
+}
+
+int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) {
+  if (size < 8) return 0;
+
+  /* Consume parameters from fuzz data */
+  int num_candidates = fuzz_byte(&data, &size, 0) % 33; /* 0..32 */
+  int max_neighbors = fuzz_byte(&data, &size, 0) % 17;  /* 0..16 */
+
+  /* Alpha: pick from interesting values */
+  uint8_t alpha_idx = fuzz_byte(&data, &size, 0) % 8;
+  float alpha_values[] = {0.0f, 0.5f, 1.0f, 1.2f, 1.5f, 2.0f, 10.0f, 100.0f};
+  float alpha = alpha_values[alpha_idx];
+
+  if (num_candidates == 0) {
+    /* Test empty case */
+    int outCount = -1;
+    int rc = diskann_prune_select(NULL, NULL, 0, alpha, max_neighbors,
+                                   NULL, &outCount);
+    assert(rc == 0 /* SQLITE_OK */);
+    assert(outCount == 0);
+    return 0;
+  }
+
+  /* Allocate arrays */
+  int n = num_candidates;
+  float *inter_distances = malloc(n * n * sizeof(float));
+  float *p_distances = malloc(n * sizeof(float));
+  int *outSelected = malloc(n * sizeof(int));
+  if (!inter_distances || !p_distances || !outSelected) {
+    free(inter_distances);
+    free(p_distances);
+    free(outSelected);
+    return 0;
+  }
+
+  /* Fill p_distances from fuzz data (sorted ascending for correct input) */
+  for (int i = 0; i < n; i++) {
+    uint8_t raw = fuzz_byte(&data, &size, (uint8_t)(i * 10));
+    p_distances[i] = (float)raw / 10.0f;
+  }
+  /* Sort p_distances ascending (prune_select expects sorted input) */
+  for (int i = 1; i < n; i++) {
+    float tmp = p_distances[i];
+    int j = i - 1;
+    while (j >= 0 && p_distances[j] > tmp) {
+      p_distances[j + 1] = p_distances[j];
+      j--;
+    }
+    p_distances[j + 1] = tmp;
+  }
+
+  /* Fill inter-distance matrix from fuzz data */
+  for (int i = 0; i < n * n; i++) {
+    uint8_t raw = fuzz_byte(&data, &size, (uint8_t)(i % 256));
+    inter_distances[i] = (float)raw / 10.0f;
+  }
+  /* Make diagonal zero */
+  for (int i = 0; i < n; i++) {
+    inter_distances[i * n + i] = 0.0f;
+  }
+
+  int outCount = -1;
+  int rc = diskann_prune_select(inter_distances, p_distances,
+                                 n, alpha, max_neighbors,
+                                 outSelected, &outCount);
+  /* Basic sanity: should not crash, count should be valid */
+  assert(rc == 0);
+  assert(outCount >= 0);
+  assert(outCount <= max_neighbors || max_neighbors == 0);
+  assert(outCount <= n);
+
+  /* Verify outSelected flags are consistent with outCount */
+  int flagCount = 0;
+  for (int i = 0; i < n; i++) {
+    if (outSelected[i]) flagCount++;
+  }
+  assert(flagCount == outCount);
+
+  free(inter_distances);
+  free(p_distances);
+  free(outSelected);
+  return 0;
+}

tests/fuzz/diskann.dict

@@ -0,0 +1,10 @@
+"neighbor_quantizer"
+"binary"
+"int8"
+"n_neighbors"
+"search_list_size"
+"search_list_size_search"
+"search_list_size_insert"
+"alpha"
+"="
+","

tests/fuzz/ivf-cell-overflow.c

@@ -0,0 +1,192 @@
+/**
+ * Fuzz target: IVF cell overflow and boundary conditions.
+ *
+ * Pushes cells past VEC0_IVF_CELL_MAX_VECTORS (64) to trigger cell
+ * splitting, then exercises blob I/O at slot boundaries.
+ *
+ * Targets:
+ * - Cell splitting when n_vectors reaches cap (64)
+ * - Blob offset arithmetic: slot * vecSize, slot / 8, slot % 8
+ * - Validity bitmap at byte boundaries (slot 7->8, 15->16, etc.)
+ * - Insert into full cell -> create new cell path
+ * - Delete from various slot positions (first, last, middle)
+ * - Multiple cells per centroid
+ * - assign-vectors command with multi-cell centroids
+ */
+#include <stdint.h>
+#include <stddef.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include "sqlite-vec.h"
+#include "sqlite3.h"
+#include <assert.h>
+
+int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) {
+  if (size < 8) return 0;
+
+  int rc;
+  sqlite3 *db;
+
+  rc = sqlite3_open(":memory:", &db);
+  assert(rc == SQLITE_OK);
+  rc = sqlite3_vec_init(db, NULL, NULL);
+  assert(rc == SQLITE_OK);
+
+  // Use small dimensions for speed but enough vectors to overflow cells
+  int dim = (data[0] % 8) + 2;  // 2..9
+  int nlist = (data[1] % 4) + 1;  // 1..4
+  // We need >64 vectors to overflow a cell
+  int num_vecs = (data[2] % 64) + 65;  // 65..128
+  int delete_pattern = data[3];  // Controls which vectors to delete
+
+  const uint8_t *payload = data + 4;
+  size_t payload_size = size - 4;
+
+  char sql[256];
+  snprintf(sql, sizeof(sql),
+    "CREATE VIRTUAL TABLE v USING vec0("
+    "emb float[%d] indexed by ivf(nlist=%d, nprobe=%d))",
+    dim, nlist, nlist);
+
+  rc = sqlite3_exec(db, sql, NULL, NULL, NULL);
+  if (rc != SQLITE_OK) { sqlite3_close(db); return 0; }
+
+  // Insert enough vectors to overflow at least one cell
+  sqlite3_stmt *stmtInsert = NULL;
+  sqlite3_prepare_v2(db,
+    "INSERT INTO v(v, emb) VALUES (?, ?)", -1, &stmtInsert, NULL);
+  if (!stmtInsert) { sqlite3_close(db); return 0; }
+
+  size_t offset = 0;
+  for (int i = 0; i < num_vecs; i++) {
+    float *vec = sqlite3_malloc(dim * sizeof(float));
+    if (!vec) break;
+    for (int d = 0; d < dim; d++) {
+      if (offset < payload_size) {
+        vec[d] = ((float)(int8_t)payload[offset++]) / 50.0f;
+      } else {
+        // Cluster vectors near specific centroids to ensure some cells overflow
+        int cluster = i % nlist;
+        vec[d] = (float)cluster + (float)(i % 10) * 0.01f + d * 0.001f;
+      }
+    }
+    sqlite3_reset(stmtInsert);
+    sqlite3_bind_int64(stmtInsert, 1, (int64_t)(i + 1));
+    sqlite3_bind_blob(stmtInsert, 2, vec, dim * sizeof(float), SQLITE_TRANSIENT);
+    sqlite3_step(stmtInsert);
+    sqlite3_free(vec);
+  }
+  sqlite3_finalize(stmtInsert);
+
+  // Train to assign vectors to centroids (triggers cell building)
+  sqlite3_exec(db,
+    "INSERT INTO v(v) VALUES ('compute-centroids')",
+    NULL, NULL, NULL);
+
+  // Delete vectors at boundary positions based on fuzz data
+  // This tests validity bitmap manipulation at different slot positions
+  for (int i = 0; i < num_vecs; i++) {
+    int byte_idx = i / 8;
+    if (byte_idx < (int)payload_size && (payload[byte_idx] & (1 << (i % 8)))) {
+      // Use delete_pattern to thin deletions
+      if ((delete_pattern + i) % 3 == 0) {
+        char delsql[64];
+        snprintf(delsql, sizeof(delsql), "DELETE FROM v WHERE rowid = %d", i + 1);
+        sqlite3_exec(db, delsql, NULL, NULL, NULL);
+      }
+    }
+  }
+
+  // Insert more vectors after deletions (into cells with holes)
+  {
+    sqlite3_stmt *si = NULL;
+    sqlite3_prepare_v2(db,
+      "INSERT INTO v(v, emb) VALUES (?, ?)", -1, &si, NULL);
+    if (si) {
+      for (int i = 0; i < 10; i++) {
+        float *vec = sqlite3_malloc(dim * sizeof(float));
+        if (!vec) break;
+        for (int d = 0; d < dim; d++)
+          vec[d] = (float)(i + 200) * 0.01f;
+        sqlite3_reset(si);
+        sqlite3_bind_int64(si, 1, (int64_t)(num_vecs + i + 1));
+        sqlite3_bind_blob(si, 2, vec, dim * sizeof(float), SQLITE_TRANSIENT);
+        sqlite3_step(si);
+        sqlite3_free(vec);
+      }
+      sqlite3_finalize(si);
+    }
+  }
+
+  // KNN query that must scan multiple cells per centroid
+  {
+    float *qvec = sqlite3_malloc(dim * sizeof(float));
+    if (qvec) {
+      for (int d = 0; d < dim; d++) qvec[d] = 0.0f;
+      sqlite3_stmt *sk = NULL;
+      snprintf(sql, sizeof(sql),
+        "SELECT rowid, distance FROM v WHERE emb MATCH ? LIMIT 20");
+      sqlite3_prepare_v2(db, sql, -1, &sk, NULL);
+      if (sk) {
+        sqlite3_bind_blob(sk, 1, qvec, dim * sizeof(float), SQLITE_TRANSIENT);
+        while (sqlite3_step(sk) == SQLITE_ROW) {}
+        sqlite3_finalize(sk);
+      }
+      sqlite3_free(qvec);
+    }
+  }
+
+  // Test assign-vectors with multi-cell state
+  // First clear centroids
+  sqlite3_exec(db,
+    "INSERT INTO v(v) VALUES ('clear-centroids')",
+    NULL, NULL, NULL);
+
+  // Set centroids manually, then assign
+  for (int c = 0; c < nlist; c++) {
+    float *cvec = sqlite3_malloc(dim * sizeof(float));
+    if (!cvec) break;
+    for (int d = 0; d < dim; d++) cvec[d] = (float)c + d * 0.1f;
+
+    char cmd[128];
+    snprintf(cmd, sizeof(cmd),
+      "INSERT INTO v(v, emb) VALUES ('set-centroid:%d', ?)", c);
+    sqlite3_stmt *sc = NULL;
+    sqlite3_prepare_v2(db, cmd, -1, &sc, NULL);
+    if (sc) {
+      sqlite3_bind_blob(sc, 1, cvec, dim * sizeof(float), SQLITE_TRANSIENT);
+      sqlite3_step(sc);
+      sqlite3_finalize(sc);
+    }
+    sqlite3_free(cvec);
+  }
+
+  sqlite3_exec(db,
+    "INSERT INTO v(v) VALUES ('assign-vectors')",
+    NULL, NULL, NULL);
+
+  // Final query after assign-vectors
+  {
+    float *qvec = sqlite3_malloc(dim * sizeof(float));
+    if (qvec) {
+      for (int d = 0; d < dim; d++) qvec[d] = 1.0f;
+      sqlite3_stmt *sk = NULL;
+      sqlite3_prepare_v2(db,
+        "SELECT rowid, distance FROM v WHERE emb MATCH ? LIMIT 5",
+        -1, &sk, NULL);
+      if (sk) {
+        sqlite3_bind_blob(sk, 1, qvec, dim * sizeof(float), SQLITE_TRANSIENT);
+        while (sqlite3_step(sk) == SQLITE_ROW) {}
+        sqlite3_finalize(sk);
+      }
+      sqlite3_free(qvec);
+    }
+  }
+
+  // Full scan
+  sqlite3_exec(db, "SELECT * FROM v", NULL, NULL, NULL);
+
+  sqlite3_close(db);
+  return 0;
+}

tests/fuzz/ivf-create.c

@@ -0,0 +1,36 @@
+#include <stdint.h>
+#include <stddef.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include "sqlite-vec.h"
+#include "sqlite3.h"
+#include <assert.h>
+
+int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) {
+  int rc;
+  sqlite3 *db;
+  sqlite3_stmt *stmt;
+
+  rc = sqlite3_open(":memory:", &db);
+  assert(rc == SQLITE_OK);
+  rc = sqlite3_vec_init(db, NULL, NULL);
+  assert(rc == SQLITE_OK);
+
+  sqlite3_str *s = sqlite3_str_new(NULL);
+  assert(s);
+  sqlite3_str_appendall(s, "CREATE VIRTUAL TABLE v USING vec0(emb float[4] indexed by ivf(");
+  sqlite3_str_appendf(s, "%.*s", (int)size, data);
+  sqlite3_str_appendall(s, "))");
+  const char *zSql = sqlite3_str_finish(s);
+  assert(zSql);
+
+  rc = sqlite3_prepare_v2(db, zSql, -1, &stmt, NULL);
+  sqlite3_free((void *)zSql);
+  if (rc == SQLITE_OK) {
+    sqlite3_step(stmt);
+  }
+  sqlite3_finalize(stmt);
+  sqlite3_close(db);
+  return 0;
+}

tests/fuzz/ivf-create.dict

@@ -0,0 +1,16 @@
+"nlist"
+"nprobe"
+"quantizer"
+"oversample"
+"binary"
+"int8"
+"none"
+"="
+","
+"("
+")"
+"0"
+"1"
+"128"
+"65536"
+"65537"

tests/fuzz/ivf-kmeans.c

@@ -0,0 +1,180 @@
+/**
+ * Fuzz target: IVF k-means clustering.
+ *
+ * Builds a table, inserts fuzz-controlled vectors, then runs
+ * compute-centroids with fuzz-controlled parameters (nlist, max_iter, seed).
+ * Targets:
+ * - kmeans with N < k (clamping), N == 1, k == 1
+ * - kmeans with duplicate/identical vectors (all distances zero)
+ * - kmeans with NaN/Inf vectors
+ * - Empty cluster reassignment path (farthest-point heuristic)
+ * - Large nlist relative to N
+ * - The compute-centroids:{json} command parsing
+ * - clear-centroids followed by compute-centroids (round-trip)
+ */
+#include <stdint.h>
+#include <stddef.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include "sqlite-vec.h"
+#include "sqlite3.h"
+#include <assert.h>
+
+int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) {
+  if (size < 10) return 0;
+
+  int rc;
+  sqlite3 *db;
+
+  rc = sqlite3_open(":memory:", &db);
+  assert(rc == SQLITE_OK);
+  rc = sqlite3_vec_init(db, NULL, NULL);
+  assert(rc == SQLITE_OK);
+
+  // Parse fuzz header
+  // Byte 0-1: dimension (1..128)
+  // Byte 2: nlist for CREATE (1..64)
+  // Byte 3: nlist override for compute-centroids (0 = use default)
+  // Byte 4: max_iter (1..50)
+  // Byte 5-8: seed
+  // Byte 9: num_vectors (1..64)
+  // Remaining: vector float data
+
+  int dim = (data[0] | (data[1] << 8)) % 128 + 1;
+  int nlist_create = (data[2] % 64) + 1;
+  int nlist_override = data[3] % 65;  // 0 means use table default
+  int max_iter = (data[4] % 50) + 1;
+  uint32_t seed = (uint32_t)data[5] | ((uint32_t)data[6] << 8) |
+                  ((uint32_t)data[7] << 16) | ((uint32_t)data[8] << 24);
+  int num_vecs = (data[9] % 64) + 1;
+
+  const uint8_t *payload = data + 10;
+  size_t payload_size = size - 10;
+
+  char sql[256];
+  snprintf(sql, sizeof(sql),
+    "CREATE VIRTUAL TABLE v USING vec0("
+    "emb float[%d] indexed by ivf(nlist=%d, nprobe=%d))",
+    dim, nlist_create, nlist_create);
+
+  rc = sqlite3_exec(db, sql, NULL, NULL, NULL);
+  if (rc != SQLITE_OK) { sqlite3_close(db); return 0; }
+
+  // Insert vectors
+  sqlite3_stmt *stmtInsert = NULL;
+  sqlite3_prepare_v2(db,
+    "INSERT INTO v(v, emb) VALUES (?, ?)", -1, &stmtInsert, NULL);
+  if (!stmtInsert) { sqlite3_close(db); return 0; }
+
+  size_t offset = 0;
+  for (int i = 0; i < num_vecs; i++) {
+    float *vec = sqlite3_malloc(dim * sizeof(float));
+    if (!vec) break;
+
+    for (int d = 0; d < dim; d++) {
+      if (offset + 4 <= payload_size) {
+        memcpy(&vec[d], payload + offset, sizeof(float));
+        offset += 4;
+      } else if (offset < payload_size) {
+        // Scale to interesting range including values > 1, < -1
+        vec[d] = ((float)(int8_t)payload[offset++]) / 5.0f;
+      } else {
+        // Reuse earlier bytes to fill remaining dimensions
+        vec[d] = (float)(i * dim + d) * 0.01f;
+      }
+    }
+
+    sqlite3_reset(stmtInsert);
+    sqlite3_bind_int64(stmtInsert, 1, (int64_t)(i + 1));
+    sqlite3_bind_blob(stmtInsert, 2, vec, dim * sizeof(float), SQLITE_TRANSIENT);
+    sqlite3_step(stmtInsert);
+    sqlite3_free(vec);
+  }
+  sqlite3_finalize(stmtInsert);
+
+  // Exercise compute-centroids with JSON options
+  {
+    char cmd[256];
+    snprintf(cmd, sizeof(cmd),
+      "INSERT INTO v(rowid) VALUES "
+      "('compute-centroids:{\"nlist\":%d,\"max_iterations\":%d,\"seed\":%u}')",
+      nlist_override, max_iter, seed);
+    sqlite3_exec(db, cmd, NULL, NULL, NULL);
+  }
+
+  // KNN query after training
+  {
+    float *qvec = sqlite3_malloc(dim * sizeof(float));
+    if (qvec) {
+      for (int d = 0; d < dim; d++) {
+        qvec[d] = (d < 3) ? 1.0f : 0.0f;
+      }
+      sqlite3_stmt *stmtKnn = NULL;
+      sqlite3_prepare_v2(db,
+        "SELECT rowid, distance FROM v WHERE emb MATCH ? LIMIT 5",
+        -1, &stmtKnn, NULL);
+      if (stmtKnn) {
+        sqlite3_bind_blob(stmtKnn, 1, qvec, dim * sizeof(float), SQLITE_TRANSIENT);
+        while (sqlite3_step(stmtKnn) == SQLITE_ROW) {}
+        sqlite3_finalize(stmtKnn);
+      }
+      sqlite3_free(qvec);
+    }
+  }
+
+  // Clear centroids and re-compute to test round-trip
+  sqlite3_exec(db,
+    "INSERT INTO v(v) VALUES ('clear-centroids')",
+    NULL, NULL, NULL);
+
+  // Insert a few more vectors in untrained state
+  {
+    sqlite3_stmt *si = NULL;
+    sqlite3_prepare_v2(db,
+      "INSERT INTO v(v, emb) VALUES (?, ?)", -1, &si, NULL);
+    if (si) {
+      for (int i = 0; i < 3; i++) {
+        float *vec = sqlite3_malloc(dim * sizeof(float));
+        if (!vec) break;
+        for (int d = 0; d < dim; d++) vec[d] = (float)(i + 100) * 0.1f;
+        sqlite3_reset(si);
+        sqlite3_bind_int64(si, 1, (int64_t)(num_vecs + i + 1));
+        sqlite3_bind_blob(si, 2, vec, dim * sizeof(float), SQLITE_TRANSIENT);
+        sqlite3_step(si);
+        sqlite3_free(vec);
+      }
+      sqlite3_finalize(si);
+    }
+  }
+
+  // Re-train
+  sqlite3_exec(db,
+    "INSERT INTO v(v) VALUES ('compute-centroids')",
+    NULL, NULL, NULL);
+
+  // Delete some rows after training, then query
+  sqlite3_exec(db, "DELETE FROM v WHERE rowid = 1", NULL, NULL, NULL);
+  sqlite3_exec(db, "DELETE FROM v WHERE rowid = 2", NULL, NULL, NULL);
+
+  // Query after deletes
+  {
+    float *qvec = sqlite3_malloc(dim * sizeof(float));
+    if (qvec) {
+      for (int d = 0; d < dim; d++) qvec[d] = 0.5f;
+      sqlite3_stmt *stmtKnn = NULL;
+      sqlite3_prepare_v2(db,
+        "SELECT rowid, distance FROM v WHERE emb MATCH ? LIMIT 10",
+        -1, &stmtKnn, NULL);
+      if (stmtKnn) {
+        sqlite3_bind_blob(stmtKnn, 1, qvec, dim * sizeof(float), SQLITE_TRANSIENT);
+        while (sqlite3_step(stmtKnn) == SQLITE_ROW) {}
+        sqlite3_finalize(stmtKnn);
+      }
+      sqlite3_free(qvec);
+    }
+  }
+
+  sqlite3_close(db);
+  return 0;
+}

