refactor: opencode

deploy/systemd/iai-mcp-daemon.service

@@ -16,13 +16,16 @@ After=default.target
 
 [Service]
 Type=simple
-ExecStart=/usr/bin/python3 -m iai_mcp.daemon
+ExecStart=/home/andreas/.venv/iai-mcp/bin/python -m iai_mcp.daemon
 Restart=on-failure
 RestartSec=30
 StartLimitIntervalSec=60
 StartLimitBurst=3
 
 Environment="IAI_MCP_STORE=%h/.iai-mcp"
+Environment="QDRANT_URL=http://192.168.0.22:6333"
+Environment="QDRANT_API_KEY=1CerixWX$3zdlj"
+Environment="IAI_MCP_EMBED_MODEL=bge-m3"
 Environment="LANG=en_US.UTF-8"
 
 StandardOutput=journal

mcp-wrapper/src/tools.ts

@@ -250,8 +250,18 @@ export const toolSchemas: Record<ToolName, ToolSchema> = {
       properties: {
         kind: {
           type: "string",
+          enum: [
+            "s4_contradiction",
+            "trajectory_metric",
+            "schema_induction_run",
+            "llm_health",
+            "curiosity_silent_log",
+            "curiosity_question",
+            "cls_consolidation_run",
+            "crypto_key_rotated",
+          ],
           description:
-            "Event kind. Must be in the whitelist (see tool description).",
+            "Event kind — must be one of the enum values above.",
         },
         since: {
           type: "string",

scripts/install.sh

@@ -31,13 +31,18 @@ die()  { printf '\n\033[0;31m✗ %s\033[0m\n' "$*" >&2; exit 1; }
 # IAI_TEST_SKIP_BUILD=1 short-circuits the whole bootstrap so the LaunchAgent
 # section (6) can be exercised in isolation by tests/test_install_uninstall.py
 # (Plan 07.1-03 Task 3) without spending ~30s on venv + npm.
+#
+# VENV_BASE: all venvs live in ~/.venv/<project-name> for central backup.
 # ---------------------------------------------------------------------------
+VENV_BASE="${HOME}/.venv"
+VENV_PATH="${VENV_BASE}/iai-mcp"
+
 if [[ "${IAI_TEST_SKIP_BUILD:-0}" == "1" ]]; then
     step "build skip (IAI_TEST_SKIP_BUILD=1)"
     ok "skipping sections 1-4 (venv/pip/npm/symlink) — test mode"
 else
     # -----------------------------------------------------------------------
-    # 1. venv
+    # 1. venv (central location: ~/.venv/iai-mcp)
     # -----------------------------------------------------------------------
     step "python venv"
     # iai-mcp requires Python 3.11 or 3.12 (torch + lancedb on 3.13/3.14
@@ -60,19 +65,19 @@ else
     fi
     [ -n "$PY" ] || die "Python 3.11 or 3.12 not found. macOS:  brew install python@3.12   |   Linux:  apt install python3.12 (or use pyenv)"
     ok "using $PY ($($PY --version))"
-    if [ ! -d .venv ]; then
+    if [ ! -d "${VENV_PATH}" ]; then
-        "$PY" -m venv .venv
+        "$PY" -m venv "${VENV_PATH}"
-        ok ".venv created"
+        ok "venv created at ${VENV_PATH}"
     else
-        ok ".venv already exists"
+        ok "venv already exists at ${VENV_PATH}"
     fi
 
     # -----------------------------------------------------------------------
     # 2. editable install
     # -----------------------------------------------------------------------
     step "editable install (pip -e .)"
-    .venv/bin/pip install --quiet --upgrade pip
+    "${VENV_PATH}"/bin/pip install --quiet --upgrade pip
-    .venv/bin/pip install --quiet -e .
+    "${VENV_PATH}"/bin/pip install --quiet -e .
     ok "iai-mcp python package installed into venv"
 
     # -----------------------------------------------------------------------
@@ -99,7 +104,7 @@ else
     step "global CLI symlink"
     LOCAL_BIN="${HOME}/.local/bin"
     LINK_PATH="${LOCAL_BIN}/iai-mcp"
-    TARGET="${REPO_ROOT}/.venv/bin/iai-mcp"
+    TARGET="${VENV_PATH}/bin/iai-mcp"
 