tests/fuzz/ivf-knn-deep.c

@@ -0,0 +1,199 @@
+/**
+ * Fuzz target: IVF KNN search deep paths.
+ *
+ * Exercises the full KNN pipeline with fuzz-controlled:
+ * - nprobe values (including > nlist, =1, =nlist)
+ * - Query vectors (including adversarial floats)
+ * - Mix of trained/untrained state
+ * - Oversample + rescore path (quantizer=int8 with oversample>1)
+ * - Multiple interleaved KNN queries
+ * - Candidate array realloc path (many vectors in probed cells)
+ *
+ * Targets:
+ * - ivf_scan_cells_from_stmt: candidate realloc, distance computation
+ * - ivf_query_knn: centroid sorting, nprobe selection
+ * - Oversample rescore: re-ranking with full-precision vectors
+ * - qsort with NaN distances
+ */
+#include <stdint.h>
+#include <stddef.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include "sqlite-vec.h"
+#include "sqlite3.h"
+#include <assert.h>
+
+static uint16_t read_u16(const uint8_t *p) {
+  return (uint16_t)(p[0] | (p[1] << 8));
+}
+
+int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) {
+  if (size < 16) return 0;
+
+  int rc;
+  sqlite3 *db;
+
+  rc = sqlite3_open(":memory:", &db);
+  assert(rc == SQLITE_OK);
+  rc = sqlite3_vec_init(db, NULL, NULL);
+  assert(rc == SQLITE_OK);
+
+  // Header
+  int dim = (data[0] % 32) + 2;  // 2..33
+  int nlist = (data[1] % 16) + 1;  // 1..16
+  int nprobe_initial = (data[2] % 20) + 1;  // 1..20 (can be > nlist)
+  int quantizer_type = data[3] % 3;  // 0=none, 1=int8, 2=binary
+  int oversample = (data[4] % 4) + 1;  // 1..4
+  int num_vecs = (data[5] % 80) + 4;  // 4..83
+  int num_queries = (data[6] % 8) + 1;  // 1..8
+  int k_limit = (data[7] % 20) + 1;  // 1..20
+
+  const uint8_t *payload = data + 8;
+  size_t payload_size = size - 8;
+
+  // For binary quantizer, dimension must be multiple of 8
+  if (quantizer_type == 2) {
+    dim = ((dim + 7) / 8) * 8;
+    if (dim == 0) dim = 8;
+  }
+
+  const char *qname;
+  switch (quantizer_type) {
+    case 1: qname = "int8"; break;
+    case 2: qname = "binary"; break;
+    default: qname = "none"; break;
+  }
+
+  // Oversample only valid with quantization
+  if (quantizer_type == 0) oversample = 1;
+
+  // Cap nprobe to nlist for CREATE (parser rejects nprobe > nlist)
+  int nprobe_create = nprobe_initial <= nlist ? nprobe_initial : nlist;
+
+  char sql[512];
+  snprintf(sql, sizeof(sql),
+    "CREATE VIRTUAL TABLE v USING vec0("
+    "emb float[%d] indexed by ivf(nlist=%d, nprobe=%d, quantizer=%s%s))",
+    dim, nlist, nprobe_create, qname,
+    oversample > 1 ? ", oversample=2" : "");
+
+  // If that fails (e.g. oversample with none), try without oversample
+  rc = sqlite3_exec(db, sql, NULL, NULL, NULL);
+  if (rc != SQLITE_OK) {
+    snprintf(sql, sizeof(sql),
+      "CREATE VIRTUAL TABLE v USING vec0("
+      "emb float[%d] indexed by ivf(nlist=%d, nprobe=%d, quantizer=%s))",
+      dim, nlist, nprobe_create, qname);
+    rc = sqlite3_exec(db, sql, NULL, NULL, NULL);
+    if (rc != SQLITE_OK) { sqlite3_close(db); return 0; }
+  }
+
+  // Insert vectors
+  sqlite3_stmt *stmtInsert = NULL;
+  sqlite3_prepare_v2(db,
+    "INSERT INTO v(v, emb) VALUES (?, ?)", -1, &stmtInsert, NULL);
+  if (!stmtInsert) { sqlite3_close(db); return 0; }
+
+  size_t offset = 0;
+  for (int i = 0; i < num_vecs; i++) {
+    float *vec = sqlite3_malloc(dim * sizeof(float));
+    if (!vec) break;
+    for (int d = 0; d < dim; d++) {
+      if (offset < payload_size) {
+        vec[d] = ((float)(int8_t)payload[offset++]) / 20.0f;
+      } else {
+        vec[d] = (float)((i * dim + d) % 256 - 128) / 128.0f;
+      }
+    }
+    sqlite3_reset(stmtInsert);
+    sqlite3_bind_int64(stmtInsert, 1, (int64_t)(i + 1));
+    sqlite3_bind_blob(stmtInsert, 2, vec, dim * sizeof(float), SQLITE_TRANSIENT);
+    sqlite3_step(stmtInsert);
+    sqlite3_free(vec);
+  }
+  sqlite3_finalize(stmtInsert);
+
+  // Query BEFORE training (flat scan path)
+  {
+    float *qvec = sqlite3_malloc(dim * sizeof(float));
+    if (qvec) {
+      for (int d = 0; d < dim; d++) qvec[d] = 0.5f;
+      sqlite3_stmt *sk = NULL;
+      snprintf(sql, sizeof(sql),
+        "SELECT rowid, distance FROM v WHERE emb MATCH ? LIMIT %d", k_limit);
+      sqlite3_prepare_v2(db, sql, -1, &sk, NULL);
+      if (sk) {
+        sqlite3_bind_blob(sk, 1, qvec, dim * sizeof(float), SQLITE_TRANSIENT);
+        while (sqlite3_step(sk) == SQLITE_ROW) {}
+        sqlite3_finalize(sk);
+      }
+      sqlite3_free(qvec);
+    }
+  }
+
+  // Train
+  sqlite3_exec(db,
+    "INSERT INTO v(v) VALUES ('compute-centroids')",
+    NULL, NULL, NULL);
+
+  // Change nprobe at runtime (can exceed nlist -- tests clamping in query)
+  {
+    char cmd[64];
+    snprintf(cmd, sizeof(cmd),
+      "INSERT INTO v(v) VALUES ('nprobe=%d')", nprobe_initial);
+    sqlite3_exec(db, cmd, NULL, NULL, NULL);
+  }
+
+  // Multiple KNN queries with different fuzz-derived query vectors
+  for (int q = 0; q < num_queries; q++) {
+    float *qvec = sqlite3_malloc(dim * sizeof(float));
+    if (!qvec) break;
+    for (int d = 0; d < dim; d++) {
+      if (offset < payload_size) {
+        qvec[d] = ((float)(int8_t)payload[offset++]) / 10.0f;
+      } else {
+        qvec[d] = (q == 0) ? 1.0f : 0.0f;
+      }
+    }
+
+    sqlite3_stmt *sk = NULL;
+    snprintf(sql, sizeof(sql),
+      "SELECT rowid, distance FROM v WHERE emb MATCH ? LIMIT %d", k_limit);
+    sqlite3_prepare_v2(db, sql, -1, &sk, NULL);
+    if (sk) {
+      sqlite3_bind_blob(sk, 1, qvec, dim * sizeof(float), SQLITE_TRANSIENT);
+      while (sqlite3_step(sk) == SQLITE_ROW) {}
+      sqlite3_finalize(sk);
+    }
+    sqlite3_free(qvec);
+  }
+
+  // Delete half the vectors then query again
+  for (int i = 1; i <= num_vecs / 2; i++) {
+    char delsql[64];
+    snprintf(delsql, sizeof(delsql), "DELETE FROM v WHERE rowid = %d", i);
+    sqlite3_exec(db, delsql, NULL, NULL, NULL);
+  }
+
+  // Query after mass deletion
+  {
+    float *qvec = sqlite3_malloc(dim * sizeof(float));
+    if (qvec) {
+      for (int d = 0; d < dim; d++) qvec[d] = -0.5f;
+      sqlite3_stmt *sk = NULL;
+      snprintf(sql, sizeof(sql),
+        "SELECT rowid, distance FROM v WHERE emb MATCH ? LIMIT %d", k_limit);
+      sqlite3_prepare_v2(db, sql, -1, &sk, NULL);
+      if (sk) {
+        sqlite3_bind_blob(sk, 1, qvec, dim * sizeof(float), SQLITE_TRANSIENT);
+        while (sqlite3_step(sk) == SQLITE_ROW) {}
+        sqlite3_finalize(sk);
+      }
+      sqlite3_free(qvec);
+    }
+  }
+
+  sqlite3_close(db);
+  return 0;
+}

tests/fuzz/ivf-operations.c

@@ -0,0 +1,121 @@
+#include <stdint.h>
+#include <stddef.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include "sqlite-vec.h"
+#include "sqlite3.h"
+#include <assert.h>
+
+int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) {
+  if (size < 6) return 0;
+
+  int rc;
+  sqlite3 *db;
+  sqlite3_stmt *stmtInsert = NULL;
+  sqlite3_stmt *stmtDelete = NULL;
+  sqlite3_stmt *stmtKnn = NULL;
+  sqlite3_stmt *stmtScan = NULL;
+
+  rc = sqlite3_open(":memory:", &db);
+  assert(rc == SQLITE_OK);
+  rc = sqlite3_vec_init(db, NULL, NULL);
+  assert(rc == SQLITE_OK);
+
+  rc = sqlite3_exec(db,
+    "CREATE VIRTUAL TABLE v USING vec0(emb float[4] indexed by ivf(nlist=4, nprobe=4))",
+    NULL, NULL, NULL);
+  if (rc != SQLITE_OK) { sqlite3_close(db); return 0; }
+
+  sqlite3_prepare_v2(db,
+    "INSERT INTO v(v, emb) VALUES (?, ?)", -1, &stmtInsert, NULL);
+  sqlite3_prepare_v2(db,
+    "DELETE FROM v WHERE rowid = ?", -1, &stmtDelete, NULL);
+  sqlite3_prepare_v2(db,
+    "SELECT rowid, distance FROM v WHERE emb MATCH ? LIMIT 3",
+    -1, &stmtKnn, NULL);
+  sqlite3_prepare_v2(db,
+    "SELECT rowid FROM v", -1, &stmtScan, NULL);
+
+  if (!stmtInsert || !stmtDelete || !stmtKnn || !stmtScan) goto cleanup;
+
+  size_t i = 0;
+  while (i + 2 <= size) {
+    uint8_t op = data[i++] % 7;
+    uint8_t rowid_byte = data[i++];
+    int64_t rowid = (int64_t)(rowid_byte % 32) + 1;
+
+    switch (op) {
+      case 0: {
+        // INSERT: consume 16 bytes for 4 floats, or use what's left
+        float vec[4] = {0.0f, 0.0f, 0.0f, 0.0f};
+        for (int j = 0; j < 4 && i < size; j++, i++) {
+          vec[j] = (float)((int8_t)data[i]) / 10.0f;
+        }
+        sqlite3_reset(stmtInsert);
+        sqlite3_bind_int64(stmtInsert, 1, rowid);
+        sqlite3_bind_blob(stmtInsert, 2, vec, sizeof(vec), SQLITE_TRANSIENT);
+        sqlite3_step(stmtInsert);
+        break;
+      }
+      case 1: {
+        // DELETE
+        sqlite3_reset(stmtDelete);
+        sqlite3_bind_int64(stmtDelete, 1, rowid);
+        sqlite3_step(stmtDelete);
+        break;
+      }
+      case 2: {
+        // KNN query with a fixed query vector
+        float qvec[4] = {1.0f, 0.0f, 0.0f, 0.0f};
+        sqlite3_reset(stmtKnn);
+        sqlite3_bind_blob(stmtKnn, 1, qvec, sizeof(qvec), SQLITE_STATIC);
+        while (sqlite3_step(stmtKnn) == SQLITE_ROW) {}
+        break;
+      }
+      case 3: {
+        // Full scan
+        sqlite3_reset(stmtScan);
+        while (sqlite3_step(stmtScan) == SQLITE_ROW) {}
+        break;
+      }
+      case 4: {
+        // compute-centroids command
+        sqlite3_exec(db,
+          "INSERT INTO v(v) VALUES ('compute-centroids')",
+          NULL, NULL, NULL);
+        break;
+      }
+      case 5: {
+        // clear-centroids command
+        sqlite3_exec(db,
+          "INSERT INTO v(v) VALUES ('clear-centroids')",
+          NULL, NULL, NULL);
+        break;
+      }
+      case 6: {
+        // nprobe=N command
+        if (i < size) {
+          uint8_t n = data[i++];
+          int nprobe = (n % 4) + 1;
+          char buf[64];
+          snprintf(buf, sizeof(buf),
+            "INSERT INTO v(v) VALUES ('nprobe=%d')", nprobe);
+          sqlite3_exec(db, buf, NULL, NULL, NULL);
+        }
+        break;
+      }
+    }
+  }
+
+  // Final operations — must not crash regardless of prior state
+  sqlite3_exec(db, "SELECT * FROM v", NULL, NULL, NULL);
+
+cleanup:
+  sqlite3_finalize(stmtInsert);
+  sqlite3_finalize(stmtDelete);
+  sqlite3_finalize(stmtKnn);
+  sqlite3_finalize(stmtScan);
+  sqlite3_close(db);
+  return 0;
+}

tests/fuzz/ivf-quantize.c

@@ -0,0 +1,129 @@
+/**
+ * Fuzz target: IVF quantization functions.
+ *
+ * Directly exercises ivf_quantize_int8 and ivf_quantize_binary with
+ * fuzz-controlled dimensions and float data. Targets:
+ * - ivf_quantize_int8: clamping, int8 overflow boundary
+ * - ivf_quantize_binary: D not divisible by 8, memset(D/8) undercount
+ * - Round-trip through CREATE TABLE + INSERT with quantized IVF
+ */
+#include <stdint.h>
+#include <stddef.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <math.h>
+#include "sqlite-vec.h"
+#include "sqlite3.h"
+#include <assert.h>
+
+int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) {
+  if (size < 8) return 0;
+
+  int rc;
+  sqlite3 *db;
+
+  rc = sqlite3_open(":memory:", &db);
+  assert(rc == SQLITE_OK);
+  rc = sqlite3_vec_init(db, NULL, NULL);
+  assert(rc == SQLITE_OK);
+
+  // Byte 0: quantizer type (0=int8, 1=binary)
+  // Byte 1: dimension (1..64, but we test edge cases)
+  // Byte 2: nlist (1..8)
+  // Byte 3: num_vectors to insert (1..32)
+  // Remaining: float data
+  int qtype = data[0] % 2;
+  int dim = (data[1] % 64) + 1;
+  int nlist = (data[2] % 8) + 1;
+  int num_vecs = (data[3] % 32) + 1;
+  const uint8_t *payload = data + 4;
+  size_t payload_size = size - 4;
+
+  // For binary quantizer, D must be multiple of 8 to avoid the D/8 bug
+  // in production. But we explicitly want to test non-multiples too to
+  // find the bug. Use dim as-is.
+  const char *quantizer = qtype ? "binary" : "int8";
+
+  // Binary quantizer needs D multiple of 8 in current code, but let's
+  // test both valid and invalid dimensions to see what happens.
+  // For binary with non-multiple-of-8, the code does memset(dst, 0, D/8)
+  // which underallocates when D%8 != 0.
+  char sql[256];
+  snprintf(sql, sizeof(sql),
+    "CREATE VIRTUAL TABLE v USING vec0("
+    "emb float[%d] indexed by ivf(nlist=%d, nprobe=%d, quantizer=%s))",
+    dim, nlist, nlist, quantizer);
+
+  rc = sqlite3_exec(db, sql, NULL, NULL, NULL);
+  if (rc != SQLITE_OK) { sqlite3_close(db); return 0; }
+
+  // Insert vectors with fuzz-controlled float values
+  sqlite3_stmt *stmtInsert = NULL;
+  sqlite3_prepare_v2(db,
+    "INSERT INTO v(v, emb) VALUES (?, ?)", -1, &stmtInsert, NULL);
+  if (!stmtInsert) { sqlite3_close(db); return 0; }
+
+  size_t offset = 0;
+  for (int i = 0; i < num_vecs && offset < payload_size; i++) {
+    // Build float vector from fuzz data
+    float *vec = sqlite3_malloc(dim * sizeof(float));
+    if (!vec) break;
+
+    for (int d = 0; d < dim; d++) {
+      if (offset + 4 <= payload_size) {
+        // Use raw bytes as float -- can produce NaN, Inf, denormals
+        memcpy(&vec[d], payload + offset, sizeof(float));
+        offset += 4;
+      } else if (offset < payload_size) {
+        // Partial: use byte as scaled value
+        vec[d] = ((float)(int8_t)payload[offset++]) / 50.0f;
+      } else {
+        vec[d] = 0.0f;
+      }
+    }
+
+    sqlite3_reset(stmtInsert);
+    sqlite3_bind_int64(stmtInsert, 1, (int64_t)(i + 1));
+    sqlite3_bind_blob(stmtInsert, 2, vec, dim * sizeof(float), SQLITE_TRANSIENT);
+    sqlite3_step(stmtInsert);
+    sqlite3_free(vec);
+  }
+  sqlite3_finalize(stmtInsert);
+
+  // Trigger compute-centroids to exercise kmeans + quantization together
+  sqlite3_exec(db,
+    "INSERT INTO v(v) VALUES ('compute-centroids')",
+    NULL, NULL, NULL);
+
+  // KNN query with fuzz-derived query vector
+  {
+    float *qvec = sqlite3_malloc(dim * sizeof(float));
+    if (qvec) {
+      for (int d = 0; d < dim; d++) {
+        if (offset < payload_size) {
+          qvec[d] = ((float)(int8_t)payload[offset++]) / 10.0f;
+        } else {
+          qvec[d] = 1.0f;
+        }
+      }
+
+      sqlite3_stmt *stmtKnn = NULL;
+      sqlite3_prepare_v2(db,
+        "SELECT rowid, distance FROM v WHERE emb MATCH ? LIMIT 5",
+        -1, &stmtKnn, NULL);
+      if (stmtKnn) {
+        sqlite3_bind_blob(stmtKnn, 1, qvec, dim * sizeof(float), SQLITE_TRANSIENT);
+        while (sqlite3_step(stmtKnn) == SQLITE_ROW) {}
+        sqlite3_finalize(stmtKnn);
+      }
+      sqlite3_free(qvec);
+    }
+  }
+
+  // Full scan
+  sqlite3_exec(db, "SELECT * FROM v", NULL, NULL, NULL);
+
+  sqlite3_close(db);
+  return 0;
+}

tests/fuzz/ivf-rescore.c

@@ -0,0 +1,182 @@
+/**
+ * Fuzz target: IVF oversample + rescore path.
+ *
+ * Specifically targets the code path where quantizer != none AND
+ * oversample > 1, which triggers:
+ * 1. Quantized KNN scan to collect oversample*k candidates
+ * 2. Full-precision vector lookup from _ivf_vectors table
+ * 3. Re-scoring with float32 distances
+ * 4. Re-sort and truncation
+ *
+ * This path has the most complex memory management in the KNN query:
+ * - Two separate distance computations (quantized + float)
+ * - Cross-table lookups (cells + vectors KV store)
+ * - Candidate array resizing
+ * - qsort over partially re-scored arrays
+ *
+ * Also tests the int8 + binary quantization round-trip fidelity
+ * under adversarial float inputs.
+ */
+#include <stdint.h>
+#include <stddef.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <math.h>
+#include "sqlite-vec.h"
+#include "sqlite3.h"
+#include <assert.h>
+
+int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) {
+  if (size < 12) return 0;
+
+  int rc;
+  sqlite3 *db;
+
+  rc = sqlite3_open(":memory:", &db);
+  assert(rc == SQLITE_OK);
+  rc = sqlite3_vec_init(db, NULL, NULL);
+  assert(rc == SQLITE_OK);
+
+  // Header
+  int quantizer_type = (data[0] % 2) + 1;  // 1=int8, 2=binary (never none)
+  int dim = (data[1] % 32) + 8;  // 8..39
+  int nlist = (data[2] % 8) + 1;  // 1..8
+  int oversample = (data[3] % 4) + 2;  // 2..5 (always > 1)
+  int num_vecs = (data[4] % 60) + 8;  // 8..67
+  int k_limit = (data[5] % 15) + 1;  // 1..15
+
+  const uint8_t *payload = data + 6;
+  size_t payload_size = size - 6;
+
+  // Binary quantizer needs D multiple of 8
+  if (quantizer_type == 2) {
+    dim = ((dim + 7) / 8) * 8;
+  }
+
+  const char *qname = (quantizer_type == 1) ? "int8" : "binary";
+
+  char sql[512];
+  snprintf(sql, sizeof(sql),
+    "CREATE VIRTUAL TABLE v USING vec0("
+    "emb float[%d] indexed by ivf(nlist=%d, nprobe=%d, quantizer=%s, oversample=%d))",
+    dim, nlist, nlist, qname, oversample);
+
+  rc = sqlite3_exec(db, sql, NULL, NULL, NULL);
+  if (rc != SQLITE_OK) { sqlite3_close(db); return 0; }
+
+  // Insert vectors with diverse values
+  sqlite3_stmt *stmtInsert = NULL;
+  sqlite3_prepare_v2(db,
+    "INSERT INTO v(v, emb) VALUES (?, ?)", -1, &stmtInsert, NULL);
+  if (!stmtInsert) { sqlite3_close(db); return 0; }
+
+  size_t offset = 0;
+  for (int i = 0; i < num_vecs; i++) {
+    float *vec = sqlite3_malloc(dim * sizeof(float));
+    if (!vec) break;
+    for (int d = 0; d < dim; d++) {
+      if (offset + 4 <= payload_size) {
+        // Use raw bytes as float for adversarial values
+        memcpy(&vec[d], payload + offset, sizeof(float));
+        offset += 4;
+        // Sanitize: replace NaN/Inf with bounded values to avoid
+        // poisoning the entire computation. We want edge values,
+        // not complete nonsense.
+        if (isnan(vec[d]) || isinf(vec[d])) {
+          vec[d] = (vec[d] > 0) ? 1e6f : -1e6f;
+          if (isnan(vec[d])) vec[d] = 0.0f;
+        }
+      } else if (offset < payload_size) {
+        vec[d] = ((float)(int8_t)payload[offset++]) / 30.0f;
+      } else {
+        vec[d] = (float)(i * dim + d) * 0.001f;
+      }
+    }
+    sqlite3_reset(stmtInsert);
+    sqlite3_bind_int64(stmtInsert, 1, (int64_t)(i + 1));
+    sqlite3_bind_blob(stmtInsert, 2, vec, dim * sizeof(float), SQLITE_TRANSIENT);
+    sqlite3_step(stmtInsert);
+    sqlite3_free(vec);
+  }
+  sqlite3_finalize(stmtInsert);
+
+  // Train
+  sqlite3_exec(db,
+    "INSERT INTO v(v) VALUES ('compute-centroids')",
+    NULL, NULL, NULL);
+
+  // Multiple KNN queries to exercise rescore path
+  for (int q = 0; q < 4; q++) {
+    float *qvec = sqlite3_malloc(dim * sizeof(float));
+    if (!qvec) break;
+    for (int d = 0; d < dim; d++) {
+      if (offset < payload_size) {
+        qvec[d] = ((float)(int8_t)payload[offset++]) / 10.0f;
+      } else {
+        qvec[d] = (q == 0) ? 1.0f : (q == 1) ? -1.0f : 0.0f;
+      }
+    }
+
+    sqlite3_stmt *sk = NULL;
+    snprintf(sql, sizeof(sql),
+      "SELECT rowid, distance FROM v WHERE emb MATCH ? LIMIT %d", k_limit);
+    sqlite3_prepare_v2(db, sql, -1, &sk, NULL);
+    if (sk) {
+      sqlite3_bind_blob(sk, 1, qvec, dim * sizeof(float), SQLITE_TRANSIENT);
+      while (sqlite3_step(sk) == SQLITE_ROW) {}
+      sqlite3_finalize(sk);
+    }
+    sqlite3_free(qvec);
+  }
+
+  // Delete some vectors, then query again (rescore with missing _ivf_vectors rows)
+  for (int i = 1; i <= num_vecs / 3; i++) {
+    char delsql[64];
+    snprintf(delsql, sizeof(delsql), "DELETE FROM v WHERE rowid = %d", i);
+    sqlite3_exec(db, delsql, NULL, NULL, NULL);
+  }
+
+  {
+    float *qvec = sqlite3_malloc(dim * sizeof(float));
+    if (qvec) {
+      for (int d = 0; d < dim; d++) qvec[d] = 0.5f;
+      sqlite3_stmt *sk = NULL;
+      snprintf(sql, sizeof(sql),
+        "SELECT rowid, distance FROM v WHERE emb MATCH ? LIMIT %d", k_limit);
+      sqlite3_prepare_v2(db, sql, -1, &sk, NULL);
+      if (sk) {
+        sqlite3_bind_blob(sk, 1, qvec, dim * sizeof(float), SQLITE_TRANSIENT);
+        while (sqlite3_step(sk) == SQLITE_ROW) {}
+        sqlite3_finalize(sk);
+      }
+      sqlite3_free(qvec);
+    }
+  }
+
+  // Retrain after deletions
+  sqlite3_exec(db,
+    "INSERT INTO v(v) VALUES ('compute-centroids')",
+    NULL, NULL, NULL);
+
+  // Query after retrain
+  {
+    float *qvec = sqlite3_malloc(dim * sizeof(float));
+    if (qvec) {
+      for (int d = 0; d < dim; d++) qvec[d] = -0.3f;
+      sqlite3_stmt *sk = NULL;
+      snprintf(sql, sizeof(sql),
+        "SELECT rowid, distance FROM v WHERE emb MATCH ? LIMIT %d", k_limit);
+      sqlite3_prepare_v2(db, sql, -1, &sk, NULL);
+      if (sk) {
+        sqlite3_bind_blob(sk, 1, qvec, dim * sizeof(float), SQLITE_TRANSIENT);
+        while (sqlite3_step(sk) == SQLITE_ROW) {}
+        sqlite3_finalize(sk);
+      }
+      sqlite3_free(qvec);
+    }
+  }
+
+  sqlite3_close(db);
+  return 0;
+}

tests/fuzz/ivf-shadow-corrupt.c

@@ -0,0 +1,228 @@
+/**
+ * Fuzz target: IVF shadow table corruption.
+ *
+ * Creates a trained IVF table, then corrupts IVF shadow table blobs
+ * (centroids, cells validity/rowids/vectors, rowid_map) with fuzz data.
+ * Then exercises all read/write paths. Must not crash.
+ *
+ * Targets:
+ * - Cell validity bitmap with wrong size
+ * - Cell rowids blob with wrong size/alignment
+ * - Cell vectors blob with wrong size
+ * - Centroid blob with wrong size
+ * - n_vectors inconsistent with validity bitmap
+ * - Missing rowid_map entries
+ * - KNN scan over corrupted cells
+ * - Insert/delete with corrupted rowid_map
+ */
+#include <stdint.h>
+#include <stddef.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include "sqlite-vec.h"
+#include "sqlite3.h"
+#include <assert.h>
+
+int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) {
+  if (size < 4) return 0;
+
+  int rc;
+  sqlite3 *db;
+
+  rc = sqlite3_open(":memory:", &db);
+  assert(rc == SQLITE_OK);
+  rc = sqlite3_vec_init(db, NULL, NULL);
+  assert(rc == SQLITE_OK);
+
+  // Create IVF table and insert enough vectors to train
+  rc = sqlite3_exec(db,
+    "CREATE VIRTUAL TABLE v USING vec0("
+    "emb float[8] indexed by ivf(nlist=2, nprobe=2))",
+    NULL, NULL, NULL);
+  if (rc != SQLITE_OK) { sqlite3_close(db); return 0; }
+
+  // Insert 10 vectors
+  {
+    sqlite3_stmt *si = NULL;
+    sqlite3_prepare_v2(db,
+      "INSERT INTO v(v, emb) VALUES (?, ?)", -1, &si, NULL);
+    if (!si) { sqlite3_close(db); return 0; }
+    for (int i = 0; i < 10; i++) {
+      float vec[8];
+      for (int d = 0; d < 8; d++) {
+        vec[d] = (float)(i * 8 + d) * 0.1f;
+      }
+      sqlite3_reset(si);
+      sqlite3_bind_int64(si, 1, i + 1);
+      sqlite3_bind_blob(si, 2, vec, sizeof(vec), SQLITE_TRANSIENT);
+      sqlite3_step(si);
+    }
+    sqlite3_finalize(si);
+  }
+
+  // Train
+  sqlite3_exec(db,
+    "INSERT INTO v(v) VALUES ('compute-centroids')",
+    NULL, NULL, NULL);
+
+  // Now corrupt shadow tables based on fuzz input
+  uint8_t target = data[0] % 10;
+  const uint8_t *payload = data + 1;
+  int payload_size = (int)(size - 1);
+
+  // Limit payload to avoid huge allocations
+  if (payload_size > 4096) payload_size = 4096;
+
+  sqlite3_stmt *stmt = NULL;
+
+  switch (target) {
+    case 0: {
+      // Corrupt cell validity blob
+      rc = sqlite3_prepare_v2(db,
+        "UPDATE v_ivf_cells00 SET validity = ? WHERE rowid = 1",
+        -1, &stmt, NULL);
+      if (rc == SQLITE_OK) {
+        sqlite3_bind_blob(stmt, 1, payload, payload_size, SQLITE_STATIC);
+        sqlite3_step(stmt); sqlite3_finalize(stmt);
+      }
+      break;
+    }
+    case 1: {
+      // Corrupt cell rowids blob
+      rc = sqlite3_prepare_v2(db,
+        "UPDATE v_ivf_cells00 SET rowids = ? WHERE rowid = 1",
+        -1, &stmt, NULL);
+      if (rc == SQLITE_OK) {
+        sqlite3_bind_blob(stmt, 1, payload, payload_size, SQLITE_STATIC);
+        sqlite3_step(stmt); sqlite3_finalize(stmt);
+      }
+      break;
+    }
+    case 2: {
+      // Corrupt cell vectors blob
+      rc = sqlite3_prepare_v2(db,
+        "UPDATE v_ivf_cells00 SET vectors = ? WHERE rowid = 1",
+        -1, &stmt, NULL);
+      if (rc == SQLITE_OK) {
+        sqlite3_bind_blob(stmt, 1, payload, payload_size, SQLITE_STATIC);
+        sqlite3_step(stmt); sqlite3_finalize(stmt);
+      }
+      break;
+    }
+    case 3: {
+      // Corrupt centroid blob
+      rc = sqlite3_prepare_v2(db,
+        "UPDATE v_ivf_centroids00 SET centroid = ? WHERE centroid_id = 0",
+        -1, &stmt, NULL);
+      if (rc == SQLITE_OK) {
+        sqlite3_bind_blob(stmt, 1, payload, payload_size, SQLITE_STATIC);
+        sqlite3_step(stmt); sqlite3_finalize(stmt);
+      }
+      break;
+    }
+    case 4: {
+      // Set n_vectors to a bogus value (larger than cell capacity)
+      int bogus_n = 99999;
+      if (payload_size >= 4) {
+        memcpy(&bogus_n, payload, 4);
+        bogus_n = abs(bogus_n) % 100000;
+      }
+      char sql[128];
+      snprintf(sql, sizeof(sql),
+        "UPDATE v_ivf_cells00 SET n_vectors = %d WHERE rowid = 1", bogus_n);
+      sqlite3_exec(db, sql, NULL, NULL, NULL);
+      break;
+    }
+    case 5: {
+      // Delete rowid_map entries (orphan vectors)
+      sqlite3_exec(db,
+        "DELETE FROM v_ivf_rowid_map00 WHERE rowid IN (1, 2, 3)",
+        NULL, NULL, NULL);
+      break;
+    }
+    case 6: {
+      // Corrupt rowid_map slot values
+      char sql[128];
+      int bogus_slot = payload_size > 0 ? (int)payload[0] * 1000 : 99999;
+      snprintf(sql, sizeof(sql),
+        "UPDATE v_ivf_rowid_map00 SET slot = %d WHERE rowid = 1", bogus_slot);
+      sqlite3_exec(db, sql, NULL, NULL, NULL);
+      break;
+    }
+    case 7: {
+      // Corrupt rowid_map cell_id values
+      sqlite3_exec(db,
+        "UPDATE v_ivf_rowid_map00 SET cell_id = 99999 WHERE rowid = 1",
+        NULL, NULL, NULL);
+      break;
+    }
+    case 8: {
+      // Delete all centroids (make trained but no centroids)
+      sqlite3_exec(db,
+        "DELETE FROM v_ivf_centroids00",
+        NULL, NULL, NULL);
+      break;
+    }
+    case 9: {
+      // Set validity to NULL
+      sqlite3_exec(db,
+        "UPDATE v_ivf_cells00 SET validity = NULL WHERE rowid = 1",
+        NULL, NULL, NULL);
+      break;
+    }
+  }
+
+  // Exercise all read paths over corrupted state — must not crash
+  float qvec[8] = {1.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f};
+
+  // KNN query
+  {
+    sqlite3_stmt *sk = NULL;
+    sqlite3_prepare_v2(db,
+      "SELECT rowid, distance FROM v WHERE emb MATCH ? LIMIT 5",
+      -1, &sk, NULL);
+    if (sk) {
+      sqlite3_bind_blob(sk, 1, qvec, sizeof(qvec), SQLITE_STATIC);
+      while (sqlite3_step(sk) == SQLITE_ROW) {}
+      sqlite3_finalize(sk);
+    }
+  }
+
+  // Full scan
+  sqlite3_exec(db, "SELECT * FROM v", NULL, NULL, NULL);
+
+  // Point query
+  sqlite3_exec(db, "SELECT * FROM v WHERE rowid = 1", NULL, NULL, NULL);
+  sqlite3_exec(db, "SELECT * FROM v WHERE rowid = 5", NULL, NULL, NULL);
+
+  // Delete
+  sqlite3_exec(db, "DELETE FROM v WHERE rowid = 3", NULL, NULL, NULL);
+
+  // Insert after corruption
+  {
+    float newvec[8] = {0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f, 0.5f};
+    sqlite3_stmt *si = NULL;
+    sqlite3_prepare_v2(db,
+      "INSERT INTO v(v, emb) VALUES (?, ?)", -1, &si, NULL);
+    if (si) {
+      sqlite3_bind_int64(si, 1, 100);
+      sqlite3_bind_blob(si, 2, newvec, sizeof(newvec), SQLITE_STATIC);
+      sqlite3_step(si);
+      sqlite3_finalize(si);
+    }
+  }
+
+  // compute-centroids over corrupted state
+  sqlite3_exec(db,
+    "INSERT INTO v(v) VALUES ('compute-centroids')",
+    NULL, NULL, NULL);
+
+  // clear-centroids
+  sqlite3_exec(db,
+    "INSERT INTO v(v) VALUES ('clear-centroids')",
+    NULL, NULL, NULL);
+
+  sqlite3_close(db);
+  return 0;
+}

tests/fuzz/rescore-create.c

@@ -1,6 +1,5 @@
 #include <stdint.h>
 #include <stddef.h>
-
 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
@@ -8,9 +7,6 @@
 #include "sqlite3.h"
 #include <assert.h>
 
-extern int sqlite3_vec_numpy_init(sqlite3 *db, char **pzErrMsg,
-                                  const sqlite3_api_routines *pApi);
-
 int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) {
   int rc = SQLITE_OK;
   sqlite3 *db;
@@ -20,17 +16,20 @@ int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) {
   assert(rc == SQLITE_OK);
   rc = sqlite3_vec_init(db, NULL, NULL);
   assert(rc == SQLITE_OK);
-  rc = sqlite3_vec_numpy_init(db, NULL, NULL);
-  assert(rc == SQLITE_OK);
 
-  rc = sqlite3_prepare_v2(db, "select * from vec_npy_each(?)", -1, &stmt, NULL);
+  sqlite3_str *s = sqlite3_str_new(NULL);
-  assert(rc == SQLITE_OK);
+  assert(s);
-  sqlite3_bind_blob(stmt, 1, data, size, SQLITE_STATIC);
+  sqlite3_str_appendall(s, "CREATE VIRTUAL TABLE v USING vec0(emb float[128] indexed by rescore(");
-  rc = sqlite3_step(stmt);
+  sqlite3_str_appendf(s, "%.*s", (int)size, data);
-  while (rc == SQLITE_ROW) {
+  sqlite3_str_appendall(s, "))");
-    rc = sqlite3_step(stmt);
+  const char *zSql = sqlite3_str_finish(s);
+  assert(zSql);
+
+  rc = sqlite3_prepare_v2(db, zSql, -1, &stmt, NULL);
+  sqlite3_free((void *)zSql);
+  if (rc == SQLITE_OK) {
+    sqlite3_step(stmt);
   }
-
   sqlite3_finalize(stmt);
   sqlite3_close(db);
   return 0;

tests/fuzz/rescore-create.dict

@@ -0,0 +1,20 @@
+"rescore"
+"quantizer"
+"bit"
+"int8"
+"oversample"
+"indexed"
+"by"
+"float"
+"("
+")"
+","
+"="
+"["
+"]"
+"1"
+"8"
+"16"
+"128"
+"256"
+"1024"

tests/fuzz/rescore-interleave.c

@@ -0,0 +1,151 @@
+#include <stdint.h>
+#include <stddef.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include "sqlite-vec.h"
+#include "sqlite3.h"
+#include <assert.h>
+
+/**
+ * Fuzz target: interleaved insert/update/delete/KNN operations on rescore
+ * tables with BOTH quantizer types, exercising the int8 quantizer path
+ * and the update code path that the existing rescore-operations.c misses.
+ *
+ * Key differences from rescore-operations.c:
+ * - Tests BOTH bit and int8 quantizers (the existing target only tests bit)
+ * - Fuzz-controlled query vectors (not fixed [1,0,0,...])
+ * - Exercises the UPDATE path (line 9080+ in sqlite-vec.c)
+ * - Tests with 16 dimensions (more realistic, exercises more of the
+ *   quantization loop)
+ * - Interleaves KNN between mutations to stress the blob_reopen path
+ *   when _rescore_vectors rows have been deleted/modified
+ */
+int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) {
+  if (size < 8) return 0;
+
+  int rc;
+  sqlite3 *db;
+  sqlite3_stmt *stmtInsert = NULL;
+  sqlite3_stmt *stmtUpdate = NULL;
+  sqlite3_stmt *stmtDelete = NULL;
+  sqlite3_stmt *stmtKnn = NULL;
+
+  rc = sqlite3_open(":memory:", &db);
+  assert(rc == SQLITE_OK);
+  rc = sqlite3_vec_init(db, NULL, NULL);
+  assert(rc == SQLITE_OK);
+
+  /* Use first byte to pick quantizer */
+  int use_int8 = data[0] & 1;
+  data++; size--;
+
+  const char *create_sql = use_int8
+    ? "CREATE VIRTUAL TABLE v USING vec0("
+      "emb float[16] indexed by rescore(quantizer=int8))"
+    : "CREATE VIRTUAL TABLE v USING vec0("
+      "emb float[16] indexed by rescore(quantizer=bit))";
+
+  rc = sqlite3_exec(db, create_sql, NULL, NULL, NULL);
+  if (rc != SQLITE_OK) { sqlite3_close(db); return 0; }
+
+  sqlite3_prepare_v2(db,
+    "INSERT INTO v(rowid, emb) VALUES (?, ?)", -1, &stmtInsert, NULL);
+  sqlite3_prepare_v2(db,
+    "UPDATE v SET emb = ? WHERE rowid = ?", -1, &stmtUpdate, NULL);
+  sqlite3_prepare_v2(db,
+    "DELETE FROM v WHERE rowid = ?", -1, &stmtDelete, NULL);
+  sqlite3_prepare_v2(db,
+    "SELECT rowid, distance FROM v WHERE emb MATCH ? "
+    "ORDER BY distance LIMIT 5", -1, &stmtKnn, NULL);
+
+  if (!stmtInsert || !stmtUpdate || !stmtDelete || !stmtKnn)
+    goto cleanup;
+
+  size_t i = 0;
+  while (i + 2 <= size) {
+    uint8_t op = data[i++] % 5; /* 5 operations now */
+    uint8_t rowid_byte = data[i++];
+    int64_t rowid = (int64_t)(rowid_byte % 24) + 1;
+
+    switch (op) {
+      case 0: {
+        /* INSERT: consume bytes for 16 floats */
+        float vec[16] = {0};
+        for (int j = 0; j < 16 && i < size; j++, i++) {
+          vec[j] = (float)((int8_t)data[i]) / 8.0f;
+        }
+        sqlite3_reset(stmtInsert);
+        sqlite3_bind_int64(stmtInsert, 1, rowid);
+        sqlite3_bind_blob(stmtInsert, 2, vec, sizeof(vec), SQLITE_TRANSIENT);
+        sqlite3_step(stmtInsert);
+        break;
+      }
+      case 1: {
+        /* DELETE */
+        sqlite3_reset(stmtDelete);
+        sqlite3_bind_int64(stmtDelete, 1, rowid);
+        sqlite3_step(stmtDelete);
+        break;
+      }
+      case 2: {
+        /* KNN with fuzz-controlled query vector */
+        float qvec[16] = {0};
+        for (int j = 0; j < 16 && i < size; j++, i++) {
+          qvec[j] = (float)((int8_t)data[i]) / 4.0f;
+        }
+        sqlite3_reset(stmtKnn);
+        sqlite3_bind_blob(stmtKnn, 1, qvec, sizeof(qvec), SQLITE_STATIC);
+        while (sqlite3_step(stmtKnn) == SQLITE_ROW) {
+          (void)sqlite3_column_int64(stmtKnn, 0);
+          (void)sqlite3_column_double(stmtKnn, 1);
+        }
+        break;
+      }
+      case 3: {
+        /* UPDATE: modify an existing vector (exercises rescore update path) */
+        float vec[16] = {0};
+        for (int j = 0; j < 16 && i < size; j++, i++) {
+          vec[j] = (float)((int8_t)data[i]) / 6.0f;
+        }
+        sqlite3_reset(stmtUpdate);
+        sqlite3_bind_blob(stmtUpdate, 1, vec, sizeof(vec), SQLITE_TRANSIENT);
+        sqlite3_bind_int64(stmtUpdate, 2, rowid);
+        sqlite3_step(stmtUpdate);
+        break;
+      }
+      case 4: {
+        /* INSERT then immediately UPDATE same row (stresses blob lifecycle) */
+        float vec1[16] = {0};
+        float vec2[16] = {0};
+        for (int j = 0; j < 16 && i < size; j++, i++) {
+          vec1[j] = (float)((int8_t)data[i]) / 10.0f;
+          vec2[j] = -vec1[j]; /* opposite direction */
+        }
+        /* Insert */
+        sqlite3_reset(stmtInsert);
+        sqlite3_bind_int64(stmtInsert, 1, rowid);
+        sqlite3_bind_blob(stmtInsert, 2, vec1, sizeof(vec1), SQLITE_TRANSIENT);
+        if (sqlite3_step(stmtInsert) == SQLITE_DONE) {
+          /* Only update if insert succeeded (rowid might already exist) */
+          sqlite3_reset(stmtUpdate);
+          sqlite3_bind_blob(stmtUpdate, 1, vec2, sizeof(vec2), SQLITE_TRANSIENT);
+          sqlite3_bind_int64(stmtUpdate, 2, rowid);
+          sqlite3_step(stmtUpdate);
+        }
+        break;
+      }
+    }
+  }
+
+  /* Final consistency check: full scan must not crash */
+  sqlite3_exec(db, "SELECT * FROM v", NULL, NULL, NULL);
+
+cleanup:
+  sqlite3_finalize(stmtInsert);
+  sqlite3_finalize(stmtUpdate);
+  sqlite3_finalize(stmtDelete);
+  sqlite3_finalize(stmtKnn);
+  sqlite3_close(db);
+  return 0;
+}

tests/fuzz/rescore-knn-deep.c

@@ -0,0 +1,178 @@
+#include <stdint.h>
+#include <stddef.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include "sqlite-vec.h"
+#include "sqlite3.h"
+#include <assert.h>
+
+/**
+ * Fuzz target: deep exercise of rescore KNN with fuzz-controlled query vectors
+ * and both quantizer types (bit + int8), multiple distance metrics.
+ *
+ * The existing rescore-operations.c only tests bit quantizer with a fixed
+ * query vector. This target:
+ * - Tests both bit and int8 quantizers
+ * - Uses fuzz-controlled query vectors (hits NaN/Inf/denormal paths)
+ * - Tests all distance metrics with int8 (L2, cosine, L1)
+ * - Exercises large LIMIT values (oversample multiplication)
+ * - Tests KNN with rowid IN constraints
+ * - Exercises the insert->query->update->query->delete->query cycle
+ */
+int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) {
+  if (size < 20) return 0;
+
+  int rc;
+  sqlite3 *db;
+
+  rc = sqlite3_open(":memory:", &db);
+  assert(rc == SQLITE_OK);
+  rc = sqlite3_vec_init(db, NULL, NULL);
+  assert(rc == SQLITE_OK);
+
+  /* Use first 4 bytes for configuration */
+  uint8_t config = data[0];
+  uint8_t num_inserts = (data[1] % 20) + 3; /* 3..22 inserts */
+  uint8_t limit_val = (data[2] % 50) + 1;   /* 1..50 for LIMIT */
+  uint8_t metric_choice = data[3] % 3;
+  data += 4;
+  size -= 4;
+
+  int use_int8 = config & 1;
+  const char *metric_str;
+  switch (metric_choice) {
+    case 0: metric_str = ""; break; /* default L2 */
+    case 1: metric_str = " distance_metric=cosine"; break;
+    case 2: metric_str = " distance_metric=l1"; break;
+    default: metric_str = ""; break;
+  }
+
+  /* Build CREATE TABLE statement */
+  char create_sql[256];
+  if (use_int8) {
+    snprintf(create_sql, sizeof(create_sql),
+      "CREATE VIRTUAL TABLE v USING vec0("
+      "emb float[16] indexed by rescore(quantizer=int8)%s)", metric_str);
+  } else {
+    /* bit quantizer ignores distance_metric for the coarse pass (always hamming),
+       but the float rescore phase uses the specified metric */
+    snprintf(create_sql, sizeof(create_sql),
+      "CREATE VIRTUAL TABLE v USING vec0("
+      "emb float[16] indexed by rescore(quantizer=bit)%s)", metric_str);
+  }
+
+  rc = sqlite3_exec(db, create_sql, NULL, NULL, NULL);
+  if (rc != SQLITE_OK) { sqlite3_close(db); return 0; }
+
+  /* Insert vectors using fuzz data */
+  {
+    sqlite3_stmt *ins = NULL;
+    sqlite3_prepare_v2(db,
+      "INSERT INTO v(rowid, emb) VALUES (?, ?)", -1, &ins, NULL);
+    if (!ins) { sqlite3_close(db); return 0; }
+
+    size_t cursor = 0;
+    for (int i = 0; i < num_inserts && cursor + 1 < size; i++) {
+      float vec[16];
+      for (int j = 0; j < 16; j++) {
+        if (cursor < size) {
+          /* Map fuzz byte to float -- includes potential for
+             interesting float values via reinterpretation */
+          int8_t sb = (int8_t)data[cursor++];
+          vec[j] = (float)sb / 5.0f;
+        } else {
+          vec[j] = 0.0f;
+        }
+      }
+      sqlite3_reset(ins);
+      sqlite3_bind_int64(ins, 1, (sqlite3_int64)(i + 1));
+      sqlite3_bind_blob(ins, 2, vec, sizeof(vec), SQLITE_TRANSIENT);
+      sqlite3_step(ins);
+    }
+    sqlite3_finalize(ins);
+  }
+
+  /* Build a fuzz-controlled query vector from remaining data */
+  float qvec[16] = {0};
+  {
+    size_t cursor = 0;
+    for (int j = 0; j < 16 && cursor < size; j++) {
+      int8_t sb = (int8_t)data[cursor++];
+      qvec[j] = (float)sb / 3.0f;
+    }
+  }
+
+  /* KNN query with fuzz-controlled vector and LIMIT */
+  {
+    char knn_sql[256];
+    snprintf(knn_sql, sizeof(knn_sql),
+      "SELECT rowid, distance FROM v WHERE emb MATCH ? "
+      "ORDER BY distance LIMIT %d", (int)limit_val);
+
+    sqlite3_stmt *knn = NULL;
+    sqlite3_prepare_v2(db, knn_sql, -1, &knn, NULL);
+    if (knn) {
+      sqlite3_bind_blob(knn, 1, qvec, sizeof(qvec), SQLITE_STATIC);
+      while (sqlite3_step(knn) == SQLITE_ROW) {
+        /* Read results to ensure distance computation didn't produce garbage
+           that crashes the cursor iteration */
+        (void)sqlite3_column_int64(knn, 0);
+        (void)sqlite3_column_double(knn, 1);
+      }
+      sqlite3_finalize(knn);
+    }
+  }
+
+  /* Update some vectors, then query again */
+  {
+    float uvec[16];
+    for (int j = 0; j < 16; j++) uvec[j] = qvec[15 - j]; /* reverse of query */
+    sqlite3_stmt *upd = NULL;
+    sqlite3_prepare_v2(db,
+      "UPDATE v SET emb = ? WHERE rowid = 1", -1, &upd, NULL);
+    if (upd) {
+      sqlite3_bind_blob(upd, 1, uvec, sizeof(uvec), SQLITE_STATIC);
+      sqlite3_step(upd);
+      sqlite3_finalize(upd);
+    }
+  }
+
+  /* Second KNN after update */
+  {
+    sqlite3_stmt *knn = NULL;
+    sqlite3_prepare_v2(db,
+      "SELECT rowid, distance FROM v WHERE emb MATCH ? "
+      "ORDER BY distance LIMIT 10", -1, &knn, NULL);
+    if (knn) {
+      sqlite3_bind_blob(knn, 1, qvec, sizeof(qvec), SQLITE_STATIC);
+      while (sqlite3_step(knn) == SQLITE_ROW) {}
+      sqlite3_finalize(knn);
+    }
+  }
+
+  /* Delete half the rows, then KNN again */
+  for (int i = 1; i <= num_inserts; i += 2) {
+    char del_sql[64];
+    snprintf(del_sql, sizeof(del_sql),
+      "DELETE FROM v WHERE rowid = %d", i);
+    sqlite3_exec(db, del_sql, NULL, NULL, NULL);
+  }
+
+  /* Third KNN after deletes -- exercises distance computation over
+     zeroed-out slots in the quantized chunk */
+  {
+    sqlite3_stmt *knn = NULL;
+    sqlite3_prepare_v2(db,
+      "SELECT rowid, distance FROM v WHERE emb MATCH ? "
+      "ORDER BY distance LIMIT 5", -1, &knn, NULL);
+    if (knn) {
+      sqlite3_bind_blob(knn, 1, qvec, sizeof(qvec), SQLITE_STATIC);
+      while (sqlite3_step(knn) == SQLITE_ROW) {}
+      sqlite3_finalize(knn);
+    }
+  }
+
+  sqlite3_close(db);
+  return 0;
+}