     [ -x "${TARGET}" ] || die "venv entry point not found at ${TARGET}"
 
@@ -114,7 +119,7 @@ else
     ok "${LINK_PATH} -> ${TARGET}"
 
     # PATH sanity check using python (grep is hook-blocked in this dev env).
-    PATH_HAS_LOCAL_BIN="$(.venv/bin/python - <<PY
+    PATH_HAS_LOCAL_BIN="$("${VENV_PATH}"/bin/python - <<PY
 import os
 print("1" if "${LOCAL_BIN}" in os.environ.get("PATH", "").split(":") else "0")
 PY
@@ -163,7 +168,7 @@ if [[ "$(uname)" != "Darwin" ]]; then
 elif [[ "${DRY_RUN:-0}" == "1" ]]; then
     ok "DRY_RUN=1 — skipping launchctl calls (test mode)"
 else
-    PYTHON_PATH="${REPO_ROOT}/.venv/bin/python"
+    PYTHON_PATH="${VENV_PATH}/bin/python"
     if [ ! -x "${PYTHON_PATH}" ]; then
         warn "venv python not found at ${PYTHON_PATH} — falling back to $(command -v python3)"
         PYTHON_PATH="$(command -v python3)"

scripts/rebuild-venv.sh

@@ -0,0 +1,43 @@
+#!/usr/bin/env bash
+# Rebuild iai-mcp venv for non-AVX CPU (no AVX support on this machine)
+# Run from repo root: bash scripts/rebuild-venv.sh
+
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+REPO_ROOT="$(cd "${SCRIPT_DIR}/.." && pwd)"
+cd "${REPO_ROOT}"
+
+VENV_PATH="${HOME}/.venv/iai-mcp"
+
+echo "==> Removing old venv"
+rm -rf "${VENV_PATH}"
+
+echo "==> Creating fresh venv"
+/usr/bin/python3.12 -m venv "${VENV_PATH}"
+
+echo "==> Upgrading pip"
+"${VENV_PATH}"/bin/pip install --quiet --upgrade pip
+
+echo "==> Installing torch CPU-only (no AVX needed)"
+"${VENV_PATH}"/bin/pip install --quiet \
+  --extra-index-url https://download.pytorch.org/whl/cpu \
+  torch torchvision torchaudio
+
+echo "==> Installing iai-mcp dependencies"
+"${VENV_PATH}"/bin/pip install --quiet -e .
+
+echo "==> Building TS wrapper"
+if [ -d mcp-wrapper ]; then
+    pushd mcp-wrapper >/dev/null
+    npm ci --silent --no-audit --no-fund
+    npm run build --silent
+    popd >/dev/null
+    echo "   ✓ mcp-wrapper/dist built"
+else
+    echo "   ! mcp-wrapper/ missing"
+fi
+
+echo "==> Done"
+echo "   venv: ${VENV_PATH}"
+echo "   test: ${VENV_PATH}/bin/python -c 'import torch; print(torch.__version__)'"

src/iai_mcp/core.py

@@ -222,7 +222,7 @@ def dispatch(store: MemoryStore, method: str, params: dict) -> dict:
 
         # Plan 02 dispatch: non-empty store -> 5-stage pipeline;
         # empty store -> baseline cosine recall (Plan 01 fallback).
-        records_count = store.db.open_table("records").count_rows()
+        records_count = store.count_rows("records")
         if records_count == 0:
             cue_embedding = params.get("cue_embedding") or [0.0] * EMBED_DIM
             resp = retrieve.recall(
@@ -675,7 +675,7 @@ def dispatch(store: MemoryStore, method: str, params: dict) -> dict:
     if method == "topology":
         from iai_mcp import sigma as sigma_mod
 
-        records_count = store.db.open_table("records").count_rows()
+        records_count = store.count_rows("records")
         if records_count == 0:
             return {
                 "N": 0, "C": 0.0, "L": 0.0, "sigma": None,
@@ -741,7 +741,7 @@ def dispatch(store: MemoryStore, method: str, params: dict) -> dict:
         # wake_depth knob reaches the assembler.
         from iai_mcp.session import assemble_session_start, SessionStartPayload
         sid = params.get("session_id", "-")
-        records_count = store.db.open_table("records").count_rows()
+        records_count = store.count_rows("records")
         if records_count == 0:
             empty = SessionStartPayload(
                 l0="",
@@ -810,7 +810,7 @@ def _schema_list_dispatch(store: MemoryStore, params: dict) -> dict:
     records = store.all_records()
     schema_records = [r for r in records if "schema" in (r.tags or [])]
 