tests/fuzz/rescore-operations.c

@@ -0,0 +1,96 @@
+#include <stdint.h>
+#include <stddef.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include "sqlite-vec.h"
+#include "sqlite3.h"
+#include <assert.h>
+
+int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) {
+  if (size < 6) return 0;
+
+  int rc;
+  sqlite3 *db;
+  sqlite3_stmt *stmtInsert = NULL;
+  sqlite3_stmt *stmtDelete = NULL;
+  sqlite3_stmt *stmtKnn = NULL;
+  sqlite3_stmt *stmtScan = NULL;
+
+  rc = sqlite3_open(":memory:", &db);
+  assert(rc == SQLITE_OK);
+  rc = sqlite3_vec_init(db, NULL, NULL);
+  assert(rc == SQLITE_OK);
+
+  rc = sqlite3_exec(db,
+    "CREATE VIRTUAL TABLE v USING vec0("
+    "emb float[8] indexed by rescore(quantizer=bit))",
+    NULL, NULL, NULL);
+  if (rc != SQLITE_OK) { sqlite3_close(db); return 0; }
+
+  sqlite3_prepare_v2(db,
+    "INSERT INTO v(rowid, emb) VALUES (?, ?)", -1, &stmtInsert, NULL);
+  sqlite3_prepare_v2(db,
+    "DELETE FROM v WHERE rowid = ?", -1, &stmtDelete, NULL);
+  sqlite3_prepare_v2(db,
+    "SELECT rowid, distance FROM v WHERE emb MATCH ? ORDER BY distance LIMIT 3",
+    -1, &stmtKnn, NULL);
+  sqlite3_prepare_v2(db,
+    "SELECT rowid FROM v", -1, &stmtScan, NULL);
+
+  if (!stmtInsert || !stmtDelete || !stmtKnn || !stmtScan) goto cleanup;
+
+  size_t i = 0;
+  while (i + 2 <= size) {
+    uint8_t op = data[i++] % 4;
+    uint8_t rowid_byte = data[i++];
+    int64_t rowid = (int64_t)(rowid_byte % 32) + 1;
+
+    switch (op) {
+      case 0: {
+        // INSERT: consume 32 bytes for 8 floats, or use what's left
+        float vec[8] = {0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f};
+        for (int j = 0; j < 8 && i < size; j++, i++) {
+          vec[j] = (float)((int8_t)data[i]) / 10.0f;
+        }
+        sqlite3_reset(stmtInsert);
+        sqlite3_bind_int64(stmtInsert, 1, rowid);
+        sqlite3_bind_blob(stmtInsert, 2, vec, sizeof(vec), SQLITE_TRANSIENT);
+        sqlite3_step(stmtInsert);
+        break;
+      }
+      case 1: {
+        // DELETE
+        sqlite3_reset(stmtDelete);
+        sqlite3_bind_int64(stmtDelete, 1, rowid);
+        sqlite3_step(stmtDelete);
+        break;
+      }
+      case 2: {
+        // KNN query with a fixed query vector
+        float qvec[8] = {1.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 0.0f};
+        sqlite3_reset(stmtKnn);
+        sqlite3_bind_blob(stmtKnn, 1, qvec, sizeof(qvec), SQLITE_STATIC);
+        while (sqlite3_step(stmtKnn) == SQLITE_ROW) {}
+        break;
+      }
+      case 3: {
+        // Full scan
+        sqlite3_reset(stmtScan);
+        while (sqlite3_step(stmtScan) == SQLITE_ROW) {}
+        break;
+      }
+    }
+  }
+
+  // Final operations -- must not crash regardless of prior state
+  sqlite3_exec(db, "SELECT * FROM v", NULL, NULL, NULL);
+
+cleanup:
+  sqlite3_finalize(stmtInsert);
+  sqlite3_finalize(stmtDelete);
+  sqlite3_finalize(stmtKnn);
+  sqlite3_finalize(stmtScan);
+  sqlite3_close(db);
+  return 0;
+}

tests/fuzz/rescore-quantize-edge.c

@@ -0,0 +1,177 @@
+#include <stdint.h>
+#include <stddef.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <math.h>
+#include "sqlite-vec.h"
+#include "sqlite3.h"
+#include <assert.h>
+
+/* Test wrappers from sqlite-vec-rescore.c (SQLITE_VEC_TEST build) */
+extern void _test_rescore_quantize_float_to_bit(const float *src, uint8_t *dst, size_t dim);
+extern void _test_rescore_quantize_float_to_int8(const float *src, int8_t *dst, size_t dim);
+extern size_t _test_rescore_quantized_byte_size_bit(size_t dimensions);
+extern size_t _test_rescore_quantized_byte_size_int8(size_t dimensions);
+
+/**
+ * Fuzz target: edge cases in rescore quantization functions.
+ *
+ * The existing rescore-quantize.c only tests dimensions that are multiples of 8
+ * and never passes special float values. This target:
+ *
+ * - Tests rescore_quantized_byte_size with arbitrary dimension values
+ *   (including 0, 1, 7, MAX values -- looking for integer division issues)
+ * - Passes raw float reinterpretation of fuzz bytes (NaN, Inf, denormals,
+ *   negative zero -- these are the values that break min/max/range logic)
+ * - Tests the int8 quantizer with all-identical values (range=0 branch)
+ * - Tests the int8 quantizer with extreme ranges (overflow in scale calc)
+ * - Tests bit quantizer with exact float threshold (0.0f boundary)
+ */
+int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) {
+  if (size < 8) return 0;
+
+  uint8_t mode = data[0] % 5;
+  data++; size--;
+
+  switch (mode) {
+    case 0: {
+      /* Test rescore_quantized_byte_size with fuzz-controlled dimensions.
+         This function does dimensions / CHAR_BIT for bit, dimensions for int8.
+         We're checking it doesn't do anything weird with edge values. */
+      if (size < sizeof(size_t)) return 0;
+      size_t dim;
+      memcpy(&dim, data, sizeof(dim));
+
+      /* These should never crash, just return values */
+      size_t bit_size = _test_rescore_quantized_byte_size_bit(dim);
+      size_t int8_size = _test_rescore_quantized_byte_size_int8(dim);
+
+      /* Verify basic invariants */
+      (void)bit_size;
+      (void)int8_size;
+      break;
+    }
+
+    case 1: {
+      /* Bit quantize with raw reinterpreted floats (NaN, Inf, denormal).
+         The key check: src[i] >= 0.0f -- NaN comparison is always false,
+         so NaN should produce 0-bits. But denormals and -0.0f are tricky. */
+      size_t num_floats = size / sizeof(float);
+      if (num_floats == 0) return 0;
+      /* Round to multiple of 8 for bit quantizer */
+      size_t dim = (num_floats / 8) * 8;
+      if (dim == 0) return 0;
+
+      const float *src = (const float *)data;
+      size_t bit_bytes = dim / 8;
+      uint8_t *dst = (uint8_t *)malloc(bit_bytes);
+      if (!dst) return 0;
+
+      _test_rescore_quantize_float_to_bit(src, dst, dim);
+
+      /* Verify: for each bit, if src >= 0 then bit should be set */
+      for (size_t i = 0; i < dim; i++) {
+        int bit_set = (dst[i / 8] >> (i % 8)) & 1;
+        if (src[i] >= 0.0f) {
+          assert(bit_set == 1);
+        } else if (src[i] < 0.0f) {
+          /* Definitely negative -- bit must be 0 */
+          assert(bit_set == 0);
+        }
+        /* NaN: comparison is false, so bit_set should be 0 */
+      }
+
+      free(dst);
+      break;
+    }
+
+    case 2: {
+      /* Int8 quantize with raw reinterpreted floats.
+         The dangerous paths:
+         - All values identical (range == 0) -> memset path
+         - vmin/vmax with NaN (NaN < anything is false, NaN > anything is false)
+         - Extreme range causing scale = 255/range to be Inf or 0
+         - denormals near the clamping boundaries */
+      size_t num_floats = size / sizeof(float);
+      if (num_floats == 0) return 0;
+
+      const float *src = (const float *)data;
+      int8_t *dst = (int8_t *)malloc(num_floats);
+      if (!dst) return 0;
+
+      _test_rescore_quantize_float_to_int8(src, dst, num_floats);
+
+      /* Output must always be in [-128, 127] (trivially true for int8_t,
+         but check the actual clamping logic worked) */
+      for (size_t i = 0; i < num_floats; i++) {
+        assert(dst[i] >= -128 && dst[i] <= 127);
+      }
+
+      free(dst);
+      break;
+    }
+
+    case 3: {
+      /* Int8 quantize stress: all-same values (range=0 branch) */
+      size_t dim = (size < 64) ? size : 64;
+      if (dim == 0) return 0;
+
+      float *src = (float *)malloc(dim * sizeof(float));
+      int8_t *dst = (int8_t *)malloc(dim);
+      if (!src || !dst) { free(src); free(dst); return 0; }
+
+      /* Fill with a single value derived from fuzz data */
+      float val;
+      memcpy(&val, data, sizeof(float) < size ? sizeof(float) : size);
+      for (size_t i = 0; i < dim; i++) src[i] = val;
+
+      _test_rescore_quantize_float_to_int8(src, dst, dim);
+
+      /* All outputs should be 0 when range == 0 */
+      for (size_t i = 0; i < dim; i++) {
+        assert(dst[i] == 0);
+      }
+
+      free(src);
+      free(dst);
+      break;
+    }
+
+    case 4: {
+      /* Int8 quantize with extreme range: one huge positive, one huge negative.
+         Tests scale = 255.0f / range overflow to Inf, then v * Inf = Inf,
+         then clamping to [-128, 127]. */
+      if (size < 2 * sizeof(float)) return 0;
+
+      float extreme[2];
+      memcpy(extreme, data, 2 * sizeof(float));
+
+      /* Only test if both are finite (NaN/Inf tested in case 2) */
+      if (!isfinite(extreme[0]) || !isfinite(extreme[1])) return 0;
+
+      /* Build a vector with these two extreme values plus some fuzz */
+      size_t dim = 16;
+      float src[16];
+      src[0] = extreme[0];
+      src[1] = extreme[1];
+      for (size_t i = 2; i < dim; i++) {
+        if (2 * sizeof(float) + (i - 2) < size) {
+          src[i] = (float)((int8_t)data[2 * sizeof(float) + (i - 2)]) * 1000.0f;
+        } else {
+          src[i] = 0.0f;
+        }
+      }
+
+      int8_t dst[16];
+      _test_rescore_quantize_float_to_int8(src, dst, dim);
+
+      for (size_t i = 0; i < dim; i++) {
+        assert(dst[i] >= -128 && dst[i] <= 127);
+      }
+      break;
+    }
+  }
+
+  return 0;
+}

tests/fuzz/rescore-quantize.c

@@ -0,0 +1,54 @@
+#include <stdint.h>
+#include <stddef.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include "sqlite-vec.h"
+#include "sqlite3.h"
+#include <assert.h>
+
+/* These are SQLITE_VEC_TEST wrappers defined in sqlite-vec-rescore.c */
+extern void _test_rescore_quantize_float_to_bit(const float *src, uint8_t *dst, size_t dim);
+extern void _test_rescore_quantize_float_to_int8(const float *src, int8_t *dst, size_t dim);
+
+int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) {
+  /* Need at least 4 bytes for one float */
+  if (size < 4) return 0;
+
+  /* Use the input as an array of floats. Dimensions must be a multiple of 8
+   * for the bit quantizer. */
+  size_t num_floats = size / sizeof(float);
+  if (num_floats == 0) return 0;
+
+  /* Round down to multiple of 8 for bit quantizer compatibility */
+  size_t dim = (num_floats / 8) * 8;
+  if (dim == 0) dim = 8;
+  if (dim > num_floats) return 0;
+
+  const float *src = (const float *)data;
+
+  /* Allocate output buffers */
+  size_t bit_bytes = dim / 8;
+  uint8_t *bit_dst = (uint8_t *)malloc(bit_bytes);
+  int8_t *int8_dst = (int8_t *)malloc(dim);
+  if (!bit_dst || !int8_dst) {
+    free(bit_dst);
+    free(int8_dst);
+    return 0;
+  }
+
+  /* Test bit quantization */
+  _test_rescore_quantize_float_to_bit(src, bit_dst, dim);
+
+  /* Test int8 quantization */
+  _test_rescore_quantize_float_to_int8(src, int8_dst, dim);
+
+  /* Verify int8 output is in range */
+  for (size_t i = 0; i < dim; i++) {
+    assert(int8_dst[i] >= -128 && int8_dst[i] <= 127);
+  }
+
+  free(bit_dst);
+  free(int8_dst);
+  return 0;
+}

tests/fuzz/rescore-shadow-corrupt.c

@@ -0,0 +1,230 @@
+#include <stdint.h>
+#include <stddef.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include "sqlite-vec.h"
+#include "sqlite3.h"
+#include <assert.h>
+
+/**
+ * Fuzz target: corrupt rescore shadow tables then exercise KNN/read/write.
+ *
+ * This targets the dangerous code paths in rescore_knn (Phase 1 + 2):
+ * - sqlite3_blob_read into baseVectors with potentially wrong-sized blobs
+ * - distance computation on corrupted/partial quantized data
+ * - blob_reopen on _rescore_vectors with missing/corrupted rows
+ * - insert/delete after corruption (blob_write to wrong offsets)
+ *
+ * The existing shadow-corrupt.c only tests vec0 without rescore.
+ */
+int LLVMFuzzerTestOneInput(const uint8_t *data, size_t size) {
+  if (size < 4) return 0;
+
+  int rc;
+  sqlite3 *db;
+
+  rc = sqlite3_open(":memory:", &db);
+  assert(rc == SQLITE_OK);
+  rc = sqlite3_vec_init(db, NULL, NULL);
+  assert(rc == SQLITE_OK);
+
+  /* Pick quantizer type from first byte */
+  int use_int8 = data[0] & 1;
+  int target = (data[1] % 8);
+  const uint8_t *payload = data + 2;
+  int payload_size = (int)(size - 2);
+
+  const char *create_sql = use_int8
+    ? "CREATE VIRTUAL TABLE v USING vec0("
+      "emb float[16] indexed by rescore(quantizer=int8))"
+    : "CREATE VIRTUAL TABLE v USING vec0("
+      "emb float[16] indexed by rescore(quantizer=bit))";
+
+  rc = sqlite3_exec(db, create_sql, NULL, NULL, NULL);
+  if (rc != SQLITE_OK) { sqlite3_close(db); return 0; }
+
+  /* Insert several vectors so there's a full chunk to corrupt */
+  {
+    sqlite3_stmt *ins = NULL;
+    sqlite3_prepare_v2(db,
+      "INSERT INTO v(rowid, emb) VALUES (?, ?)", -1, &ins, NULL);
+    if (!ins) { sqlite3_close(db); return 0; }
+
+    for (int i = 1; i <= 8; i++) {
+      float vec[16];
+      for (int j = 0; j < 16; j++) vec[j] = (float)(i * 10 + j) / 100.0f;
+      sqlite3_reset(ins);
+      sqlite3_bind_int64(ins, 1, i);
+      sqlite3_bind_blob(ins, 2, vec, sizeof(vec), SQLITE_TRANSIENT);
+      sqlite3_step(ins);
+    }
+    sqlite3_finalize(ins);
+  }
+
+  /* Now corrupt rescore shadow tables based on fuzz input */
+  sqlite3_stmt *stmt = NULL;
+
+  switch (target) {
+    case 0: {
+      /* Corrupt _rescore_chunks00 vectors blob with fuzz data */
+      rc = sqlite3_prepare_v2(db,
+        "UPDATE v_rescore_chunks00 SET vectors = ? WHERE rowid = 1",
+        -1, &stmt, NULL);
+      if (rc == SQLITE_OK) {
+        sqlite3_bind_blob(stmt, 1, payload, payload_size, SQLITE_STATIC);
+        sqlite3_step(stmt);
+        sqlite3_finalize(stmt);
+        stmt = NULL;
+      }
+      break;
+    }
+    case 1: {
+      /* Corrupt _rescore_vectors00 vector blob for a specific row */
+      rc = sqlite3_prepare_v2(db,
+        "UPDATE v_rescore_vectors00 SET vector = ? WHERE rowid = 3",
+        -1, &stmt, NULL);
+      if (rc == SQLITE_OK) {
+        sqlite3_bind_blob(stmt, 1, payload, payload_size, SQLITE_STATIC);
+        sqlite3_step(stmt);
+        sqlite3_finalize(stmt);
+        stmt = NULL;
+      }
+      break;
+    }
+    case 2: {
+      /* Truncate the quantized chunk blob to wrong size */
+      rc = sqlite3_prepare_v2(db,
+        "UPDATE v_rescore_chunks00 SET vectors = X'DEADBEEF' WHERE rowid = 1",
+        -1, &stmt, NULL);
+      if (rc == SQLITE_OK) {
+        sqlite3_step(stmt);
+        sqlite3_finalize(stmt);
+        stmt = NULL;
+      }
+      break;
+    }
+    case 3: {
+      /* Delete rows from _rescore_vectors (orphan the float vectors) */
+      sqlite3_exec(db,
+        "DELETE FROM v_rescore_vectors00 WHERE rowid IN (2, 4, 6)",
+        NULL, NULL, NULL);
+      break;
+    }
+    case 4: {
+      /* Delete the chunk row entirely from _rescore_chunks */
+      sqlite3_exec(db,
+        "DELETE FROM v_rescore_chunks00 WHERE rowid = 1",
+        NULL, NULL, NULL);
+      break;
+    }
+    case 5: {
+      /* Set vectors to NULL in _rescore_chunks */
+      sqlite3_exec(db,
+        "UPDATE v_rescore_chunks00 SET vectors = NULL WHERE rowid = 1",
+        NULL, NULL, NULL);
+      break;
+    }
+    case 6: {
+      /* Set vector to NULL in _rescore_vectors */
+      sqlite3_exec(db,
+        "UPDATE v_rescore_vectors00 SET vector = NULL WHERE rowid = 3",
+        NULL, NULL, NULL);
+      break;
+    }
+    case 7: {
+      /* Corrupt BOTH tables with fuzz data */
+      int half = payload_size / 2;
+      rc = sqlite3_prepare_v2(db,
+        "UPDATE v_rescore_chunks00 SET vectors = ? WHERE rowid = 1",
+        -1, &stmt, NULL);
+      if (rc == SQLITE_OK) {
+        sqlite3_bind_blob(stmt, 1, payload, half, SQLITE_STATIC);
+        sqlite3_step(stmt);
+        sqlite3_finalize(stmt);
+        stmt = NULL;
+      }
+      rc = sqlite3_prepare_v2(db,
+        "UPDATE v_rescore_vectors00 SET vector = ? WHERE rowid = 1",
+        -1, &stmt, NULL);
+      if (rc == SQLITE_OK) {
+        sqlite3_bind_blob(stmt, 1, payload + half,
+                          payload_size - half, SQLITE_STATIC);
+        sqlite3_step(stmt);
+        sqlite3_finalize(stmt);
+        stmt = NULL;
+      }
+      break;
+    }
+  }
+
+  /* Exercise ALL read/write paths -- NONE should crash */
+
+  /* KNN query (triggers rescore_knn Phase 1 + Phase 2) */
+  {
+    float qvec[16] = {1,0,0,0, 0,1,0,0, 0,0,1,0, 0,0,0,1};
+    sqlite3_stmt *knn = NULL;
+    sqlite3_prepare_v2(db,
+      "SELECT rowid, distance FROM v WHERE emb MATCH ? "
+      "ORDER BY distance LIMIT 5", -1, &knn, NULL);
+    if (knn) {
+      sqlite3_bind_blob(knn, 1, qvec, sizeof(qvec), SQLITE_STATIC);
+      while (sqlite3_step(knn) == SQLITE_ROW) {}
+      sqlite3_finalize(knn);
+    }
+  }
+
+  /* Full scan (triggers reading from _rescore_vectors) */
+  sqlite3_exec(db, "SELECT * FROM v", NULL, NULL, NULL);
+
+  /* Point lookups */
+  sqlite3_exec(db, "SELECT * FROM v WHERE rowid = 1", NULL, NULL, NULL);
+  sqlite3_exec(db, "SELECT * FROM v WHERE rowid = 3", NULL, NULL, NULL);
+
+  /* Insert after corruption */
+  {
+    float vec[16] = {0};
+    sqlite3_stmt *ins = NULL;
+    sqlite3_prepare_v2(db,
+      "INSERT INTO v(rowid, emb) VALUES (99, ?)", -1, &ins, NULL);
+    if (ins) {
+      sqlite3_bind_blob(ins, 1, vec, sizeof(vec), SQLITE_STATIC);
+      sqlite3_step(ins);
+      sqlite3_finalize(ins);
+    }
+  }
+
+  /* Delete after corruption */
+  sqlite3_exec(db, "DELETE FROM v WHERE rowid = 5", NULL, NULL, NULL);
+
+  /* Update after corruption */
+  {
+    float vec[16] = {1,1,1,1, 1,1,1,1, 1,1,1,1, 1,1,1,1};
+    sqlite3_stmt *upd = NULL;
+    sqlite3_prepare_v2(db,
+      "UPDATE v SET emb = ? WHERE rowid = 1", -1, &upd, NULL);
+    if (upd) {
+      sqlite3_bind_blob(upd, 1, vec, sizeof(vec), SQLITE_STATIC);
+      sqlite3_step(upd);
+      sqlite3_finalize(upd);
+    }
+  }
+
+  /* KNN again after modifications to corrupted state */
+  {
+    float qvec[16] = {0,0,0,0, 0,0,0,0, 1,1,1,1, 1,1,1,1};
+    sqlite3_stmt *knn = NULL;
+    sqlite3_prepare_v2(db,
+      "SELECT rowid, distance FROM v WHERE emb MATCH ? "
+      "ORDER BY distance LIMIT 3", -1, &knn, NULL);
+    if (knn) {
+      sqlite3_bind_blob(knn, 1, qvec, sizeof(qvec), SQLITE_STATIC);
+      while (sqlite3_step(knn) == SQLITE_ROW) {}
+      sqlite3_finalize(knn);
+    }
+  }
+
+  sqlite3_exec(db, "DROP TABLE v", NULL, NULL, NULL);
+  sqlite3_close(db);
+  return 0;
+}

tests/generate_legacy_db.py

@@ -0,0 +1,81 @@
+# /// script
+# requires-python = ">=3.10"
+# dependencies = ["sqlite-vec==0.1.6"]
+# ///
+"""Generate a legacy sqlite-vec database for backwards-compat testing.
+
+Usage:
+  uv run --script generate_legacy_db.py
+
+Creates tests/fixtures/legacy-v0.1.6.db with a vec0 table containing
+test data that can be read by the current version of sqlite-vec.
+"""
+import sqlite3
+import sqlite_vec
+import struct
+import os
+
+FIXTURE_DIR = os.path.join(os.path.dirname(__file__), "fixtures")
+DB_PATH = os.path.join(FIXTURE_DIR, "legacy-v0.1.6.db")
+
+DIMS = 4
+N_ROWS = 50
+
+
+def _f32(vals):
+    return struct.pack(f"{len(vals)}f", *vals)
+
+
+def main():
+    os.makedirs(FIXTURE_DIR, exist_ok=True)
+    if os.path.exists(DB_PATH):
+        os.remove(DB_PATH)
+
+    db = sqlite3.connect(DB_PATH)
+    db.enable_load_extension(True)
+    sqlite_vec.load(db)
+
+    # Print version for verification
+    version = db.execute("SELECT vec_version()").fetchone()[0]
+    print(f"sqlite-vec version: {version}")
+
+    # Create a basic vec0 table — flat index, no fancy features
+    db.execute(f"CREATE VIRTUAL TABLE legacy_vectors USING vec0(emb float[{DIMS}])")
+
+    # Insert test data: vectors where element[0] == rowid for easy verification
+    for i in range(1, N_ROWS + 1):
+        vec = [float(i), 0.0, 0.0, 0.0]
+        db.execute("INSERT INTO legacy_vectors(rowid, emb) VALUES (?, ?)", [i, _f32(vec)])
+
+    db.commit()
+
+    # Verify
+    count = db.execute("SELECT count(*) FROM legacy_vectors").fetchone()[0]
+    print(f"Inserted {count} rows")
+
+    # Test KNN works
+    query = _f32([1.0, 0.0, 0.0, 0.0])
+    rows = db.execute(
+        "SELECT rowid, distance FROM legacy_vectors WHERE emb MATCH ? AND k = 5",
+        [query],
+    ).fetchall()
+    print(f"KNN top 5: {[(r[0], round(r[1], 4)) for r in rows]}")
+    assert rows[0][0] == 1  # closest to [1,0,0,0]
+    assert len(rows) == 5
+
+    # Also create a table with name == column name (the conflict case)
+    # This was allowed in old versions — new code must not break on reconnect
+    db.execute("CREATE VIRTUAL TABLE emb USING vec0(emb float[4])")
+    for i in range(1, 11):
+        db.execute("INSERT INTO emb(rowid, emb) VALUES (?, ?)", [i, _f32([float(i), 0, 0, 0])])
+    db.commit()
+
+    count2 = db.execute("SELECT count(*) FROM emb").fetchone()[0]
+    print(f"Table 'emb' with column 'emb': {count2} rows (name conflict case)")
+
+    db.close()
+    print(f"\nGenerated: {DB_PATH}")
+
+
+if __name__ == "__main__":
+    main()

tests/sqlite-vec-internal.h

@@ -3,6 +3,11 @@
 
 #include <stdlib.h>
 #include <stddef.h>
+#include <stdint.h>
+
+#ifndef SQLITE_VEC_ENABLE_IVF
+#define SQLITE_VEC_ENABLE_IVF 1
+#endif
 
 int min_idx(
   const float *distances,
@@ -62,12 +67,81 @@ enum Vec0DistanceMetrics {
   VEC0_DISTANCE_METRIC_L1 = 3,
 };
 