-    edges_df = store.db.open_table("edges").to_pandas()
+    edges_df = store.edges_as_dataframe()
     if not edges_df.empty:
         schema_edges = edges_df[edges_df["edge_type"] == "schema_instance_of"]
     else:

src/iai_mcp/daemon.py

@@ -51,7 +51,7 @@ from iai_mcp.quiet_window import (
     should_relearn,
 )
 from iai_mcp.socket_server import SocketServer
-from iai_mcp.store import MemoryStore
+from iai_mcp.store import MemoryStore, get_store, _use_qdrant
 from iai_mcp.tz import load_user_tz
 
 # ---------------------------------------------------------------------------
@@ -1076,7 +1076,7 @@ async def main() -> int:
     # consistency — each read re-checks the latest committed version at
     # negligible cost (one manifest stat per query) and restores the
     # tick body's ability to see work.
-    store = MemoryStore(read_consistency_interval=timedelta(seconds=0))
+    store = get_store(read_consistency_interval=timedelta(seconds=0))
 
     try:
         from iai_mcp.crypto_key_watch import check_crypto_key_file_rotation_event
@@ -1098,7 +1098,10 @@ async def main() -> int:
     #   - partial_swap_inconsistent -> STOP daemon; surface remediation prompt
     #                                  (manual recovery; no rollback anchor).
     from iai_mcp.migrate import detect_partial_migration
-    _migration_state = detect_partial_migration(store.db)
+    if _use_qdrant():
+        _migration_state = {"state": "clean"}
+    else:
+        _migration_state = detect_partial_migration(store.db)
     if _migration_state["state"] == "partial_swap_inconsistent":
         try:
             sys.stderr.write(

src/iai_mcp/embed.py

@@ -1,4 +1,4 @@
-"""Embedding layer -- configurable embedder with a 3-model registry.
+"""Embedding layer -- configurable embedder with a 4-model registry + remote.
 
 Plan 05-08 (2026-04-20): the DEFAULT is now ``bge-small-en-v1.5`` (384d
 English-only), reverting the Phase-2 deviation. PROJECT.md line
@@ -8,11 +8,12 @@ swapped in bge-m3 (1024d multilingual) as D-08a. User directive
 job. bge-m3 stays selectable via env var / kwarg for anyone who needs
 multilingual semantic match at the 5x RAM cost.
 
-Configurable 4-model registry:
+Configurable 4-model registry (local) + remote OpenAI-compatible endpoint:
 - "bge-m3"                 -> BAAI/bge-m3               -> 1024d (opt-in, multilingual)
 - "multilingual-e5-small"  -> intfloat/multilingual-e5-small -> 384d (compromise)
 - "bge-small-en-v1.5"      -> BAAI/bge-small-en-v1.5    -> 384d (DEFAULT, English)
 - "all-MiniLM-L6-v2"       -> sentence-transformers/all-MiniLM-L6-v2 -> 384d (English alternative embedder option; included for compatibility testing)
+- "remote-bge-m3"          -> OpenAI-compatible API     -> 1024d (remote, no local model load)
 
 Selection priority at Embedder() instantiation:
 1. Explicit `model_key` constructor arg
@@ -31,14 +32,23 @@ from __future__ import annotations
 import os
 import threading
 
+import httpx
 from sentence_transformers import SentenceTransformer
 
 
-# 4-model registry. Name convention: short logical key -> HF repo id + dim.
+# 4-model registry + remote entry. Name convention: short logical key -> HF
+# repo id / endpoint + dim.
 # (2026-04-29): all-MiniLM-L6-v2 added as additive ablation entry;
 # DEFAULT_MODEL_KEY unchanged (English-Only Brain lock from / Plan 05-08).
+# (2026-05-11): bge-m3 configured as remote (non-AVX CPU) — delegates embedding
+# to an OpenAI-compatible server (bge-m3 @ 1024d).
 MODEL_REGISTRY: dict[str, dict] = {
-    "bge-m3": {"hf": "BAAI/bge-m3", "dim": 1024},
+    "bge-m3": {
+        "endpoint": "http://192.168.0.50:12434/v1/embeddings",
+        "model": "bge-m3",
+        "dim": 1024,
+        "remote": True,
+    },
     "multilingual-e5-small": {"hf": "intfloat/multilingual-e5-small", "dim": 384},
     "bge-small-en-v1.5": {"hf": "BAAI/bge-small-en-v1.5", "dim": 384},
     "all-MiniLM-L6-v2": {"hf": "sentence-transformers/all-MiniLM-L6-v2", "dim": 384},
@@ -64,6 +74,11 @@ def _resolve_model_key(model_key: str | None = None) -> str:
     return DEFAULT_MODEL_KEY
 