+enum Vec0IndexType {
+  VEC0_INDEX_TYPE_FLAT = 1,
+#ifdef SQLITE_VEC_ENABLE_RESCORE
+  VEC0_INDEX_TYPE_RESCORE = 2,
+#endif
+  VEC0_INDEX_TYPE_IVF = 3,
+  VEC0_INDEX_TYPE_DISKANN = 4,
+};
+
+enum Vec0RescoreQuantizerType {
+  VEC0_RESCORE_QUANTIZER_BIT = 1,
+  VEC0_RESCORE_QUANTIZER_INT8 = 2,
+};
+
+struct Vec0RescoreConfig {
+  enum Vec0RescoreQuantizerType quantizer_type;
+  int oversample;
+};
+
+#if SQLITE_VEC_ENABLE_IVF
+enum Vec0IvfQuantizer {
+  VEC0_IVF_QUANTIZER_NONE = 0,
+  VEC0_IVF_QUANTIZER_INT8 = 1,
+  VEC0_IVF_QUANTIZER_BINARY = 2,
+};
+
+struct Vec0IvfConfig {
+  int nlist;
+  int nprobe;
+  int quantizer;
+  int oversample;
+};
+#else
+struct Vec0IvfConfig { char _unused; };
+#endif
+
+#ifdef SQLITE_VEC_ENABLE_RESCORE
+enum Vec0RescoreQuantizerType {
+  VEC0_RESCORE_QUANTIZER_BIT = 1,
+  VEC0_RESCORE_QUANTIZER_INT8 = 2,
+};
+
+struct Vec0RescoreConfig {
+  enum Vec0RescoreQuantizerType quantizer_type;
+  int oversample;
+};
+#endif
+
+enum Vec0DiskannQuantizerType {
+  VEC0_DISKANN_QUANTIZER_BINARY = 1,
+  VEC0_DISKANN_QUANTIZER_INT8   = 2,
+};
+
+struct Vec0DiskannConfig {
+  enum Vec0DiskannQuantizerType quantizer_type;
+  int n_neighbors;
+  int search_list_size;
+  int search_list_size_search;
+  int search_list_size_insert;
+  float alpha;
+  int buffer_threshold;
+};
+
 struct VectorColumnDefinition {
   char *name;
   int name_length;
   size_t dimensions;
   enum VectorElementType element_type;
   enum Vec0DistanceMetrics distance_metric;
+  enum Vec0IndexType index_type;
+#ifdef SQLITE_VEC_ENABLE_RESCORE
+  struct Vec0RescoreConfig rescore;
+#endif
+  struct Vec0IvfConfig ivf;
+  struct Vec0DiskannConfig diskann;
 };
 
 int vec0_parse_vector_column(const char *source, int source_length,
@@ -78,10 +152,90 @@ int vec0_parse_partition_key_definition(const char *source, int source_length,
                                         int *out_column_name_length,
                                         int *out_column_type);
 
+size_t diskann_quantized_vector_byte_size(
+    enum Vec0DiskannQuantizerType quantizer_type, size_t dimensions);
+
+int diskann_validity_byte_size(int n_neighbors);
+size_t diskann_neighbor_ids_byte_size(int n_neighbors);
+size_t diskann_neighbor_qvecs_byte_size(
+    int n_neighbors, enum Vec0DiskannQuantizerType quantizer_type,
+    size_t dimensions);
+int diskann_node_init(
+    int n_neighbors, enum Vec0DiskannQuantizerType quantizer_type,
+    size_t dimensions,
+    unsigned char **outValidity, int *outValiditySize,
+    unsigned char **outNeighborIds, int *outNeighborIdsSize,
+    unsigned char **outNeighborQvecs, int *outNeighborQvecsSize);
+int diskann_validity_get(const unsigned char *validity, int i);
+void diskann_validity_set(unsigned char *validity, int i, int value);
+int diskann_validity_count(const unsigned char *validity, int n_neighbors);
+long long diskann_neighbor_id_get(const unsigned char *neighbor_ids, int i);
+void diskann_neighbor_id_set(unsigned char *neighbor_ids, int i, long long rowid);
+const unsigned char *diskann_neighbor_qvec_get(
+    const unsigned char *qvecs, int i,
+    enum Vec0DiskannQuantizerType quantizer_type, size_t dimensions);
+void diskann_neighbor_qvec_set(
+    unsigned char *qvecs, int i, const unsigned char *src_qvec,
+    enum Vec0DiskannQuantizerType quantizer_type, size_t dimensions);
+void diskann_node_set_neighbor(
+    unsigned char *validity, unsigned char *neighbor_ids, unsigned char *qvecs, int i,
+    long long neighbor_rowid, const unsigned char *neighbor_qvec,
+    enum Vec0DiskannQuantizerType quantizer_type, size_t dimensions);
+void diskann_node_clear_neighbor(
+    unsigned char *validity, unsigned char *neighbor_ids, unsigned char *qvecs, int i,
+    enum Vec0DiskannQuantizerType quantizer_type, size_t dimensions);
+int diskann_quantize_vector(
+    const float *src, size_t dimensions,
+    enum Vec0DiskannQuantizerType quantizer_type,
+    unsigned char *out);
+
+int diskann_prune_select(
+    const float *inter_distances, const float *p_distances,
+    int num_candidates, float alpha, int max_neighbors,
+    int *outSelected, int *outCount);
+
 #ifdef SQLITE_VEC_TEST
 float _test_distance_l2_sqr_float(const float *a, const float *b, size_t dims);
 float _test_distance_cosine_float(const float *a, const float *b, size_t dims);
 float _test_distance_hamming(const unsigned char *a, const unsigned char *b, size_t dims);
+
+#ifdef SQLITE_VEC_ENABLE_RESCORE
+void _test_rescore_quantize_float_to_bit(const float *src, uint8_t *dst, size_t dim);
+void _test_rescore_quantize_float_to_int8(const float *src, int8_t *dst, size_t dim);
+size_t _test_rescore_quantized_byte_size_bit(size_t dimensions);
+size_t _test_rescore_quantized_byte_size_int8(size_t dimensions);
+#endif
+#if SQLITE_VEC_ENABLE_IVF
+void ivf_quantize_int8(const float *src, int8_t *dst, int D);
+void ivf_quantize_binary(const float *src, uint8_t *dst, int D);
+#endif
+// DiskANN candidate list (opaque struct, use accessors)
+struct DiskannCandidateList {
+  void *items;  // opaque
+  int count;
+  int capacity;
+};
+
+int _test_diskann_candidate_list_init(struct DiskannCandidateList *list, int capacity);
+void _test_diskann_candidate_list_free(struct DiskannCandidateList *list);
+int _test_diskann_candidate_list_insert(struct DiskannCandidateList *list, long long rowid, float distance);
+int _test_diskann_candidate_list_next_unvisited(const struct DiskannCandidateList *list);
+int _test_diskann_candidate_list_count(const struct DiskannCandidateList *list);
+long long _test_diskann_candidate_list_rowid(const struct DiskannCandidateList *list, int i);
+float _test_diskann_candidate_list_distance(const struct DiskannCandidateList *list, int i);
+void _test_diskann_candidate_list_set_visited(struct DiskannCandidateList *list, int i);
+
+// DiskANN visited set (opaque struct, use accessors)
+struct DiskannVisitedSet {
+  void *slots;  // opaque
+  int capacity;
+  int count;
+};
+
+int _test_diskann_visited_set_init(struct DiskannVisitedSet *set, int capacity);
+void _test_diskann_visited_set_free(struct DiskannVisitedSet *set);
+int _test_diskann_visited_set_contains(const struct DiskannVisitedSet *set, long long rowid);
+int _test_diskann_visited_set_insert(struct DiskannVisitedSet *set, long long rowid);
 #endif
 
 #endif /* SQLITE_VEC_INTERNAL_H */

tests/test-auxiliary.py

@@ -1,5 +1,7 @@
 import sqlite3
-from helpers import exec, vec0_shadow_table_contents
+import struct
+import pytest
+from helpers import exec, vec0_shadow_table_contents, _f32
 
 
 def test_constructor_limit(db, snapshot):
@@ -126,3 +128,198 @@ def test_knn(db, snapshot):
     ) == snapshot(name="illegal KNN w/ aux")
 
 
+# ======================================================================
+# Auxiliary columns with non-flat indexes
+# ======================================================================
+
+
+def test_rescore_aux_shadow_tables(db, snapshot):
+    """Rescore + aux column: verify shadow tables are created correctly."""
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0("
+        "  emb float[128] indexed by rescore(quantizer=bit),"
+        "  +label text,"
+        "  +score float"
+        ")"
+    )
+    assert exec(db, "SELECT name, sql FROM sqlite_master WHERE type='table' AND name LIKE 't_%' ORDER BY name") == snapshot(
+        name="rescore aux shadow tables"
+    )
+
+
+def test_rescore_aux_insert_knn(db, snapshot):
+    """Insert with aux data, KNN should return aux column values."""
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0("
+        "  emb float[128] indexed by rescore(quantizer=bit),"
+        "  +label text"
+        ")"
+    )
+    import random
+    random.seed(77)
+    data = [
+        ("alpha", [random.gauss(0, 1) for _ in range(128)]),
+        ("beta", [random.gauss(0, 1) for _ in range(128)]),
+        ("gamma", [random.gauss(0, 1) for _ in range(128)]),
+    ]
+    for label, vec in data:
+        db.execute(
+            "INSERT INTO t(emb, label) VALUES (?, ?)",
+            [_f32(vec), label],
+        )
+
+    assert exec(db, "SELECT rowid, label FROM t ORDER BY rowid") == snapshot(
+        name="rescore aux select all"
+    )
+    assert vec0_shadow_table_contents(db, "t", skip_info=True) == snapshot(
+        name="rescore aux shadow contents"
+    )
+
+    # KNN should include aux column, "alpha" closest to its own vector
+    rows = db.execute(
+        "SELECT label, distance FROM t WHERE emb MATCH ? ORDER BY distance LIMIT 3",
+        [_f32(data[0][1])],
+    ).fetchall()
+    assert len(rows) == 3
+    assert rows[0][0] == "alpha"
+
+
+def test_rescore_aux_update(db):
+    """UPDATE aux column on rescore table should work without affecting vectors."""
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0("
+        "  emb float[128] indexed by rescore(quantizer=bit),"
+        "  +label text"
+        ")"
+    )
+    import random
+    random.seed(88)
+    vec = [random.gauss(0, 1) for _ in range(128)]
+    db.execute("INSERT INTO t(rowid, emb, label) VALUES (1, ?, 'original')", [_f32(vec)])
+    db.execute("UPDATE t SET label = 'updated' WHERE rowid = 1")
+
+    assert db.execute("SELECT label FROM t WHERE rowid = 1").fetchone()[0] == "updated"
+
+    # KNN still works with updated aux
+    rows = db.execute(
+        "SELECT rowid, label FROM t WHERE emb MATCH ? ORDER BY distance LIMIT 1",
+        [_f32(vec)],
+    ).fetchall()
+    assert rows[0][0] == 1
+    assert rows[0][1] == "updated"
+
+
+def test_rescore_aux_delete(db, snapshot):
+    """DELETE should remove aux data from shadow table."""
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0("
+        "  emb float[128] indexed by rescore(quantizer=bit),"
+        "  +label text"
+        ")"
+    )
+    import random
+    random.seed(99)
+    for i in range(5):
+        db.execute(
+            "INSERT INTO t(rowid, emb, label) VALUES (?, ?, ?)",
+            [i + 1, _f32([random.gauss(0, 1) for _ in range(128)]), f"item-{i+1}"],
+        )
+
+    db.execute("DELETE FROM t WHERE rowid = 3")
+
+    assert exec(db, "SELECT rowid, label FROM t ORDER BY rowid") == snapshot(
+        name="rescore aux after delete"
+    )
+    assert exec(db, "SELECT rowid, value00 FROM t_auxiliary ORDER BY rowid") == snapshot(
+        name="rescore aux shadow after delete"
+    )
+
+
+def test_diskann_aux_shadow_tables(db, snapshot):
+    """DiskANN + aux column: verify shadow tables are created correctly."""
+    db.execute("""
+        CREATE VIRTUAL TABLE t USING vec0(
+            emb float[8] INDEXED BY diskann(neighbor_quantizer=binary, n_neighbors=8),
+            +label text,
+            +score float
+        )
+    """)
+    assert exec(db, "SELECT name, sql FROM sqlite_master WHERE type='table' AND name LIKE 't_%' ORDER BY name") == snapshot(
+        name="diskann aux shadow tables"
+    )
+
+
+def test_diskann_aux_insert_knn(db, snapshot):
+    """DiskANN + aux: insert, KNN, verify aux values returned."""
+    db.execute("""
+        CREATE VIRTUAL TABLE t USING vec0(
+            emb float[8] INDEXED BY diskann(neighbor_quantizer=binary, n_neighbors=8),
+            +label text
+        )
+    """)
+    data = [
+        ("red", [1, 0, 0, 0, 0, 0, 0, 0]),
+        ("green", [0, 1, 0, 0, 0, 0, 0, 0]),
+        ("blue", [0, 0, 1, 0, 0, 0, 0, 0]),
+    ]
+    for label, vec in data:
+        db.execute("INSERT INTO t(emb, label) VALUES (?, ?)", [_f32(vec), label])
+
+    assert exec(db, "SELECT rowid, label FROM t ORDER BY rowid") == snapshot(
+        name="diskann aux select all"
+    )
+    assert vec0_shadow_table_contents(db, "t", skip_info=True) == snapshot(
+        name="diskann aux shadow contents"
+    )
+
+    rows = db.execute(
+        "SELECT label, distance FROM t WHERE emb MATCH ? AND k = 3",
+        [_f32([1, 0, 0, 0, 0, 0, 0, 0])],
+    ).fetchall()
+    assert len(rows) >= 1
+    assert rows[0][0] == "red"
+
+
+def test_diskann_aux_update_and_delete(db, snapshot):
+    """DiskANN + aux: update aux column, delete row, verify cleanup."""
+    db.execute("""
+        CREATE VIRTUAL TABLE t USING vec0(
+            emb float[8] INDEXED BY diskann(neighbor_quantizer=binary, n_neighbors=8),
+            +label text
+        )
+    """)
+    for i in range(5):
+        vec = [0.0] * 8
+        vec[i % 8] = 1.0
+        db.execute(
+            "INSERT INTO t(rowid, emb, label) VALUES (?, ?, ?)",
+            [i + 1, _f32(vec), f"item-{i+1}"],
+        )
+
+    db.execute("UPDATE t SET label = 'UPDATED' WHERE rowid = 2")
+    db.execute("DELETE FROM t WHERE rowid = 3")
+
+    assert exec(db, "SELECT rowid, label FROM t ORDER BY rowid") == snapshot(
+        name="diskann aux after update+delete"
+    )
+    assert exec(db, "SELECT rowid, value00 FROM t_auxiliary ORDER BY rowid") == snapshot(
+        name="diskann aux shadow after update+delete"
+    )
+
+
+def test_diskann_aux_drop_cleans_all(db):
+    """DROP TABLE should remove aux shadow table too."""
+    db.execute("""
+        CREATE VIRTUAL TABLE t USING vec0(
+            emb float[8] INDEXED BY diskann(neighbor_quantizer=binary, n_neighbors=8),
+            +label text
+        )
+    """)
+    db.execute("INSERT INTO t(emb, label) VALUES (?, 'test')", [_f32([1]*8)])
+    db.execute("DROP TABLE t")
+
+    tables = [r[0] for r in db.execute(
+        "SELECT name FROM sqlite_master WHERE name LIKE 't_%'"
+    ).fetchall()]
+    assert "t_auxiliary" not in tables
+

tests/test-general.py

@@ -27,3 +27,15 @@ def test_info(db, snapshot):
     assert exec(db, "select key, typeof(value) from v_info order by 1") == snapshot()
 
 
+def test_command_column_name_conflict(db):
+    """Table name matching a column name should error (command column conflict)."""
+    # This would conflict: hidden command column 'embeddings' vs vector column 'embeddings'
+    with pytest.raises(sqlite3.OperationalError, match="conflicts with table name"):
+        db.execute(
+            "create virtual table embeddings using vec0(embeddings float[4])"
+        )
+
+    # Different names should work fine
+    db.execute("create virtual table t using vec0(embeddings float[4])")
+
+

tests/test-insert-delete.py

@@ -483,3 +483,171 @@ def test_delete_one_chunk_of_two_shrinks_pages(tmp_path):
         row = db.execute("select emb from v where rowid = ?", [i]).fetchone()
         assert row[0] == _f32([float(i)] * dims)
     db.close()
+
+
+def test_wal_concurrent_reader_during_write(tmp_path):
+    """In WAL mode, a reader should see a consistent snapshot while a writer inserts."""
+    dims = 4
+    db_path = str(tmp_path / "test.db")
+
+    # Writer: create table, insert initial rows, enable WAL
+    writer = sqlite3.connect(db_path)
+    writer.enable_load_extension(True)
+    writer.load_extension("dist/vec0")
+    writer.execute("PRAGMA journal_mode=WAL")
+    writer.execute(
+        f"CREATE VIRTUAL TABLE v USING vec0(emb float[{dims}])"
+    )
+    for i in range(1, 11):
+        writer.execute("INSERT INTO v(rowid, emb) VALUES (?, ?)", [i, _f32([float(i)] * dims)])
+    writer.commit()
+
+    # Reader: open separate connection, start read
+    reader = sqlite3.connect(db_path)
+    reader.enable_load_extension(True)
+    reader.load_extension("dist/vec0")
+
+    # Reader sees 10 rows
+    count_before = reader.execute("SELECT count(*) FROM v").fetchone()[0]
+    assert count_before == 10
+
+    # Writer inserts more rows (not yet committed)
+    writer.execute("BEGIN")
+    for i in range(11, 21):
+        writer.execute("INSERT INTO v(rowid, emb) VALUES (?, ?)", [i, _f32([float(i)] * dims)])
+
+    # Reader still sees 10 (WAL snapshot isolation)
+    count_during = reader.execute("SELECT count(*) FROM v").fetchone()[0]
+    assert count_during == 10
+
+    # KNN during writer's transaction should work on reader's snapshot
+    rows = reader.execute(
+        "SELECT rowid FROM v WHERE emb MATCH ? AND k = 5",
+        [_f32([1.0] * dims)],
+    ).fetchall()
+    assert len(rows) == 5
+    assert all(r[0] <= 10 for r in rows)  # only original rows
+
+    # Writer commits
+    writer.commit()
+
+    # Reader sees new rows after re-query (new snapshot)
+    count_after = reader.execute("SELECT count(*) FROM v").fetchone()[0]
+    assert count_after == 20
+
+    writer.close()
+    reader.close()
+
+
+def test_insert_or_replace_integer_pk(db):
+    """INSERT OR REPLACE should update vector when rowid already exists."""
+    db.execute("create virtual table v using vec0(emb float[4], chunk_size=8)")
+
+    db.execute(
+        "insert into v(rowid, emb) values (1, ?)", [_f32([1.0, 2.0, 3.0, 4.0])]
+    )
+    # Replace with new vector
+    db.execute(
+        "insert or replace into v(rowid, emb) values (1, ?)",
+        [_f32([10.0, 20.0, 30.0, 40.0])],
+    )
+
+    # Should still have exactly 1 row
+    count = db.execute("select count(*) from v").fetchone()[0]
+    assert count == 1
+
+    # Vector should be the replaced value
+    row = db.execute("select emb from v where rowid = 1").fetchone()
+    assert row[0] == _f32([10.0, 20.0, 30.0, 40.0])
+
+
+def test_insert_or_replace_new_row(db):
+    """INSERT OR REPLACE with a new rowid should just insert normally."""
+    db.execute("create virtual table v using vec0(emb float[4], chunk_size=8)")
+
+    db.execute(
+        "insert or replace into v(rowid, emb) values (1, ?)",
+        [_f32([1.0, 2.0, 3.0, 4.0])],
+    )
+
+    count = db.execute("select count(*) from v").fetchone()[0]
+    assert count == 1
+
+    row = db.execute("select emb from v where rowid = 1").fetchone()
+    assert row[0] == _f32([1.0, 2.0, 3.0, 4.0])
+
+
+def test_insert_or_replace_text_pk(db):
+    """INSERT OR REPLACE should work with text primary keys."""
+    db.execute(
+        "create virtual table v using vec0("
+        "id text primary key, emb float[4], chunk_size=8"
+        ")"
+    )
+
+    db.execute(
+        "insert into v(id, emb) values ('doc_a', ?)",
+        [_f32([1.0, 2.0, 3.0, 4.0])],
+    )
+    db.execute(
+        "insert or replace into v(id, emb) values ('doc_a', ?)",
+        [_f32([10.0, 20.0, 30.0, 40.0])],
+    )
+
+    count = db.execute("select count(*) from v").fetchone()[0]
+    assert count == 1
+
+    row = db.execute("select emb from v where id = 'doc_a'").fetchone()
+    assert row[0] == _f32([10.0, 20.0, 30.0, 40.0])
+
+
+def test_insert_or_replace_with_auxiliary(db):
+    """INSERT OR REPLACE should also replace auxiliary column values."""
+    db.execute(
+        "create virtual table v using vec0("
+        "emb float[4], +label text, chunk_size=8"
+        ")"
+    )
+
+    db.execute(
+        "insert into v(rowid, emb, label) values (1, ?, 'old')",
+        [_f32([1.0, 2.0, 3.0, 4.0])],
+    )
+    db.execute(
+        "insert or replace into v(rowid, emb, label) values (1, ?, 'new')",
+        [_f32([10.0, 20.0, 30.0, 40.0])],
+    )
+
+    count = db.execute("select count(*) from v").fetchone()[0]
+    assert count == 1
+
+    row = db.execute("select emb, label from v where rowid = 1").fetchone()
+    assert row[0] == _f32([10.0, 20.0, 30.0, 40.0])
+    assert row[1] == "new"
+
+
+def test_insert_or_replace_knn_uses_new_vector(db):
+    """After INSERT OR REPLACE, KNN should find the new vector, not the old one."""
+    db.execute("create virtual table v using vec0(emb float[4], chunk_size=8)")
+
+    db.execute(
+        "insert into v(rowid, emb) values (1, ?)", [_f32([1.0, 0.0, 0.0, 0.0])]
+    )
+    db.execute(
+        "insert into v(rowid, emb) values (2, ?)", [_f32([0.0, 1.0, 0.0, 0.0])]
+    )
+
+    # Replace row 1's vector to be very close to row 2
+    db.execute(
+        "insert or replace into v(rowid, emb) values (1, ?)",
+        [_f32([0.0, 0.9, 0.0, 0.0])],
+    )
+
+    # KNN for [0, 1, 0, 0] should return row 2 first (exact), then row 1 (close)
+    rows = db.execute(
+        "select rowid, distance from v where emb match ? and k = 2",
+        [_f32([0.0, 1.0, 0.0, 0.0])],
+    ).fetchall()
+    assert rows[0][0] == 2
+    assert rows[1][0] == 1
+    assert rows[1][1] < 0.11  # should be close (L2 distance ≈ 0.1)

tests/test-ivf-mutations.py

@@ -0,0 +1,589 @@
+"""
+Thorough IVF mutation tests: insert, delete, update, KNN correctness,
+error cases, edge cases, and cell overflow scenarios.
+"""
+import pytest
+import sqlite3
+import struct
+import math
+from helpers import _f32, exec
+
+
+@pytest.fixture()
+def db():
+    db = sqlite3.connect(":memory:")
+    db.row_factory = sqlite3.Row
+    db.enable_load_extension(True)
+    db.load_extension("dist/vec0")
+    db.enable_load_extension(False)
+    return db
+
+
+def ivf_total_vectors(db, table="t", col=0):
+    """Count total vectors across all IVF cells."""
+    return db.execute(
+        f"SELECT COALESCE(SUM(n_vectors), 0) FROM {table}_ivf_cells{col:02d}"
+    ).fetchone()[0]
+
+
+def ivf_unassigned_count(db, table="t", col=0):
+    return db.execute(
+        f"SELECT COALESCE(SUM(n_vectors), 0) FROM {table}_ivf_cells{col:02d} WHERE centroid_id = -1"
+    ).fetchone()[0]
+
+
+def ivf_assigned_count(db, table="t", col=0):
+    return db.execute(
+        f"SELECT COALESCE(SUM(n_vectors), 0) FROM {table}_ivf_cells{col:02d} WHERE centroid_id >= 0"
+    ).fetchone()[0]
+
+
+def knn(db, query, k, table="t", col="v"):
+    """Run a KNN query and return list of (rowid, distance) tuples."""
+    rows = db.execute(
+        f"SELECT rowid, distance FROM {table} WHERE {col} MATCH ? AND k = ?",
+        [_f32(query), k],
+    ).fetchall()
+    return [(r[0], r[1]) for r in rows]
+
+
+# ============================================================================
+# Single row insert + KNN
+# ============================================================================
+
+
+def test_insert_single_row_knn(db):
+    db.execute("CREATE VIRTUAL TABLE t USING vec0(v float[4] indexed by ivf())")
+    db.execute("INSERT INTO t(rowid, v) VALUES (1, ?)", [_f32([1, 0, 0, 0])])
+    results = knn(db, [1, 0, 0, 0], 5)
+    assert len(results) == 1
+    assert results[0][0] == 1
+    assert results[0][1] < 0.001
+
+
+# ============================================================================
+# Batch insert + KNN recall
+# ============================================================================
+
+
+def test_batch_insert_knn_recall(db):
+    """Insert 200 vectors, train, verify KNN recall with nprobe=nlist."""
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0(v float[4] indexed by ivf(nlist=8, nprobe=8))"
+    )
+    for i in range(200):
+        db.execute(
+            "INSERT INTO t(rowid, v) VALUES (?, ?)",
+            [i, _f32([float(i), 0, 0, 0])],
+        )
+    assert ivf_total_vectors(db) == 200
+
+    db.execute("INSERT INTO t(t) VALUES ('compute-centroids')")
+    assert ivf_assigned_count(db) == 200
+
+    # Query near 100 -- closest should be rowid 100
+    results = knn(db, [100.0, 0, 0, 0], 10)
+    assert len(results) == 10
+    assert results[0][0] == 100
+    assert results[0][1] < 0.01
+
+    # All results should be near 100
+    rowids = {r[0] for r in results}
+    assert all(95 <= r <= 105 for r in rowids)
+
+
+# ============================================================================
+# Delete rows, verify they're gone from KNN
+# ============================================================================
+
+
+def test_delete_rows_gone_from_knn(db):
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0(v float[4] indexed by ivf(nlist=2, nprobe=2))"
+    )
+    for i in range(20):
+        db.execute(
+            "INSERT INTO t(rowid, v) VALUES (?, ?)",
+            [i, _f32([float(i), 0, 0, 0])],
+        )
+
+    db.execute("INSERT INTO t(t) VALUES ('compute-centroids')")
+
+    # Delete rowid 10
+    db.execute("DELETE FROM t WHERE rowid = 10")
+
+    results = knn(db, [10.0, 0, 0, 0], 20)
+    rowids = {r[0] for r in results}
+    assert 10 not in rowids
+
+
+def test_delete_all_rows_empty_results(db):
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0(v float[4] indexed by ivf(nlist=2, nprobe=2))"
+    )
+    for i in range(10):
+        db.execute(
+            "INSERT INTO t(rowid, v) VALUES (?, ?)",
+            [i, _f32([float(i), 0, 0, 0])],
+        )
+
+    db.execute("INSERT INTO t(t) VALUES ('compute-centroids')")
+
+    for i in range(10):
+        db.execute("DELETE FROM t WHERE rowid = ?", [i])
+
+    assert ivf_total_vectors(db) == 0
+    results = knn(db, [5.0, 0, 0, 0], 10)
+    assert len(results) == 0
+
+
+# ============================================================================
+# Insert after delete (reuse rowids)
+# ============================================================================
+
+
+def test_insert_after_delete_reuse_rowid(db):
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0(v float[4] indexed by ivf(nlist=2, nprobe=2))"
+    )
+    for i in range(10):
+        db.execute(
+            "INSERT INTO t(rowid, v) VALUES (?, ?)",
+            [i, _f32([float(i), 0, 0, 0])],
+        )
+
+    db.execute("INSERT INTO t(t) VALUES ('compute-centroids')")
+
+    # Delete rowid 5
+    db.execute("DELETE FROM t WHERE rowid = 5")
+
+    # Re-insert rowid 5 with a very different vector
+    db.execute(
+        "INSERT INTO t(rowid, v) VALUES (5, ?)", [_f32([999.0, 0, 0, 0])]
+    )
+
+    # KNN near 999 should find rowid 5
+    results = knn(db, [999.0, 0, 0, 0], 1)
+    assert len(results) >= 1
+    assert results[0][0] == 5
+
+
+# ============================================================================
+# Update vectors (INSERT OR REPLACE), verify KNN reflects new values
+# ============================================================================
+
+
+def test_update_vector_via_delete_insert(db):
+    """vec0 IVF update: delete then re-insert with new vector."""
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0(v float[4] indexed by ivf(nlist=2, nprobe=2))"
+    )
+    for i in range(10):
+        db.execute(
+            "INSERT INTO t(rowid, v) VALUES (?, ?)",
+            [i, _f32([float(i), 0, 0, 0])],
+        )
+
+    db.execute("INSERT INTO t(t) VALUES ('compute-centroids')")
+
+    # "Update" rowid 3: delete and re-insert with new vector
+    db.execute("DELETE FROM t WHERE rowid = 3")
+    db.execute(
+        "INSERT INTO t(rowid, v) VALUES (3, ?)",
+        [_f32([100.0, 0, 0, 0])],
+    )
+
+    # KNN near 100 should find rowid 3
+    results = knn(db, [100.0, 0, 0, 0], 1)
+    assert results[0][0] == 3
+
+
+# ============================================================================
+# Error cases: IVF + auxiliary/metadata/partition key columns
+# ============================================================================
+
+
+def test_ivf_with_auxiliary_column(db):
+    """IVF should support auxiliary columns."""
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0(v float[4] indexed by ivf(), +extra text)"
+    )
+    tables = [r[0] for r in db.execute(
+        "SELECT name FROM sqlite_master WHERE name LIKE 't_%' ORDER BY 1"
+    ).fetchall()]
+    assert "t_auxiliary" in tables
+
+
+def test_error_ivf_with_metadata_column(db):
+    result = exec(
+        db,
+        "CREATE VIRTUAL TABLE t USING vec0(v float[4] indexed by ivf(), genre text)",
+    )
+    assert "error" in result
+    assert "metadata" in result.get("message", "").lower()
+
+
+def test_error_ivf_with_partition_key(db):
+    result = exec(
+        db,
+        "CREATE VIRTUAL TABLE t USING vec0(v float[4] indexed by ivf(), user_id integer partition key)",
+    )
+    assert "error" in result
+    assert "partition" in result.get("message", "").lower()
+
+
+def test_flat_with_auxiliary_still_works(db):
+    """Regression guard: flat-indexed tables with aux columns should still work."""
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0(v float[4], +extra text)"
+    )
+    db.execute(
+        "INSERT INTO t(rowid, v, extra) VALUES (1, ?, 'hello')",
+        [_f32([1, 0, 0, 0])],
+    )
+    row = db.execute("SELECT extra FROM t WHERE rowid = 1").fetchone()
+    assert row[0] == "hello"
+
+
+def test_flat_with_metadata_still_works(db):
+    """Regression guard: flat-indexed tables with metadata columns should still work."""
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0(v float[4], genre text)"
+    )
+    db.execute(
+        "INSERT INTO t(rowid, v, genre) VALUES (1, ?, 'rock')",
+        [_f32([1, 0, 0, 0])],
+    )
+    row = db.execute("SELECT genre FROM t WHERE rowid = 1").fetchone()
+    assert row[0] == "rock"
+
+
+def test_flat_with_partition_key_still_works(db):
+    """Regression guard: flat-indexed tables with partition key should still work."""
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0(v float[4], user_id integer partition key)"
+    )
+    db.execute(
+        "INSERT INTO t(rowid, v, user_id) VALUES (1, ?, 42)",
+        [_f32([1, 0, 0, 0])],
+    )
+    row = db.execute("SELECT user_id FROM t WHERE rowid = 1").fetchone()
+    assert row[0] == 42
+
+
+# ============================================================================
+# Edge cases
+# ============================================================================
+
+
+def test_zero_vectors(db):
+    """Insert zero vectors, verify KNN still works."""
+    db.execute("CREATE VIRTUAL TABLE t USING vec0(v float[4] indexed by ivf())")
+    for i in range(5):
+        db.execute(
+            "INSERT INTO t(rowid, v) VALUES (?, ?)",
+            [i, _f32([0, 0, 0, 0])],
+        )
+    results = knn(db, [0, 0, 0, 0], 5)
+    assert len(results) == 5
+    # All distances should be 0
+    for _, dist in results:
+        assert dist < 0.001
+
+
+def test_large_values(db):
+    """Insert vectors with very large and small values."""
+    db.execute("CREATE VIRTUAL TABLE t USING vec0(v float[4] indexed by ivf())")
+    db.execute(
+        "INSERT INTO t(rowid, v) VALUES (1, ?)", [_f32([1e6, 1e6, 1e6, 1e6])]
+    )
+    db.execute(
+        "INSERT INTO t(rowid, v) VALUES (2, ?)", [_f32([1e-6, 1e-6, 1e-6, 1e-6])]
+    )
+    db.execute(
+        "INSERT INTO t(rowid, v) VALUES (3, ?)", [_f32([-1e6, -1e6, -1e6, -1e6])]
+    )
+
+    results = knn(db, [1e6, 1e6, 1e6, 1e6], 3)
+    assert results[0][0] == 1
+
+
+def test_single_row_compute_centroids(db):
+    """Single row table, compute-centroids should still work."""
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0(v float[4] indexed by ivf(nlist=1))"
+    )
+    db.execute(
+        "INSERT INTO t(rowid, v) VALUES (1, ?)", [_f32([1, 2, 3, 4])]
+    )
+    db.execute("INSERT INTO t(t) VALUES ('compute-centroids')")
+    assert ivf_assigned_count(db) == 1
+
+    results = knn(db, [1, 2, 3, 4], 1)
+    assert len(results) == 1
+    assert results[0][0] == 1
+
+
+# ============================================================================
+# Cell overflow (many vectors in one cell)
+# ============================================================================
+
+
+def test_cell_overflow_many_vectors(db):
+    """Insert >64 vectors that all go to same centroid. Should create multiple cells."""
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0(v float[4] indexed by ivf(nlist=0))"
+    )
+    # Insert 100 very similar vectors
+    for i in range(100):
+        db.execute(
+            "INSERT INTO t(rowid, v) VALUES (?, ?)",
+            [i, _f32([1.0 + i * 0.001, 0, 0, 0])],
+        )
+
+    # Set a single centroid so all vectors go there
+    db.execute(
+        "INSERT INTO t(t, v) VALUES ('set-centroid:0', ?)",
+        [_f32([1.0, 0, 0, 0])],
+    )
+    db.execute("INSERT INTO t(t) VALUES ('assign-vectors')")
+
+    assert ivf_assigned_count(db) == 100
+
+    # Should have more than 1 cell (64 max per cell)
+    cell_count = db.execute(
+        "SELECT count(*) FROM t_ivf_cells00 WHERE centroid_id = 0"
+    ).fetchone()[0]
+    assert cell_count >= 2  # 100 / 64 = 2 cells needed
+
+    # All vectors should be queryable
+    results = knn(db, [1.0, 0, 0, 0], 100)
+    assert len(results) == 100
+
+
+# ============================================================================
+# Large batch with training
+# ============================================================================
+
+
+def test_large_batch_with_training(db):
+    """Insert 500, train, insert 500 more, verify total is 1000."""
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0(v float[4] indexed by ivf(nlist=16, nprobe=16))"
+    )
+    for i in range(500):
+        db.execute(
+            "INSERT INTO t(rowid, v) VALUES (?, ?)",
+            [i, _f32([float(i), 0, 0, 0])],
+        )
+
+    db.execute("INSERT INTO t(t) VALUES ('compute-centroids')")
+
+    for i in range(500, 1000):
+        db.execute(
+            "INSERT INTO t(rowid, v) VALUES (?, ?)",
+            [i, _f32([float(i), 0, 0, 0])],
+        )
+
+    assert ivf_total_vectors(db) == 1000
+
+    # KNN should still work
+    results = knn(db, [750.0, 0, 0, 0], 5)
+    assert len(results) == 5
+    assert results[0][0] == 750
+
+
+# ============================================================================
+# KNN after interleaved insert/delete
+# ============================================================================
+
+
+def test_knn_after_interleaved_insert_delete(db):
+    """Insert 20, train, delete 10 closest to query, verify remaining."""
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0(v float[4] indexed by ivf(nlist=4, nprobe=4))"
+    )
+    for i in range(20):
+        db.execute(
+            "INSERT INTO t(rowid, v) VALUES (?, ?)",
+            [i, _f32([float(i), 0, 0, 0])],
+        )
+
+    db.execute("INSERT INTO t(t) VALUES ('compute-centroids')")
+
+    # Delete rowids 0-9 (closest to query at 5.0)
+    for i in range(10):
+        db.execute("DELETE FROM t WHERE rowid = ?", [i])
+
+    results = knn(db, [5.0, 0, 0, 0], 10)
+    rowids = {r[0] for r in results}
+    # None of the deleted rowids should appear
+    assert all(r >= 10 for r in rowids)
+    assert len(results) == 10
+
+
+def test_knn_empty_centroids_after_deletes(db):
+    """Some centroids may become empty after deletes. Should not crash."""
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0(v float[4] indexed by ivf(nlist=4, nprobe=2))"
+    )
+    # Insert vectors clustered near 4 points
+    for i in range(40):
+        db.execute(
+            "INSERT INTO t(rowid, v) VALUES (?, ?)",
+            [i, _f32([float(i % 10) * 10, 0, 0, 0])],
+        )
+
+    db.execute("INSERT INTO t(t) VALUES ('compute-centroids')")
+
+    # Delete a bunch, potentially emptying some centroids
+    for i in range(30):
+        db.execute("DELETE FROM t WHERE rowid = ?", [i])
+
+    # Should not crash even with empty centroids
+    results = knn(db, [50.0, 0, 0, 0], 20)
+    assert len(results) <= 10  # only 10 left
+
+
+# ============================================================================
+# KNN returns correct distances
+# ============================================================================
+
+
+def test_knn_correct_distances(db):
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0(v float[4] indexed by ivf(nlist=2, nprobe=2))"
+    )
+    db.execute("INSERT INTO t(rowid, v) VALUES (1, ?)", [_f32([0, 0, 0, 0])])
+    db.execute("INSERT INTO t(rowid, v) VALUES (2, ?)", [_f32([3, 0, 0, 0])])
+    db.execute("INSERT INTO t(rowid, v) VALUES (3, ?)", [_f32([0, 4, 0, 0])])
+
+    db.execute("INSERT INTO t(t) VALUES ('compute-centroids')")
+
+    results = knn(db, [0, 0, 0, 0], 3)
+    result_map = {r[0]: r[1] for r in results}
+
+    # L2 distances (sqrt of sum of squared differences)
+    assert abs(result_map[1] - 0.0) < 0.01
+    assert abs(result_map[2] - 3.0) < 0.01   # sqrt(3^2) = 3
+    assert abs(result_map[3] - 4.0) < 0.01   # sqrt(4^2) = 4
+
+
+# ============================================================================
+# Delete in flat mode leaves no orphan rowid_map entries
+# ============================================================================
+
+
+def test_delete_flat_mode_rowid_map_count(db):
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0(v float[4] indexed by ivf(nlist=4))"
+    )
+    for i in range(5):
+        db.execute(
+            "INSERT INTO t(rowid, v) VALUES (?, ?)",
+            [i, _f32([float(i), 0, 0, 0])],
+        )
+
+    db.execute("DELETE FROM t WHERE rowid = 0")
+    db.execute("DELETE FROM t WHERE rowid = 2")
+    db.execute("DELETE FROM t WHERE rowid = 4")
+
+    assert db.execute("SELECT count(*) FROM t_ivf_rowid_map00").fetchone()[0] == 2
+    assert ivf_unassigned_count(db) == 2
+
+
+# ============================================================================
+# Duplicate rowid insert
+# ============================================================================
+
+
+def test_delete_reinsert_as_update(db):
+    """Simulate update via delete + insert on same rowid."""
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0(v float[4] indexed by ivf(nlist=2, nprobe=2))"
+    )
+    db.execute("INSERT INTO t(rowid, v) VALUES (1, ?)", [_f32([1, 0, 0, 0])])
+
+    # Delete then re-insert as "update"
+    db.execute("DELETE FROM t WHERE rowid = 1")
+    db.execute("INSERT INTO t(rowid, v) VALUES (1, ?)", [_f32([99, 0, 0, 0])])
+
+    results = knn(db, [99, 0, 0, 0], 1)
+    assert len(results) == 1
+    assert results[0][0] == 1
+    assert results[0][1] < 0.01
+
+
+def test_duplicate_rowid_insert_fails(db):
+    """Inserting a duplicate rowid should fail with a constraint error."""
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0(v float[4] indexed by ivf(nlist=2, nprobe=2))"
+    )
+    db.execute("INSERT INTO t(rowid, v) VALUES (1, ?)", [_f32([1, 0, 0, 0])])
+
+    result = exec(
+        db,
+        "INSERT INTO t(rowid, v) VALUES (1, ?)",
+        [_f32([99, 0, 0, 0])],
+    )
+    assert "error" in result
+
+
+# ============================================================================
+# Interleaved insert/delete with KNN correctness
+# ============================================================================
+
+
+def test_interleaved_ops_correctness(db):
+    """Complex sequence of inserts and deletes, verify KNN is always correct."""
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0(v float[4] indexed by ivf(nlist=4, nprobe=4))"
+    )
+
+    # Phase 1: Insert 50 vectors
+    for i in range(50):
+        db.execute(
+            "INSERT INTO t(rowid, v) VALUES (?, ?)",
+            [i, _f32([float(i), 0, 0, 0])],
+        )
+
+    db.execute("INSERT INTO t(t) VALUES ('compute-centroids')")
+
+    # Phase 2: Delete even-numbered rowids
+    for i in range(0, 50, 2):
+        db.execute("DELETE FROM t WHERE rowid = ?", [i])
+
+    # Phase 3: Insert new vectors with higher rowids
+    for i in range(50, 75):
+        db.execute(
+            "INSERT INTO t(rowid, v) VALUES (?, ?)",
+            [i, _f32([float(i), 0, 0, 0])],
+        )
+
+    # Phase 4: Delete some of the new ones
+    for i in range(60, 70):
+        db.execute("DELETE FROM t WHERE rowid = ?", [i])
+
+    # KNN query: should only find existing vectors
+    results = knn(db, [25.0, 0, 0, 0], 50)
+    rowids = {r[0] for r in results}
+
+    # Verify no deleted rowids appear
+    deleted = set(range(0, 50, 2)) | set(range(60, 70))
+    assert len(rowids & deleted) == 0
+
+    # Verify we get the right count (25 odd + 15 new - 10 deleted new = 30)
+    expected_alive = set(range(1, 50, 2)) | set(range(50, 60)) | set(range(70, 75))
+    assert rowids.issubset(expected_alive)
+
+
+def test_ivf_update_vector_blocked(db):
+    """UPDATE on a vector column with IVF index should error (index would become stale)."""
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0(emb float[4] indexed by ivf(nlist=2))"
+    )
+    db.execute("INSERT INTO t(rowid, emb) VALUES (1, ?)", [_f32([1, 0, 0, 0])])
+    db.execute("INSERT INTO t(rowid, emb) VALUES (2, ?)", [_f32([0, 1, 0, 0])])
+
+    with pytest.raises(sqlite3.OperationalError, match="UPDATE on vector column.*not supported for IVF"):
+        db.execute("UPDATE t SET emb = ? WHERE rowid = 1", [_f32([0, 0, 1, 0])])

tests/test-ivf-quantization.py

@@ -0,0 +1,272 @@
+import pytest
+import sqlite3
+from helpers import _f32, exec
+
+
+@pytest.fixture()
+def db():
+    db = sqlite3.connect(":memory:")
+    db.row_factory = sqlite3.Row
+    db.enable_load_extension(True)
+    db.load_extension("dist/vec0")
+    db.enable_load_extension(False)
+    return db
+
+
+# ============================================================================
+# Parser tests — quantizer and oversample options
+# ============================================================================
+
+
+def test_ivf_quantizer_binary(db):
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0("
+        "v float[768] indexed by ivf(nlist=64, quantizer=binary, oversample=10))"
+    )
+    tables = [
+        r[0]
+        for r in db.execute(
+            "SELECT name FROM sqlite_master WHERE type='table' ORDER BY 1"
+        ).fetchall()
+    ]
+    assert "t_ivf_centroids00" in tables
+    assert "t_ivf_cells00" in tables
+
+
+def test_ivf_quantizer_int8(db):
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0("
+        "v float[768] indexed by ivf(nlist=64, quantizer=int8))"
+    )
+    tables = [
+        r[0]
+        for r in db.execute(
+            "SELECT name FROM sqlite_master WHERE type='table' ORDER BY 1"
+        ).fetchall()
+    ]
+    assert "t_ivf_centroids00" in tables
+
+
+def test_ivf_quantizer_none_explicit(db):
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0("
+        "v float[768] indexed by ivf(quantizer=none))"
+    )
+    # Should work — same as no quantizer
+    tables = [
+        r[0]
+        for r in db.execute(
+            "SELECT name FROM sqlite_master WHERE type='table' ORDER BY 1"
+        ).fetchall()
+    ]
+    assert "t_ivf_centroids00" in tables
+
+
+def test_ivf_quantizer_all_params(db):
+    """All params together: nlist, nprobe, quantizer, oversample."""
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0("
+        "v float[768] distance_metric=cosine "
+        "indexed by ivf(nlist=128, nprobe=16, quantizer=int8, oversample=4))"
+    )
+    tables = [
+        r[0]
+        for r in db.execute(
+            "SELECT name FROM sqlite_master WHERE type='table' ORDER BY 1"
+        ).fetchall()
+    ]
+    assert "t_ivf_centroids00" in tables
+
+
+def test_ivf_error_oversample_without_quantizer(db):
+    """oversample > 1 without quantizer should error."""
+    result = exec(
+        db,
+        "CREATE VIRTUAL TABLE t USING vec0("
+        "v float[768] indexed by ivf(oversample=10))",
+    )
+    assert "error" in result
+
+
+def test_ivf_error_unknown_quantizer(db):
+    result = exec(
+        db,
+        "CREATE VIRTUAL TABLE t USING vec0("
+        "v float[768] indexed by ivf(quantizer=pq))",
+    )
+    assert "error" in result
+
+
+def test_ivf_oversample_1_without_quantizer_ok(db):
+    """oversample=1 (default) is fine without quantizer."""
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0("
+        "v float[768] indexed by ivf(nlist=64))"
+    )
+    # Should succeed — oversample defaults to 1
+
+
+# ============================================================================
+# Functional tests — insert, train, query with quantized IVF
+# ============================================================================
+
+
+def test_ivf_int8_insert_and_query(db):
+    """int8 quantized IVF: insert, train, query."""
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0("
+        "v float[4] indexed by ivf(nlist=2, quantizer=int8, oversample=4))"
+    )
+    for i in range(20):
+        db.execute(
+            "INSERT INTO t(rowid, v) VALUES (?, ?)", [i, _f32([i, 0, 0, 0])]
+        )
+
+    db.execute("INSERT INTO t(t) VALUES ('compute-centroids')")
+
+    # Should be able to query
+    rows = db.execute(
+        "SELECT rowid, distance FROM t WHERE v MATCH ? AND k = 5",
+        [_f32([10.0, 0, 0, 0])],
+    ).fetchall()
+    assert len(rows) == 5
+    # Top result should be close to 10
+    assert rows[0][0] in range(8, 13)
+
+    # Full vectors should be in _ivf_vectors table
+    fv_count = db.execute("SELECT count(*) FROM t_ivf_vectors00").fetchone()[0]
+    assert fv_count == 20
+
+
+def test_ivf_binary_insert_and_query(db):
+    """Binary quantized IVF: insert, train, query."""
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0("
+        "v float[8] indexed by ivf(nlist=2, quantizer=binary, oversample=4))"
+    )
+    for i in range(20):
+        # Vectors with varying sign patterns
+        v = [(i * 0.1 - 1.0) + j * 0.3 for j in range(8)]
+        db.execute(
+            "INSERT INTO t(rowid, v) VALUES (?, ?)", [i, _f32(v)]
+        )
+
+    db.execute("INSERT INTO t(t) VALUES ('compute-centroids')")
+
+    rows = db.execute(
+        "SELECT rowid FROM t WHERE v MATCH ? AND k = 5",
+        [_f32([0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5])],
+    ).fetchall()
+    assert len(rows) == 5
+
+    # Full vectors stored
+    fv_count = db.execute("SELECT count(*) FROM t_ivf_vectors00").fetchone()[0]
+    assert fv_count == 20
+
+
+def test_ivf_int8_cell_sizes_smaller(db):
+    """Cell blobs should be smaller with int8 quantization."""
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0("
+        "v float[768] indexed by ivf(nlist=2, quantizer=int8, oversample=1))"
+    )
+    for i in range(10):
+        db.execute(
+            "INSERT INTO t(rowid, v) VALUES (?, ?)",
+            [i, _f32([float(x) / 768 for x in range(768)])],
+        )
+
+    # Cell vectors blob: 10 vectors at int8 = 10 * 768 = 7680 bytes
+    # vs float32 = 10 * 768 * 4 = 30720 bytes
+    # But cells have capacity 64, so blob = 64 * 768 = 49152 (int8) vs 64*768*4=196608 (float32)
+    blob_size = db.execute(
+        "SELECT length(vectors) FROM t_ivf_cells00 LIMIT 1"
+    ).fetchone()[0]
+    # int8: 64 slots * 768 bytes = 49152
+    assert blob_size == 64 * 768
+
+
+def test_ivf_binary_cell_sizes_smaller(db):
+    """Cell blobs should be much smaller with binary quantization."""
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0("
+        "v float[768] indexed by ivf(nlist=2, quantizer=binary, oversample=1))"
+    )
+    for i in range(10):
+        db.execute(
+            "INSERT INTO t(rowid, v) VALUES (?, ?)",
+            [i, _f32([float(x) / 768 for x in range(768)])],
+        )
+
+    blob_size = db.execute(
+        "SELECT length(vectors) FROM t_ivf_cells00 LIMIT 1"
+    ).fetchone()[0]
+    # binary: 64 slots * 768/8 bytes = 6144
+    assert blob_size == 64 * (768 // 8)
+
+
+def test_ivf_int8_oversample_improves_recall(db):
+    """Oversample re-ranking should improve recall over oversample=1."""
+    # Create two tables: one with oversample=1, one with oversample=10
+    db.execute(
+        "CREATE VIRTUAL TABLE t1 USING vec0("
+        "v float[4] indexed by ivf(nlist=4, quantizer=int8, oversample=1))"
+    )
+    db.execute(
+        "CREATE VIRTUAL TABLE t2 USING vec0("
+        "v float[4] indexed by ivf(nlist=4, quantizer=int8, oversample=10))"
+    )
+    for i in range(100):
+        v = _f32([i * 0.1, (i * 0.1) % 3, (i * 0.3) % 5, i * 0.01])
+        db.execute("INSERT INTO t1(rowid, v) VALUES (?, ?)", [i, v])
+        db.execute("INSERT INTO t2(rowid, v) VALUES (?, ?)", [i, v])
+
+    db.execute("INSERT INTO t1(t1) VALUES ('compute-centroids')")
+    db.execute("INSERT INTO t2(t2) VALUES ('compute-centroids')")
+    db.execute("INSERT INTO t1(t1) VALUES ('nprobe=4')")
+    db.execute("INSERT INTO t2(t2) VALUES ('nprobe=4')")
+
+    query = _f32([5.0, 1.5, 2.5, 0.5])
+    r1 = db.execute("SELECT rowid FROM t1 WHERE v MATCH ? AND k=10", [query]).fetchall()
+    r2 = db.execute("SELECT rowid FROM t2 WHERE v MATCH ? AND k=10", [query]).fetchall()
+
+    # Both should return 10 results
+    assert len(r1) == 10
+    assert len(r2) == 10
+    # oversample=10 should have at least as good recall (same or better ordering)
+
+
+def test_ivf_quantized_delete(db):
+    """Delete should remove from both cells and _ivf_vectors."""
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0("
+        "v float[4] indexed by ivf(nlist=2, quantizer=int8, oversample=1))"
+    )
+    for i in range(10):
+        db.execute(
+            "INSERT INTO t(rowid, v) VALUES (?, ?)", [i, _f32([i, 0, 0, 0])]
+        )
+
+    db.execute("INSERT INTO t(t) VALUES ('compute-centroids')")
+    assert db.execute("SELECT count(*) FROM t_ivf_vectors00").fetchone()[0] == 10
+
+    db.execute("DELETE FROM t WHERE rowid = 5")
+    # _ivf_vectors should have 9 rows
+    assert db.execute("SELECT count(*) FROM t_ivf_vectors00").fetchone()[0] == 9
+
+
+def test_ivf_binary_rejects_non_multiple_of_8_dims(db):
+    """Binary quantizer requires dimensions divisible by 8."""
+    with pytest.raises(sqlite3.OperationalError):
+        db.execute(
+            "CREATE VIRTUAL TABLE t USING vec0("
+            "  v float[12] indexed by ivf(quantizer=binary)"
+            ")"
+        )
+
+    # Dimensions divisible by 8 should work
+    db.execute(
+        "CREATE VIRTUAL TABLE t2 USING vec0("
+        "  v float[16] indexed by ivf(quantizer=binary)"
+        ")"
+    )