 
+def _is_remote_model(model_key: str) -> bool:
+    """Check if a model key refers to a remote embedder."""
+    return MODEL_REGISTRY.get(model_key, {}).get("remote", False)
+
+
 _MODEL_LOCK = threading.Lock()
 _MODEL_CACHE: dict[str, SentenceTransformer] = {}
 
@@ -158,7 +173,90 @@ class Embedder:
         return [v.tolist() for v in vecs]
 
 
-def embedder_for_store(store) -> "Embedder":
+class RemoteEmbedder:
+    """Embedder that delegates to an OpenAI-compatible remote endpoint.
+
+    Used when the local CPU cannot run sentence-transformers (e.g. no AVX).
+    Sends text to a remote bge-m3 instance and returns L2-normalised 1024d
+    vectors.
+
+    The remote endpoint must speak the OpenAI `/v1/embeddings` protocol:
+      POST /v1/embeddings
+      {"model": "bge-m3", "input": ["text"]}
+      -> {"data": [{"embedding": [0.0, ...], ...}]}
+    """
+
+    def __init__(
+        self,
+        model_key: str | None = None,
+        *,
+        endpoint: str | None = None,
+        model_name: str | None = None,
+    ) -> None:
+        if model_key is not None and model_key in MODEL_REGISTRY:
+            spec = MODEL_REGISTRY[model_key]
+            self.model_key: str = model_key
+            self._endpoint: str = spec["endpoint"]
+            self._model_name: str = spec["model"]
+            self.DIM: int = int(spec["dim"])
+        elif endpoint is not None and model_name is not None:
+            self.model_key = "custom-remote"
+            self._endpoint = endpoint
+            self._model_name = model_name
+            # Discover dim from a probe call
+            self.DIM = self._probe_dim()
+        else:
+            raise ValueError(
+                "RemoteEmbedder requires model_key from MODEL_REGISTRY "
+                "or explicit endpoint + model_name"
+            )
+
+        self._client = httpx.Client(timeout=30.0)
+
+    def _probe_dim(self) -> int:
+        """Make a single embedding call to discover the output dimension."""
+        resp = self._client.post(
+            self._endpoint,
+            json={"model": self._model_name, "input": ["probe"]},
+        )
+        resp.raise_for_status()
+        data = resp.json()
+        return len(data["data"][0]["embedding"])
+
+    def embed(self, text: str) -> list[float]:
+        """Encode a single string. Returns L2-normalised vector."""
+        resp = self._client.post(
+            self._endpoint,
+            json={"model": self._model_name, "input": [text]},
+        )
+        resp.raise_for_status()
+        data = resp.json()
+        vec = data["data"][0]["embedding"]
+        # Normalise if not already (bge-m3 on Ollama returns normalised)
+        norm = (sum(x * x for x in vec)) ** 0.5
+        if norm > 0:
+            vec = [x / norm for x in vec]
+        return vec
+
+    def embed_batch(self, texts: list[str]) -> list[list[float]]:
+        """Batch-encode preserving input order."""
+        resp = self._client.post(
+            self._endpoint,
+            json={"model": self._model_name, "input": texts},
+        )
+        resp.raise_for_status()
+        data = resp.json()
+        results = []
+        for item in data["data"]:
+            vec = item["embedding"]
+            norm = (sum(x * x for x in vec)) ** 0.5
+            if norm > 0:
+                vec = [x / norm for x in vec]
+            results.append(vec)
+        return results
+
+
+def embedder_for_store(store) -> "Embedder | RemoteEmbedder":
     """Store-aware Embedder factory. Picks the model whose output dim matches
     the existing LanceDB records schema, so a legacy 1024d store from the
     pre-Plan-05-08 bge-m3 era stays queryable until it is re-embedded down to
@@ -168,14 +266,24 @@ def embedder_for_store(store) -> "Embedder":
     1. If store.embed_dim has an exact match in MODEL_REGISTRY, prefer the
        model whose logical key name indicates the canonical model at that dim
        (bge-small-en-v1.5 for 384d default; bge-m3 for legacy/opt-in 1024d).
-    2. Otherwise fall through to the env/registry default via Embedder().
+    2. If IAI_MCP_EMBED_MODEL points to a remote model, use RemoteEmbedder.
+    3. Otherwise fall through to the env/registry default via Embedder().
 