tests/test-ivf.py

@@ -0,0 +1,426 @@
+import pytest
+import sqlite3
+import struct
+import math
+from helpers import _f32, exec
+
+
+@pytest.fixture()
+def db():
+    db = sqlite3.connect(":memory:")
+    db.row_factory = sqlite3.Row
+    db.enable_load_extension(True)
+    db.load_extension("dist/vec0")
+    db.enable_load_extension(False)
+    return db
+
+
+def ivf_total_vectors(db, table="t", col=0):
+    """Count total vectors across all IVF cells."""
+    return db.execute(
+        f"SELECT COALESCE(SUM(n_vectors), 0) FROM {table}_ivf_cells{col:02d}"
+    ).fetchone()[0]
+
+
+def ivf_unassigned_count(db, table="t", col=0):
+    """Count vectors in unassigned cells (centroid_id=-1)."""
+    return db.execute(
+        f"SELECT COALESCE(SUM(n_vectors), 0) FROM {table}_ivf_cells{col:02d} WHERE centroid_id = -1"
+    ).fetchone()[0]
+
+
+def ivf_assigned_count(db, table="t", col=0):
+    """Count vectors in trained cells (centroid_id >= 0)."""
+    return db.execute(
+        f"SELECT COALESCE(SUM(n_vectors), 0) FROM {table}_ivf_cells{col:02d} WHERE centroid_id >= 0"
+    ).fetchone()[0]
+
+
+# ============================================================================
+# Parser tests
+# ============================================================================
+
+
+def test_ivf_create_defaults(db):
+    """ivf() with no args uses defaults."""
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0(v float[4] indexed by ivf())"
+    )
+    tables = [
+        r[0]
+        for r in db.execute(
+            "SELECT name FROM sqlite_master WHERE type='table' ORDER BY 1"
+        ).fetchall()
+    ]
+    assert "t_ivf_centroids00" in tables
+    assert "t_ivf_cells00" in tables
+    assert "t_ivf_rowid_map00" in tables
+
+
+def test_ivf_create_custom_params(db):
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0(v float[4] indexed by ivf(nlist=64, nprobe=8))"
+    )
+    tables = [
+        r[0]
+        for r in db.execute(
+            "SELECT name FROM sqlite_master WHERE type='table' ORDER BY 1"
+        ).fetchall()
+    ]
+    assert "t_ivf_centroids00" in tables
+    assert "t_ivf_cells00" in tables
+
+
+def test_ivf_create_with_distance_metric(db):
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0(v float[4] distance_metric=cosine indexed by ivf(nlist=16))"
+    )
+    tables = [
+        r[0]
+        for r in db.execute(
+            "SELECT name FROM sqlite_master WHERE type='table' ORDER BY 1"
+        ).fetchall()
+    ]
+    assert "t_ivf_centroids00" in tables
+
+
+def test_ivf_create_error_unknown_key(db):
+    result = exec(
+        db,
+        "CREATE VIRTUAL TABLE t USING vec0(v float[4] indexed by ivf(bogus=1))",
+    )
+    assert "error" in result
+
+
+def test_ivf_create_error_nprobe_gt_nlist(db):
+    result = exec(
+        db,
+        "CREATE VIRTUAL TABLE t USING vec0(v float[4] indexed by ivf(nlist=4, nprobe=10))",
+    )
+    assert "error" in result
+
+
+# ============================================================================
+# Shadow table tests
+# ============================================================================
+
+
+def test_ivf_shadow_tables_created(db):
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0(v float[4] indexed by ivf(nlist=8))"
+    )
+    trained = db.execute(
+        "SELECT value FROM t_info WHERE key = 'ivf_trained_0'"
+    ).fetchone()[0]
+    assert str(trained) == "0"
+
+    # No cells yet (created lazily on first insert)
+    count = db.execute(
+        "SELECT count(*) FROM t_ivf_cells00"
+    ).fetchone()[0]
+    assert count == 0
+
+
+def test_ivf_drop_cleans_up(db):
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0(v float[4] indexed by ivf(nlist=4))"
+    )
+    db.execute("DROP TABLE t")
+    tables = [
+        r[0]
+        for r in db.execute(
+            "SELECT name FROM sqlite_master WHERE type='table'"
+        ).fetchall()
+    ]
+    assert not any("ivf" in t for t in tables)
+
+
+# ============================================================================
+# Insert tests (flat mode)
+# ============================================================================
+
+
+def test_ivf_insert_flat_mode(db):
+    """Before training, vectors go to unassigned cell."""
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0(v float[4] indexed by ivf(nlist=4))"
+    )
+    db.execute("INSERT INTO t(rowid, v) VALUES (1, ?)", [_f32([1, 2, 3, 4])])
+    db.execute("INSERT INTO t(rowid, v) VALUES (2, ?)", [_f32([5, 6, 7, 8])])
+
+    assert ivf_unassigned_count(db) == 2
+    assert ivf_assigned_count(db) == 0
+
+    # rowid_map should have 2 entries
+    assert db.execute("SELECT count(*) FROM t_ivf_rowid_map00").fetchone()[0] == 2
+
+
+def test_ivf_delete_flat_mode(db):
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0(v float[4] indexed by ivf(nlist=4))"
+    )
+    db.execute("INSERT INTO t(rowid, v) VALUES (1, ?)", [_f32([1, 2, 3, 4])])
+    db.execute("INSERT INTO t(rowid, v) VALUES (2, ?)", [_f32([5, 6, 7, 8])])
+    db.execute("DELETE FROM t WHERE rowid = 1")
+
+    assert ivf_unassigned_count(db) == 1
+    assert db.execute("SELECT count(*) FROM t_ivf_rowid_map00").fetchone()[0] == 1
+
+
+# ============================================================================
+# KNN flat mode tests
+# ============================================================================
+
+
+def test_ivf_knn_flat_mode(db):
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0(v float[4] indexed by ivf(nlist=4))"
+    )
+    db.execute("INSERT INTO t(rowid, v) VALUES (1, ?)", [_f32([1, 0, 0, 0])])
+    db.execute("INSERT INTO t(rowid, v) VALUES (2, ?)", [_f32([2, 0, 0, 0])])
+    db.execute("INSERT INTO t(rowid, v) VALUES (3, ?)", [_f32([9, 0, 0, 0])])
+
+    rows = db.execute(
+        "SELECT rowid, distance FROM t WHERE v MATCH ? AND k = 2",
+        [_f32([1.5, 0, 0, 0])],
+    ).fetchall()
+    assert len(rows) == 2
+    rowids = {r[0] for r in rows}
+    assert rowids == {1, 2}
+
+
+def test_ivf_knn_flat_empty(db):
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0(v float[4] indexed by ivf(nlist=4))"
+    )
+    rows = db.execute(
+        "SELECT rowid FROM t WHERE v MATCH ? AND k = 5",
+        [_f32([1, 0, 0, 0])],
+    ).fetchall()
+    assert len(rows) == 0
+
+
+# ============================================================================
+# compute-centroids tests
+# ============================================================================
+
+
+def test_compute_centroids(db):
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0(v float[4] indexed by ivf(nlist=4))"
+    )
+    for i in range(40):
+        db.execute(
+            "INSERT INTO t(rowid, v) VALUES (?, ?)",
+            [i, _f32([i % 10, i // 10, 0, 0])],
+        )
+
+    assert ivf_unassigned_count(db) == 40
+
+    db.execute("INSERT INTO t(t) VALUES ('compute-centroids')")
+
+    # After training: unassigned cell should be gone (or empty), vectors in trained cells
+    assert ivf_unassigned_count(db) == 0
+    assert ivf_assigned_count(db) == 40
+    assert db.execute("SELECT count(*) FROM t_ivf_centroids00").fetchone()[0] == 4
+    trained = db.execute(
+        "SELECT value FROM t_info WHERE key='ivf_trained_0'"
+    ).fetchone()[0]
+    assert str(trained) == "1"
+
+
+def test_compute_centroids_recompute(db):
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0(v float[4] indexed by ivf(nlist=2))"
+    )
+    for i in range(20):
+        db.execute(
+            "INSERT INTO t(rowid, v) VALUES (?, ?)", [i, _f32([i, 0, 0, 0])]
+        )
+
+    db.execute("INSERT INTO t(t) VALUES ('compute-centroids')")
+    assert db.execute("SELECT count(*) FROM t_ivf_centroids00").fetchone()[0] == 2
+
+    db.execute("INSERT INTO t(t) VALUES ('compute-centroids')")
+    assert db.execute("SELECT count(*) FROM t_ivf_centroids00").fetchone()[0] == 2
+    assert ivf_assigned_count(db) == 20
+
+
+# ============================================================================
+# Insert after training (assigned mode)
+# ============================================================================
+
+
+def test_ivf_insert_after_training(db):
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0(v float[4] indexed by ivf(nlist=2))"
+    )
+    for i in range(20):
+        db.execute(
+            "INSERT INTO t(rowid, v) VALUES (?, ?)", [i, _f32([i, 0, 0, 0])]
+        )
+
+    db.execute("INSERT INTO t(t) VALUES ('compute-centroids')")
+
+    db.execute(
+        "INSERT INTO t(rowid, v) VALUES (100, ?)", [_f32([5, 0, 0, 0])]
+    )
+
+    # Should be in a trained cell, not unassigned
+    row = db.execute(
+        "SELECT m.cell_id, c.centroid_id FROM t_ivf_rowid_map00 m "
+        "JOIN t_ivf_cells00 c ON c.rowid = m.cell_id "
+        "WHERE m.rowid = 100"
+    ).fetchone()
+    assert row is not None
+    assert row[1] >= 0  # centroid_id >= 0 means trained cell
+
+
+# ============================================================================
+# KNN after training (IVF probe mode)
+# ============================================================================
+
+
+def test_ivf_knn_after_training(db):
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0(v float[4] indexed by ivf(nlist=4, nprobe=4))"
+    )
+    for i in range(100):
+        db.execute(
+            "INSERT INTO t(rowid, v) VALUES (?, ?)", [i, _f32([i, 0, 0, 0])]
+        )
+
+    db.execute("INSERT INTO t(t) VALUES ('compute-centroids')")
+
+    rows = db.execute(
+        "SELECT rowid, distance FROM t WHERE v MATCH ? AND k = 5",
+        [_f32([50.0, 0, 0, 0])],
+    ).fetchall()
+    assert len(rows) == 5
+    assert rows[0][0] == 50
+    assert rows[0][1] < 0.01
+
+
+def test_ivf_knn_k_larger_than_n(db):
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0(v float[4] indexed by ivf(nlist=2, nprobe=2))"
+    )
+    for i in range(5):
+        db.execute(
+            "INSERT INTO t(rowid, v) VALUES (?, ?)", [i, _f32([i, 0, 0, 0])]
+        )
+
+    db.execute("INSERT INTO t(t) VALUES ('compute-centroids')")
+
+    rows = db.execute(
+        "SELECT rowid FROM t WHERE v MATCH ? AND k = 100",
+        [_f32([0, 0, 0, 0])],
+    ).fetchall()
+    assert len(rows) == 5
+
+
+# ============================================================================
+# Manual centroid import (set-centroid, assign-vectors)
+# ============================================================================
+
+
+def test_set_centroid_and_assign(db):
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0(v float[4] indexed by ivf(nlist=0))"
+    )
+    for i in range(20):
+        db.execute(
+            "INSERT INTO t(rowid, v) VALUES (?, ?)", [i, _f32([i, 0, 0, 0])]
+        )
+
+    db.execute(
+        "INSERT INTO t(t, v) VALUES ('set-centroid:0', ?)",
+        [_f32([5, 0, 0, 0])],
+    )
+    db.execute(
+        "INSERT INTO t(t, v) VALUES ('set-centroid:1', ?)",
+        [_f32([15, 0, 0, 0])],
+    )
+
+    assert db.execute("SELECT count(*) FROM t_ivf_centroids00").fetchone()[0] == 2
+
+    db.execute("INSERT INTO t(t) VALUES ('assign-vectors')")
+
+    assert ivf_unassigned_count(db) == 0
+    assert ivf_assigned_count(db) == 20
+
+
+# ============================================================================
+# clear-centroids
+# ============================================================================
+
+
+def test_clear_centroids(db):
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0(v float[4] indexed by ivf(nlist=2))"
+    )
+    for i in range(20):
+        db.execute(
+            "INSERT INTO t(rowid, v) VALUES (?, ?)", [i, _f32([i, 0, 0, 0])]
+        )
+
+    db.execute("INSERT INTO t(t) VALUES ('compute-centroids')")
+    assert db.execute("SELECT count(*) FROM t_ivf_centroids00").fetchone()[0] == 2
+
+    db.execute("INSERT INTO t(t) VALUES ('clear-centroids')")
+    assert db.execute("SELECT count(*) FROM t_ivf_centroids00").fetchone()[0] == 0
+    assert ivf_unassigned_count(db) == 20
+    trained = db.execute(
+        "SELECT value FROM t_info WHERE key='ivf_trained_0'"
+    ).fetchone()[0]
+    assert str(trained) == "0"
+
+
+# ============================================================================
+# Delete after training
+# ============================================================================
+
+
+def test_ivf_delete_after_training(db):
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0(v float[4] indexed by ivf(nlist=2))"
+    )
+    for i in range(10):
+        db.execute(
+            "INSERT INTO t(rowid, v) VALUES (?, ?)", [i, _f32([i, 0, 0, 0])]
+        )
+
+    db.execute("INSERT INTO t(t) VALUES ('compute-centroids')")
+    assert ivf_assigned_count(db) == 10
+
+    db.execute("DELETE FROM t WHERE rowid = 5")
+    assert ivf_assigned_count(db) == 9
+    assert db.execute("SELECT count(*) FROM t_ivf_rowid_map00").fetchone()[0] == 9
+
+
+# ============================================================================
+# Recall test
+# ============================================================================
+
+
+def test_ivf_recall_nprobe_equals_nlist(db):
+    db.execute(
+        "CREATE VIRTUAL TABLE t USING vec0(v float[4] indexed by ivf(nlist=8, nprobe=8))"
+    )
+    for i in range(100):
+        db.execute(
+            "INSERT INTO t(rowid, v) VALUES (?, ?)", [i, _f32([i, 0, 0, 0])]
+        )
+
+    db.execute("INSERT INTO t(t) VALUES ('compute-centroids')")
+
+    rows = db.execute(
+        "SELECT rowid FROM t WHERE v MATCH ? AND k = 10",
+        [_f32([50.0, 0, 0, 0])],
+    ).fetchall()
+    rowids = {r[0] for r in rows}
+
+    # 45 and 55 are equidistant from 50, so either may appear in top 10
+    assert 50 in rowids
+    assert len(rowids) == 10
+    assert all(45 <= r <= 55 for r in rowids)

tests/test-legacy-compat.py

@@ -0,0 +1,138 @@
+"""Backwards compatibility tests: current sqlite-vec reading legacy databases.
+
+The fixture file tests/fixtures/legacy-v0.1.6.db was generated by
+tests/generate_legacy_db.py using sqlite-vec v0.1.6. These tests verify
+that the current version can fully read, query, insert into, and delete
+from tables created by older versions.
+"""
+import sqlite3
+import struct
+import os
+import shutil
+import pytest
+
+FIXTURE_PATH = os.path.join(os.path.dirname(__file__), "fixtures", "legacy-v0.1.6.db")
+
+
+def _f32(vals):
+    return struct.pack(f"{len(vals)}f", *vals)
+
+
+@pytest.fixture()
+def legacy_db(tmp_path):
+    """Copy the legacy fixture to a temp dir so tests can modify it."""
+    if not os.path.exists(FIXTURE_PATH):
+        pytest.skip("Legacy fixture not found — run: uv run --script tests/generate_legacy_db.py")
+    db_path = str(tmp_path / "legacy.db")
+    shutil.copy2(FIXTURE_PATH, db_path)
+    db = sqlite3.connect(db_path)
+    db.row_factory = sqlite3.Row
+    db.enable_load_extension(True)
+    db.load_extension("dist/vec0")
+    return db
+
+
+def test_legacy_select_count(legacy_db):
+    """Basic SELECT count should return all rows."""
+    count = legacy_db.execute("SELECT count(*) FROM legacy_vectors").fetchone()[0]
+    assert count == 50
+
+
+def test_legacy_point_query(legacy_db):
+    """Point query by rowid should return correct vector."""
+    row = legacy_db.execute(
+        "SELECT rowid, emb FROM legacy_vectors WHERE rowid = 1"
+    ).fetchone()
+    assert row["rowid"] == 1
+    vec = struct.unpack("4f", row["emb"])
+    assert vec[0] == pytest.approx(1.0)
+
+
+def test_legacy_knn(legacy_db):
+    """KNN query on legacy table should return correct results."""
+    query = _f32([1.0, 0.0, 0.0, 0.0])
+    rows = legacy_db.execute(
+        "SELECT rowid, distance FROM legacy_vectors "
+        "WHERE emb MATCH ? AND k = 5",
+        [query],
+    ).fetchall()
+    assert len(rows) == 5
+    assert rows[0]["rowid"] == 1
+    assert rows[0]["distance"] == pytest.approx(0.0)
+    for i in range(len(rows) - 1):
+        assert rows[i]["distance"] <= rows[i + 1]["distance"]
+
+
+def test_legacy_insert(legacy_db):
+    """INSERT into legacy table should work."""
+    legacy_db.execute(
+        "INSERT INTO legacy_vectors(rowid, emb) VALUES (100, ?)",
+        [_f32([100.0, 0.0, 0.0, 0.0])],
+    )
+    count = legacy_db.execute("SELECT count(*) FROM legacy_vectors").fetchone()[0]
+    assert count == 51
+
+    rows = legacy_db.execute(
+        "SELECT rowid FROM legacy_vectors WHERE emb MATCH ? AND k = 1",
+        [_f32([100.0, 0.0, 0.0, 0.0])],
+    ).fetchall()
+    assert rows[0]["rowid"] == 100
+
+
+def test_legacy_delete(legacy_db):
+    """DELETE from legacy table should work."""
+    legacy_db.execute("DELETE FROM legacy_vectors WHERE rowid = 1")
+    count = legacy_db.execute("SELECT count(*) FROM legacy_vectors").fetchone()[0]
+    assert count == 49
+
+    rows = legacy_db.execute(
+        "SELECT rowid FROM legacy_vectors WHERE emb MATCH ? AND k = 5",
+        [_f32([1.0, 0.0, 0.0, 0.0])],
+    ).fetchall()
+    assert 1 not in [r["rowid"] for r in rows]
+
+
+def test_legacy_fullscan(legacy_db):
+    """Full scan should work."""
+    rows = legacy_db.execute(
+        "SELECT rowid FROM legacy_vectors ORDER BY rowid LIMIT 5"
+    ).fetchall()
+    assert [r["rowid"] for r in rows] == [1, 2, 3, 4, 5]
+
+
+def test_legacy_name_conflict_table(legacy_db):
+    """Legacy table where column name == table name should work.
+
+    The v0.1.6 DB has: CREATE VIRTUAL TABLE emb USING vec0(emb float[4])
+    Current code should NOT add the command column for this table
+    (detected via _info version check), avoiding the name conflict.
+    """
+    count = legacy_db.execute("SELECT count(*) FROM emb").fetchone()[0]
+    assert count == 10
+
+    rows = legacy_db.execute(
+        "SELECT rowid, distance FROM emb WHERE emb MATCH ? AND k = 3",
+        [_f32([1.0, 0.0, 0.0, 0.0])],
+    ).fetchall()
+    assert len(rows) == 3
+    assert rows[0]["rowid"] == 1
+
+
+def test_legacy_name_conflict_insert_delete(legacy_db):
+    """INSERT and DELETE on legacy name-conflict table."""
+    legacy_db.execute(
+        "INSERT INTO emb(rowid, emb) VALUES (100, ?)",
+        [_f32([100.0, 0.0, 0.0, 0.0])],
+    )
+    assert legacy_db.execute("SELECT count(*) FROM emb").fetchone()[0] == 11
+
+    legacy_db.execute("DELETE FROM emb WHERE rowid = 5")
+    assert legacy_db.execute("SELECT count(*) FROM emb").fetchone()[0] == 10
+
+
+def test_legacy_no_command_column(legacy_db):
+    """Legacy tables should NOT have the command column."""
+    with pytest.raises(sqlite3.OperationalError):
+        legacy_db.execute(
+            "INSERT INTO legacy_vectors(legacy_vectors) VALUES ('some_command')"
+        )

tests/test-loadable.py

@@ -119,151 +119,9 @@ FUNCTIONS = [
 MODULES = [
     "vec0",
     "vec_each",
-    # "vec_static_blob_entries",
-    # "vec_static_blobs",
 ]
 