     This decouples runtime model selection from a global env var so a single
     process can operate multiple stores at different dims while the migration
     from a legacy 1024d store down to 384d completes.
     """
     target_dim = getattr(store, "embed_dim", None)
+    env_key = os.environ.get("IAI_MCP_EMBED_MODEL")
+
+    # Check if user explicitly requested remote embedder
+    if env_key and _is_remote_model(env_key):
+        return RemoteEmbedder(model_key=env_key)
+
     if target_dim is None:
+        # No existing store — check if remote is requested
+        if env_key and _is_remote_model(env_key):
+            return RemoteEmbedder(model_key=env_key)
         return Embedder()
     preferred = {384: "bge-small-en-v1.5", 1024: "bge-m3"}
     key = preferred.get(int(target_dim))
@@ -184,10 +292,16 @@ def embedder_for_store(store) -> "Embedder":
     # stays compatible; real production code still respects store.embed_dim.
     try:
         if key is not None and key in MODEL_REGISTRY:
+            if _is_remote_model(key):
+                return RemoteEmbedder(model_key=key)
             return Embedder(model_key=key)
         for reg_key, spec in MODEL_REGISTRY.items():
             if int(spec["dim"]) == int(target_dim):
+                if _is_remote_model(reg_key):
+                    return RemoteEmbedder(model_key=reg_key)
                 return Embedder(model_key=reg_key)
     except TypeError:
         pass
+    if env_key and _is_remote_model(env_key):
+        return RemoteEmbedder(model_key=env_key)
     return Embedder()

src/iai_mcp/events.py

@@ -92,17 +92,17 @@ def write_event(
     data_plain = json.dumps(data)
     ad = str(event_id).encode("ascii")
     data_ct = encrypt_field(data_plain, store._key(), associated_data=ad)
-    row = {
+    ts = datetime.now(timezone.utc)
-        "id": str(event_id),
+    store.events_add(
-        "kind": kind,
+        event_id=event_id,
-        "severity": severity or "",
+        kind=kind,
-        "domain": domain or "",
+        severity=severity or "",
-        "ts": datetime.now(timezone.utc),
+        domain=domain or "",
-        "data_json": data_ct,
+        ts=ts,
-        "session_id": session_id,
+        data_json=data_ct,
-        "source_ids_json": json.dumps([str(x) for x in (source_ids or [])]),
+        session_id=session_id,
-    }
+        source_ids_json=json.dumps([str(x) for x in (source_ids or [])]),
-    store.db.open_table(EVENTS_TABLE).add([row])
+    )
     return event_id
 
 
@@ -132,27 +132,12 @@ def query_events(
     Returns a list of dicts with keys: id, kind, severity, domain, ts, data,
     session_id, source_ids. data and source_ids are decoded from JSON.
     """
-    tbl = store.db.open_table(EVENTS_TABLE)
+    rows = store.events_query(kind=kind, since=since, severity=severity, limit=limit)
-    df = tbl.to_pandas()
-    if df.empty:
-        return []
-    if kind is not None:
-        df = df[df["kind"] == kind]
-    if severity is not None:
-        df = df[df["severity"] == severity]
-    if since is not None:
-        # Ensure tz-aware comparison
-        since_cmp = since if since.tzinfo is not None else since.replace(tzinfo=timezone.utc)
-        # Pandas Timestamp compares naturally with tz-aware datetimes
-        df = df[df["ts"] >= since_cmp]
-    if df.empty:
-        return []
-    df = df.sort_values("ts", ascending=False).head(limit)
     out: list[dict] = []
-    for _, row in df.iterrows():
+    for row in rows:
         # decrypt data_json when it carries the iai:enc:v1: prefix.
         # Pre-02-08 rows stay plaintext; migration rewrites them lazily.
-        raw_data = row["data_json"] or "{}"
+        raw_data = row["data"] if isinstance(row.get("data"), str) else json.dumps(row.get("data", {}))
         if is_encrypted(raw_data):
             ad = str(row["id"]).encode("ascii")
             try:
@@ -165,20 +150,14 @@ def query_events(
             data = json.loads(raw_data)
         except (TypeError, json.JSONDecodeError):
             data = {}
-        try:
+        out.append({
-            source_ids = json.loads(row["source_ids_json"] or "[]")
+            "id": row["id"],
-        except (TypeError, json.JSONDecodeError):
+            "kind": row["kind"],
-            source_ids = []
+            "severity": row["severity"] or None,
-        out.append(
+            "domain": row["domain"] or None,
-            {
+            "ts": row["ts"],
-                "id": row["id"],
+            "data": data,
-                "kind": row["kind"],
+            "session_id": row["session_id"],
-                "severity": row["severity"] or None,
+            "source_ids": row["source_ids"],
-                "domain": row["domain"] or None,
+        })
-                "ts": row["ts"],
-                "data": data,
-                "session_id": row["session_id"],
-                "source_ids": source_ids,
-            }
-        )
     return out

src/iai_mcp/migrate.py

@@ -61,10 +61,7 @@ from pathlib import Path
 from typing import Callable, Optional
 from uuid import UUID
 
-import pyarrow as pa
-
 from iai_mcp.crypto import encrypt_field, is_encrypted
-from iai_mcp.embed import Embedder
 from iai_mcp.events import write_event
 from iai_mcp.store import (
     EVENTS_TABLE,
@@ -123,7 +120,7 @@ def _detect_language(text: str) -> str:
 
 def migrate_v1_to_v2(
     store: MemoryStore,
-    embedder: Optional[Embedder] = None,
+    embedder: Optional["Embedder"] = None,
     dry_run: bool = False,
     progress: Optional[Callable[[int, int], None]] = None,
 ) -> dict:
@@ -237,7 +234,7 @@ def migrate_v1_to_v2(
     }
 
 
-def _records_schema_at_dim(dim: int) -> pa.Schema:
+def _records_schema_at_dim(dim: int) -> "pa.Schema":
     """Build the records-table Arrow schema at an explicit embedding dim.
 
     Mirrors `MemoryStore._ensure_tables` lines 249-281 byte-for-byte except
@@ -247,6 +244,7 @@ def _records_schema_at_dim(dim: int) -> pa.Schema:
     is not parameterised on dim. Plan 07.11-03 / file-disjoint
     constraint forbids store.py changes; inlining is the conservative path.
     """
+    import pyarrow as pa
     return pa.schema(
         [
             ("id", pa.string()),