 
-def register_numpy(db, name: str, array):
-    ptr = array.__array_interface__["data"][0]
-    nvectors, dimensions = array.__array_interface__["shape"]
-    element_type = array.__array_interface__["typestr"]
-
-    assert element_type == "<f4"
-
-    name_escaped = db.execute("select printf('%w', ?)", [name]).fetchone()[0]
-
-    db.execute(
-        """
-          insert into temp.vec_static_blobs(name, data)
-          select ?, vec_static_blob_from_raw(?, ?, ?, ?)
-        """,
-        [name, ptr, element_type, dimensions, nvectors],
-    )
-
-    db.execute(
-        f'create virtual table "{name_escaped}" using vec_static_blob_entries({name_escaped})'
-    )
-
-
-def test_vec_static_blob_entries():
-    db = connect(EXT_PATH, extra_entrypoint="sqlite3_vec_static_blobs_init")
-
-    x = np.array([[0.1, 0.2, 0.3, 0.4], [0.9, 0.8, 0.7, 0.6]], dtype=np.float32)
-    y = np.array([[0.2, 0.3], [0.9, 0.8], [0.6, 0.5]], dtype=np.float32)
-    z = np.array(
-        [
-            [0.1, 0.1, 0.1, 0.1],
-            [0.2, 0.2, 0.2, 0.2],
-            [0.3, 0.3, 0.3, 0.3],
-            [0.4, 0.4, 0.4, 0.4],
-            [0.5, 0.5, 0.5, 0.5],
-        ],
-        dtype=np.float32,
-    )
-
-    register_numpy(db, "x", x)
-    register_numpy(db, "y", y)
-    register_numpy(db, "z", z)
-    assert execute_all(
-        db, "select *, dimensions, count from temp.vec_static_blobs;"
-    ) == [
-        {
-            "count": 2,
-            "data": None,
-            "dimensions": 4,
-            "name": "x",
-        },
-        {
-            "count": 3,
-            "data": None,
-            "dimensions": 2,
-            "name": "y",
-        },
-        {
-            "count": 5,
-            "data": None,
-            "dimensions": 4,
-            "name": "z",
-        },
-    ]
-
-    assert execute_all(db, "select vec_to_json(vector) from x;") == [
-        {
-            "vec_to_json(vector)": "[0.100000,0.200000,0.300000,0.400000]",
-        },
-        {
-            "vec_to_json(vector)": "[0.900000,0.800000,0.700000,0.600000]",
-        },
-    ]
-    assert execute_all(db, "select (vector) from y limit 2;") == [
-        {
-            "vector": b"\xcd\xccL>\x9a\x99\x99>",
-        },
-        {
-            "vector": b"fff?\xcd\xccL?",
-        },
-    ]
-    assert execute_all(db, "select rowid, (vector) from z") == [
-        {
-            "rowid": 0,
-            "vector": b"\xcd\xcc\xcc=\xcd\xcc\xcc=\xcd\xcc\xcc=\xcd\xcc\xcc=",
-        },
-        {
-            "rowid": 1,
-            "vector": b"\xcd\xccL>\xcd\xccL>\xcd\xccL>\xcd\xccL>",
-        },
-        {
-            "rowid": 2,
-            "vector": b"\x9a\x99\x99>\x9a\x99\x99>\x9a\x99\x99>\x9a\x99\x99>",
-        },
-        {
-            "rowid": 3,
-            "vector": b"\xcd\xcc\xcc>\xcd\xcc\xcc>\xcd\xcc\xcc>\xcd\xcc\xcc>",
-        },
-        {
-            "rowid": 4,
-            "vector": b"\x00\x00\x00?\x00\x00\x00?\x00\x00\x00?\x00\x00\x00?",
-        },
-    ]
-    assert execute_all(
-        db,
-        "select rowid, vec_to_json(vector) as v from z where vector match ? and k = 3 order by distance;",
-        [np.array([0.3, 0.3, 0.3, 0.3], dtype=np.float32)],
-    ) == [
-        {
-            "rowid": 2,
-            "v": "[0.300000,0.300000,0.300000,0.300000]",
-        },
-        {
-            "rowid": 3,
-            "v": "[0.400000,0.400000,0.400000,0.400000]",
-        },
-        {
-            "rowid": 1,
-            "v": "[0.200000,0.200000,0.200000,0.200000]",
-        },
-    ]
-    assert execute_all(
-        db,
-        "select rowid, vec_to_json(vector) as v from z where vector match ? and k = 3 order by distance;",
-        [np.array([0.6, 0.6, 0.6, 0.6], dtype=np.float32)],
-    ) == [
-        {
-            "rowid": 4,
-            "v": "[0.500000,0.500000,0.500000,0.500000]",
-        },
-        {
-            "rowid": 3,
-            "v": "[0.400000,0.400000,0.400000,0.400000]",
-        },
-        {
-            "rowid": 2,
-            "v": "[0.300000,0.300000,0.300000,0.300000]",
-        },
-    ]
-
-
 def test_limits():
     db = connect(EXT_PATH)
     with _raises(
@@ -507,6 +365,34 @@ def test_vec_distance_l1():
     )
 
 
+def test_vec_reject_nan_inf():
+    """NaN and Inf in float32 vectors should be rejected."""
+    import struct, math
+
+    # NaN via blob
+    nan_blob = struct.pack("4f", 1.0, float("nan"), 3.0, 4.0)
+    with pytest.raises(sqlite3.OperationalError, match="NaN"):
+        db.execute("SELECT vec_length(?)", [nan_blob])
+
+    # Inf via blob
+    inf_blob = struct.pack("4f", 1.0, float("inf"), 3.0, 4.0)
+    with pytest.raises(sqlite3.OperationalError, match="Inf"):
+        db.execute("SELECT vec_length(?)", [inf_blob])
+
+    # -Inf via blob
+    ninf_blob = struct.pack("4f", 1.0, float("-inf"), 3.0, 4.0)
+    with pytest.raises(sqlite3.OperationalError, match="Inf"):
+        db.execute("SELECT vec_length(?)", [ninf_blob])
+
+    # NaN via JSON
+    # Note: JSON doesn't have NaN literal, but strtod may parse "NaN"
+    # This tests the blob path which is the primary input method
+
+    # Valid vectors still work
+    ok_blob = struct.pack("4f", 1.0, 2.0, 3.0, 4.0)
+    assert db.execute("SELECT vec_length(?)", [ok_blob]).fetchone()[0] == 4
+
+
 def test_vec_distance_l2():
     vec_distance_l2 = lambda *args, a="?", b="?": db.execute(
         f"select vec_distance_l2({a}, {b})", args
@@ -523,11 +409,17 @@ def test_vec_distance_l2():
 
         x = vec_distance_l2(a_sql_t, b_sql_t, a=transform, b=transform)
         y = npy_l2(np.array(a), np.array(b))
-        assert isclose(x, y, abs_tol=1e-6)
+        assert isclose(x, y, rel_tol=1e-5, abs_tol=1e-6)
 
     check([1.2, 0.1], [0.4, -0.4])
     check([-1.2, -0.1], [-0.4, 0.4])
     check([1, 2, 3], [-9, -8, -7], dtype=np.int8)
+    # Extreme int8 values: diff=255, squared=65025 which overflows i16
+    # This tests the NEON widening multiply fix (slight float rounding expected)
+    check([-128] * 8, [127] * 8, dtype=np.int8)
+    check([-128] * 16, [127] * 16, dtype=np.int8)
+    check([-128, 127, -128, 127, -128, 127, -128, 127],
+          [127, -128, 127, -128, 127, -128, 127, -128], dtype=np.int8)
 
 
 def test_vec_length():
@@ -1618,231 +1510,6 @@ def test_vec_each():
       vec_each_f32(None)
 
 
-import io
-
-
-def to_npy(arr):
-    buf = io.BytesIO()
-    np.save(buf, arr)
-    buf.seek(0)
-    return buf.read()
-
-
-def test_vec_npy_each():
-    db = connect(EXT_PATH, extra_entrypoint="sqlite3_vec_numpy_init")
-    vec_npy_each = lambda *args: execute_all(
-        db, "select rowid, * from vec_npy_each(?)", args
-    )
-    assert vec_npy_each(to_npy(np.array([1.1, 2.2, 3.3], dtype=np.float32))) == [
-        {
-            "rowid": 0,
-            "vector": _f32([1.1, 2.2, 3.3]),
-        },
-    ]
-    assert vec_npy_each(to_npy(np.array([[1.1, 2.2, 3.3]], dtype=np.float32))) == [
-        {
-            "rowid": 0,
-            "vector": _f32([1.1, 2.2, 3.3]),
-        },
-    ]
-    assert vec_npy_each(
-        to_npy(np.array([[1.1, 2.2, 3.3], [9.9, 8.8, 7.7]], dtype=np.float32))
-    ) == [
-        {
-            "rowid": 0,
-            "vector": _f32([1.1, 2.2, 3.3]),
-        },
-        {
-            "rowid": 1,
-            "vector": _f32([9.9, 8.8, 7.7]),
-        },
-    ]
-
-    assert vec_npy_each(to_npy(np.array([], dtype=np.float32))) == []
-
-
-def test_vec_npy_each_errors():
-    db = connect(EXT_PATH, extra_entrypoint="sqlite3_vec_numpy_init")
-    vec_npy_each = lambda *args: execute_all(
-        db, "select rowid, * from vec_npy_each(?)", args
-    )
-
-    full = b"\x93NUMPY\x01\x00v\x00{'descr': '<f4', 'fortran_order': False, 'shape': (2, 4), }                                                          \n\xcd\xcc\x8c?\xcd\xcc\x0c@33S@\xcd\xcc\x8c@ff\x1eA\xcd\xcc\x0cAff\xf6@33\xd3@"
-
-    # EVIDENCE-OF: V03312_20150 numpy validation too short
-    with _raises("numpy array too short"):
-        vec_npy_each(b"")
-    # EVIDENCE-OF: V11954_28792 numpy validate magic
-    with _raises("numpy array does not contain the 'magic' header"):
-        vec_npy_each(b"\x93NUMPX\x01\x00v\x00")
-
-    with _raises("numpy array header length is invalid"):
-        vec_npy_each(b"\x93NUMPY\x01\x00v\x00")
-
-    with _raises("numpy header did not start with '{'"):
-        vec_npy_each(
-            b"\x93NUMPY\x01\x00v\x00c'descr': '<f4', 'fortran_order': False, 'shape': (2, 4), }                                                          \n\xcd\xcc\x8c?\xcd\xcc\x0c@33S@\xcd\xcc\x8c@ff\x1eA\xcd\xcc\x0cAff\xf6@33\xd3@"
-        )
-
-    with _raises("expected key in numpy header"):
-        vec_npy_each(
-            b"\x93NUMPY\x01\x00v\x00{                                                                                                                    \n\xcd\xcc\x8c?\xcd\xcc\x0c@33S@\xcd\xcc\x8c@ff\x1eA\xcd\xcc\x0cAff\xf6@33\xd3@"
-        )
-
-    with _raises("expected a string as key in numpy header"):
-        vec_npy_each(
-            b"\x93NUMPY\x01\x00v\x00{False: '<f4', 'fortran_order': False, 'shape': (2, 4), }                                                          \n\xcd\xcc\x8c?\xcd\xcc\x0c@33S@\xcd\xcc\x8c@ff\x1eA\xcd\xcc\x0cAff\xf6@33\xd3@"
-        )
-
-    with _raises("expected a ':' after key in numpy header"):
-        vec_npy_each(
-            b"\x93NUMPY\x01\x00v\x00{'descr'                                                                                                           \n\xcd\xcc\x8c?\xcd\xcc\x0c@33S@\xcd\xcc\x8c@ff\x1eA\xcd\xcc\x0cAff\xf6@33\xd3@"
-        )
-    with _raises("expected a ':' after key in numpy header"):
-        vec_npy_each(
-            b"\x93NUMPY\x01\x00v\x00{'descr' False                                                                                                           \n\xcd\xcc\x8c?\xcd\xcc\x0c@33S@\xcd\xcc\x8c@ff\x1eA\xcd\xcc\x0cAff\xf6@33\xd3@"
-        )
-
-    with _raises("expected a string value after 'descr' key"):
-        vec_npy_each(
-            b"\x93NUMPY\x01\x00v\x00{'descr':                                                                                                  \n\xcd\xcc\x8c?\xcd\xcc\x0c@33S@\xcd\xcc\x8c@ff\x1eA\xcd\xcc\x0cAff\xf6@33\xd3@"
-        )
-
-    with _raises("Only '<f4' values are supported in sqlite-vec numpy functions"):
-        vec_npy_each(
-            b"\x93NUMPY\x01\x00v\x00{'descr': '=f4', 'fortran_order': False, 'shape': (2, 4), }                                                          \n\xcd\xcc\x8c?\xcd\xcc\x0c@33S@\xcd\xcc\x8c@ff\x1eA\xcd\xcc\x0cAff\xf6@33\xd3@"
-        )
-
-    with _raises(
-        "Only fortran_order = False is supported in sqlite-vec numpy functions"
-    ):
-        vec_npy_each(
-            b"\x93NUMPY\x01\x00v\x00{'descr': '<f4', 'fortran_order': True, 'shape': (2, 4), }                                                          \n\xcd\xcc\x8c?\xcd\xcc\x0c@33S@\xcd\xcc\x8c@ff\x1eA\xcd\xcc\x0cAff\xf6@33\xd3@"
-        )
-
-    with _raises(
-        "Error parsing numpy array: Expected left parenthesis '(' after shape key"
-    ):
-        vec_npy_each(
-            b"\x93NUMPY\x01\x00v\x00{'shape':  2, 'descr': '<f4', 'fortran_order': False, }                                                          \n\xcd\xcc\x8c?\xcd\xcc\x0c@33S@\xcd\xcc\x8c@ff\x1eA\xcd\xcc\x0cAff\xf6@33\xd3@"
-        )
-
-    with _raises(
-        "Error parsing numpy array: Expected an initial number in shape value"
-    ):
-        vec_npy_each(
-            b"\x93NUMPY\x01\x00v\x00{'shape':  (, 'descr': '<f4', 'fortran_order': False, }                                                          \n\xcd\xcc\x8c?\xcd\xcc\x0c@33S@\xcd\xcc\x8c@ff\x1eA\xcd\xcc\x0cAff\xf6@33\xd3@"
-        )
-
-    with _raises("Error parsing numpy array: Expected comma after first shape value"):
-        vec_npy_each(
-            b"\x93NUMPY\x01\x00v\x00{'shape':  (2), 'descr': '<f4', 'fortran_order': False, }                                                          \n\xcd\xcc\x8c?\xcd\xcc\x0c@33S@\xcd\xcc\x8c@ff\x1eA\xcd\xcc\x0cAff\xf6@33\xd3@"
-        )
-
-    with _raises(
-        "Error parsing numpy array: unexpected header EOF while parsing shape"
-    ):
-        vec_npy_each(
-            b"\x93NUMPY\x01\x00v\x00{'shape':  (2,                                                                                             \n\xcd\xcc\x8c?\xcd\xcc\x0c@33S@\xcd\xcc\x8c@ff\x1eA\xcd\xcc\x0cAff\xf6@33\xd3@"
-        )
-
-    with _raises("Error parsing numpy array: unknown type in shape value"):
-        vec_npy_each(
-            b"\x93NUMPY\x01\x00v\x00{'shape':  (2, 'nope'                                                                                          \n\xcd\xcc\x8c?\xcd\xcc\x0c@33S@\xcd\xcc\x8c@ff\x1eA\xcd\xcc\x0cAff\xf6@33\xd3@"
-        )
-
-    with _raises(
-        "Error parsing numpy array: expected right parenthesis after shape value"
-    ):
-        vec_npy_each(
-            b"\x93NUMPY\x01\x00v\x00{'shape':  (2,4 (                                                                                          \n\xcd\xcc\x8c?\xcd\xcc\x0c@33S@\xcd\xcc\x8c@ff\x1eA\xcd\xcc\x0cAff\xf6@33\xd3@"
-        )
-
-    with _raises("Error parsing numpy array: unknown key in numpy header"):
-        vec_npy_each(
-            b"\x93NUMPY\x01\x00v\x00{'no': '<f4', 'fortran_order': False, 'shape': (2, 4), }                                                          \n\xcd\xcc\x8c?\xcd\xcc\x0c@33S@\xcd\xcc\x8c@ff\x1eA\xcd\xcc\x0cAff\xf6@33\xd3@"
-        )
-
-    with _raises("Error parsing numpy array: unknown extra token after value"):
-        vec_npy_each(
-            b"\x93NUMPY\x01\x00v\x00{'descr': '<f4' 'asdf', 'fortran_order': False, 'shape': (2, 4), }                                                          \n\xcd\xcc\x8c?\xcd\xcc\x0c@33S@\xcd\xcc\x8c@ff\x1eA\xcd\xcc\x0cAff\xf6@33\xd3@"
-        )
-
-    with _raises("numpy array error: Expected a data size of 32, found 31"):
-        vec_npy_each(
-            b"\x93NUMPY\x01\x00v\x00{'descr': '<f4', 'fortran_order': False, 'shape': (2, 4), }                                                          \n\xcd\xcc\x8c?\xcd\xcc\x0c@33S@\xcd\xcc\x8c@ff\x1eA\xcd\xcc\x0cAff\xf6@33\xd3"
-        )
-
-    # with _raises("XXX"):
-    #    vec_npy_each(b"\x93NUMPY\x01\x00v\x00{'descr': '<f4', 'fortran_order': False, 'shape': (2, 4), }                                                          \n\xcd\xcc\x8c?\xcd\xcc\x0c@33S@\xcd\xcc\x8c@ff\x1eA\xcd\xcc\x0cAff\xf6@33\xd3@")
-
-
-import tempfile
-
-
-def test_vec_npy_each_errors_files():
-    db = connect(EXT_PATH, extra_entrypoint="sqlite3_vec_numpy_init")
-
-    def vec_npy_each(data):
-        with tempfile.NamedTemporaryFile(delete_on_close=False) as f:
-            f.write(data)
-            f.close()
-            try:
-                return execute_all(
-                    db, "select rowid, * from vec_npy_each(vec_npy_file(?))", [f.name]
-                )
-            finally:
-                f.close()
-
-    with _raises("Could not open numpy file"):
-        db.execute('select * from vec_npy_each(vec_npy_file("not exist"))')
-
-    with _raises("numpy array file too short"):
-        vec_npy_each(b"\x93NUMPY\x01\x00v")
-
-    with _raises("numpy array file does not contain the 'magic' header"):
-        vec_npy_each(b"\x93XUMPY\x01\x00v\x00")
-
-    with _raises("numpy array file header length is invalid"):
-        vec_npy_each(b"\x93NUMPY\x01\x00v\x00")
-
-    with _raises(
-        "Error parsing numpy array: Only fortran_order = False is supported in sqlite-vec numpy functions"
-    ):
-        vec_npy_each(
-            b"\x93NUMPY\x01\x00v\x00{'descr': '<f4', 'fortran_order': True, 'shape': (2, 4), }                                                          \n\xcd\xcc\x8c?\xcd\xcc\x0c@33S@\xcd\xcc\x8c@ff\x1eA\xcd\xcc\x0cAff\xf6@33\xd3@"
-        )
-
-    with _raises("numpy array file error: Expected a data size of 32, found 31"):
-        vec_npy_each(
-            b"\x93NUMPY\x01\x00v\x00{'descr': '<f4', 'fortran_order': False, 'shape': (2, 4), }                                                          \n\xcd\xcc\x8c?\xcd\xcc\x0c@33S@\xcd\xcc\x8c@ff\x1eA\xcd\xcc\x0cAff\xf6@33\xd3"
-        )
-
-    assert vec_npy_each(to_npy(np.array([1.1, 2.2, 3.3], dtype=np.float32))) == [
-        {
-            "rowid": 0,
-            "vector": _f32([1.1, 2.2, 3.3]),
-        },
-    ]
-    assert vec_npy_each(
-        to_npy(np.array([[1.1, 2.2, 3.3], [4.4, 5.5, 6.6]], dtype=np.float32))
-    ) == [
-        {
-            "rowid": 0,
-            "vector": _f32([1.1, 2.2, 3.3]),
-        },
-        {
-            "rowid": 1,
-            "vector": _f32([4.4, 5.5, 6.6]),
-        },
-    ]
-    assert vec_npy_each(to_npy(np.array([], dtype=np.float32))) == []
-    x1025 = vec_npy_each(to_npy(np.array([[0.1, 0.2, 0.3]] * 1025, dtype=np.float32)))
-    assert len(x1025) == 1025
-
-    # np.array([[.1, .2, 3]] * 99, dtype=np.float32).shape
-
-
 def test_vec0_constructor():
     vec_constructor_error_prefix = "vec0 constructor error: {}"
     vec_col_error_prefix = "vec0 constructor error: could not parse vector column '{}'"
@@ -1923,6 +1590,54 @@ def test_vec0_constructor():
         db.execute("create virtual table v using vec0(4)")
 
 
+def test_vec0_indexed_by_flat():
+    db.execute("drop table if exists t_ibf")
+    db.execute("drop table if exists t_ibf2")
+    db.execute("drop table if exists t_ibf3")
+    db.execute("drop table if exists t_ibf4")
+
+    # indexed by flat() should succeed and behave identically to no index clause
+    db.execute("create virtual table t_ibf using vec0(emb float[4] indexed by flat())")
+    db.execute(
+        "insert into t_ibf(rowid, emb) values (1, X'00000000000000000000000000000000')"
+    )
+    rows = db.execute("select rowid from t_ibf where emb match X'00000000000000000000000000000000' and k = 1").fetchall()
+    assert len(rows) == 1
+    assert rows[0][0] == 1
+    db.execute("drop table t_ibf")
+
+    # indexed by flat() with distance_metric
+    db.execute(
+        "create virtual table t_ibf2 using vec0(emb float[4] distance_metric=cosine indexed by flat())"
+    )
+    db.execute("drop table t_ibf2")
+
+    # indexed by flat() on int8
+    db.execute("create virtual table t_ibf3 using vec0(emb int8[4] indexed by flat())")
+    db.execute("drop table t_ibf3")
+
+    # indexed by flat() on bit
+    db.execute("create virtual table t_ibf4 using vec0(emb bit[8] indexed by flat())")
+    db.execute("drop table t_ibf4")
+
+    # Error: unknown index type
+    with _raises(
+        "vec0 constructor error: could not parse vector column 'emb float[4] indexed by unknown()'",
+        sqlite3.DatabaseError,
+    ):
+        db.execute("create virtual table v using vec0(emb float[4] indexed by unknown())")
+
+    # Error: indexed by (missing type)
+    with _raises(
+        "vec0 constructor error: could not parse vector column 'emb float[4] indexed by'",
+        sqlite3.DatabaseError,
+    ):
+        db.execute("create virtual table v using vec0(emb float[4] indexed by)")
+
+    if db.in_transaction:
+        db.rollback()
+
+
 def test_vec0_create_errors():
     # EVIDENCE-OF: V17740_01811 vec0 create _chunks error handling
     db.set_authorizer(authorizer_deny_on(sqlite3.SQLITE_CREATE_TABLE, "t1_chunks"))

tests/test-metadata.py

@@ -265,6 +265,35 @@ def test_deletes(db, snapshot):
     assert vec0_shadow_table_contents(db, "v") == snapshot()
 
 
+def test_delete_by_metadata_with_long_text(db):
+    """Regression for https://github.com/asg017/sqlite-vec/issues/274.
+
+    ClearMetadata left rc=SQLITE_DONE after the long-text DELETE, which
+    propagated as an error and silently aborted the DELETE scan.
+    """
+    db.execute(
+        "create virtual table v using vec0("
+        "  tag text, embedding float[4], chunk_size=8"
+        ")"
+    )
+    for i in range(6):
+        db.execute(
+            "insert into v(tag, embedding) values (?, zeroblob(16))",
+            [f"long_text_value_{i}"],
+        )
+    for i in range(4):
+        db.execute(
+            "insert into v(tag, embedding) values (?, zeroblob(16))",
+            [f"long_text_value_0"],
+        )
+    assert db.execute("select count(*) from v").fetchone()[0] == 10
+
+    # DELETE by metadata WHERE — the pattern from the issue
+    db.execute("delete from v where tag = 'long_text_value_0'")
+    assert db.execute("select count(*) from v where tag = 'long_text_value_0'").fetchone()[0] == 0
+    assert db.execute("select count(*) from v").fetchone()[0] == 5
+
+
 def test_knn(db, snapshot):
     db.execute(
         "create virtual table v using vec0(vector float[1], name text, chunk_size=8)"