src/iai_mcp/migrate_qdrant.py

@@ -0,0 +1,202 @@
+"""Migration script: move data from 5 Qdrant collections → 2 collections.
+
+Old structure (5 collections):
+- `records`        : MemoryRecord rows (1024-dim vectors)
+- `edges`          : Graph edges (1-dim dummy vectors)
+- `events`         : Runtime events (1-dim dummy vectors)
+- `budget_ledger`  : D-GUARD spend tracking (1-dim dummy vectors)
+- `ratelimit_ledger`: D-GUARD rate limit history (1-dim dummy vectors)
+
+New structure (2 collections, per Qdrant best practices):
+- `records`        : MemoryRecord rows (1024-dim cosine vectors)
+                     All points carry `table: "records"` + `group_id` payload.
+- `metadata`       : Payload-only (no vectors) containing edges, events,
+                     budget_ledger, ratelimit_ledger.
+                     Each point carries `table` + `group_id` payload.
+
+Both collections use keyword indexes on `table` for co-located storage.
+
+Usage:
+    python -m iai_mcp.migrate_qdrant
+
+Environment:
+    QDRANT_URL    : Qdrant server URL (default: http://192.168.0.22:6333)
+    QDRANT_API_KEY: Qdrant API key
+"""
+from __future__ import annotations
+
+import base64
+import json
+import os
+import sys
+import time
+from datetime import datetime, timezone
+from uuid import UUID
+
+from qdrant_client import QdrantClient
+from qdrant_client.models import Distance, PointStruct, VectorParams
+from qdrant_client.http.exceptions import UnexpectedResponse
+
+# --------------------------------------------------------------------------- env
+QDRANT_URL = os.environ.get("QDRANT_URL", "http://192.168.0.22:6333")
+QDRANT_API_KEY = os.environ.get("QDRANT_API_KEY")
+GROUP_ID = os.environ.get("IAI_MCP_USER_ID", "default")
+
+
+def setup_client() -> QdrantClient:
+    """Create Qdrant client with API key."""
+    return QdrantClient(url=QDRANT_URL, api_key=QDRANT_API_KEY, timeout=30)
+
+
+def ensure_new_collections(client: QdrantClient) -> None:
+    """Create the 2 new collections if they don't exist."""
+    # Collection 1: records (vectors)
+    try:
+        client.get_collection("records")
+        print("  records collection already exists")
+    except Exception:
+        print("  creating records collection...")
+        client.create_collection(
+            collection_name="records",
+            vectors_config=VectorParams(size=1024, distance=Distance.COSINE),
+        )
+
+    # Collection 2: metadata (payload-only)
+    try:
+        client.get_collection("metadata")
+        print("  metadata collection already exists")
+    except Exception:
+        print("  creating metadata collection...")
+        client.create_collection(collection_name="metadata")
+
+    # Create payload indexes
+    for collection_name in ("records", "metadata"):
+        for field_name in ("table", "group_id"):
+            try:
+                client.create_payload_index(
+                    collection_name=collection_name,
+                    field_name=field_name,
+                    field_schema="keyword",
+                )
+            except Exception:
+                pass  # index may already exist
+
+
+def scroll_all(client: QdrantClient, collection_name: str, batch_size: int = 1000) -> list:
+    """Scroll through all points in a collection."""
+    offset = None
+    all_points = []
+    while True:
+        points, next_offset = client.scroll(
+            collection_name=collection_name,
+            limit=batch_size,
+            offset=offset,
+            with_payload=True,
+            with_vectors=True,
+        )
+        all_points.extend(points)
+        if next_offset is None:
+            break
+        offset = next_offset
+    return all_points
+
+
+def migrate_records(client: QdrantClient) -> int:
+    """Migrate records from old `records` collection to new `records` collection."""
+    print("\nMigrating records...")
+    old_points = scroll_all(client, "records")
+    if not old_points:
+        print("  no records to migrate")
+        return 0
+
+    new_points = []
+    for pt in old_points:
+        payload = pt.payload or {}
+        # Add table and group_id
+        payload["table"] = "records"
+        payload["group_id"] = GROUP_ID
+        new_points.append(PointStruct(
+            id=pt.id,
+            vector=list(pt.vector) if pt.vector else [],
+            payload=payload,
+        ))
+
+    client.upsert(collection_name="records", points=new_points)
+    print(f"  migrated {len(new_points)} records")
+    return len(new_points)
+
+
+def migrate_metadata(client: QdrantClient, table_name: str) -> int:
+    """Migrate points from an old collection to the new `metadata` collection."""
+    print(f"\nMigrating {table_name}...")
+    old_points = scroll_all(client, table_name)
+    if not old_points:
+        print(f"  no {table_name} points to migrate")
+        return 0
+
+    new_points = []
+    for pt in old_points:
+        payload = pt.payload or {}
+        # Add table and group_id
+        payload["table"] = table_name
+        payload["group_id"] = GROUP_ID
+        new_points.append(PointStruct(
+            id=pt.id,
+            vector={},  # payload-only (empty dict for no-vector collection)
+            payload=payload,
+        ))
+
+    client.upsert(collection_name="metadata", points=new_points)
+    print(f"  migrated {len(new_points)} {table_name} points")
+    return len(new_points)
+
+
+def drop_old_collections(client: QdrantClient) -> None:
+    """Drop the old collections after migration."""
+    old_collections = ["edges", "events", "budget_ledger", "ratelimit_ledger"]
+    for col_name in old_collections:
+        try:
+            client.delete_collection(collection_name=col_name, timeout=30)
+            print(f"  dropped {col_name} collection")
+        except Exception as e:
+            print(f"  warning: could not drop {col_name}: {e}")
+
+
+def main() -> int:
+    """Run the migration."""
+    print(f"Qdrant migration: 5 collections → 2 collections")
+    print(f"  QDRANT_URL: {QDRANT_URL}")
+    print(f"  GROUP_ID: {GROUP_ID}")
+
+    client = setup_client()
+    print("\nStep 1: Ensure new collections exist...")
+    ensure_new_collections(client)
+
+    print("\nStep 2: Migrate data...")
+    t0 = time.time()
+    total = 0
+    total += migrate_records(client)
+    total += migrate_metadata(client, "edges")
+    total += migrate_metadata(client, "events")
+    total += migrate_metadata(client, "budget_ledger")
+    total += migrate_metadata(client, "ratelimit_ledger")
+    print(f"\n  total migrated: {total} points in {time.time() - t0:.1f}s")
+
+    print("\nStep 3: Drop old collections...")
+    drop_old_collections(client)
+
+    print("\nStep 4: Verify...")
+    try:
+        rec_count = client.get_collection("records").points_count
+        meta_points = client.scroll("metadata", limit=1, with_payload=True)[0]
+        print(f"  records collection: {rec_count} points")
+        print(f"  metadata collection: exists")
+    except Exception as e:
+        print(f"  verification warning: {e}")
+
+    print("\nMigration complete!")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

src/iai_mcp/store.py

@@ -47,8 +47,7 @@ from collections.abc import Sequence
 from typing import Callable
 from uuid import UUID
 
-import lancedb
+
-import pyarrow as pa
 
 # W5: cached AESGCM cipher per store; reuse safe per
 # https://cryptography.io/en/latest/hazmat/primitives/aead/ — single AESGCM
@@ -209,6 +208,7 @@ class MemoryStore:
         connect_kwargs: dict[str, object] = {}
         if read_consistency_interval is not None:
             connect_kwargs["read_consistency_interval"] = read_consistency_interval
+        import lancedb
         self.db = lancedb.connect(str(self.root / "lancedb"), **connect_kwargs)
         # Resolve the embedding dimension once so records table + insert guard agree.
         self._embed_dim: int = _resolve_embed_dim()
@@ -1596,3 +1596,91 @@ class MemoryStore:
         if language == "__LEGACY_EMPTY__":
             rec.language = ""  # post-construction: signal to migration path
         return rec
+
+    # ---------------------------------------------------------- events
+
+    def events_add(self, event_id: UUID, kind: str, severity: str, domain: str,
+                   ts: datetime, data_json: str, session_id: str, source_ids_json: str) -> None:
+        """Add a single event row to the events table."""
+        row = {
+            "id": str(event_id),
+            "kind": kind,
+            "severity": severity,
+            "domain": domain,
+            "ts": ts,
+            "data_json": data_json,
+            "session_id": session_id,
+            "source_ids_json": source_ids_json,
+        }
+        self.db.open_table(EVENTS_TABLE).add([row])
+
+    def events_query(self, kind: str | None = None, since: datetime | None = None,
+                     severity: str | None = None, limit: int = 100) -> list[dict]:
+        """Query events matching filters, newest first."""
+        tbl = self.db.open_table(EVENTS_TABLE)
+        df = tbl.to_pandas()
+        if df.empty:
+            return []
+        if kind is not None:
+            df = df[df["kind"] == kind]
+        if severity is not None:
+            df = df[df["severity"] == severity]
+        if since is not None:
+            since_cmp = since if since.tzinfo is not None else since.replace(tzinfo=timezone.utc)
+            df = df[df["ts"] >= since_cmp]
+        if df.empty:
+            return []
+        df = df.sort_values("ts", ascending=False).head(limit)
+        out: list[dict] = []
+        for _, row in df.iterrows():
+            raw_data = row["data_json"] or "{}"
+            try:
+                data = json.loads(raw_data)
+            except (TypeError, json.JSONDecodeError):
+                data = {}
+            try:
+                source_ids = json.loads(row["source_ids_json"] or "[]")
+            except (TypeError, json.JSONDecodeError):
+                source_ids = []
+            out.append({
+                "id": row["id"],
+                "kind": row["kind"],
+                "severity": row["severity"] or None,
+                "domain": row["domain"] or None,
+                "ts": row["ts"],
+                "data": data,
+                "session_id": row["session_id"],
+                "source_ids": source_ids,
+            })
+        return out
+
+
+# --------------------------------------------------------------------------- Qdrant backend
+
+def _use_qdrant() -> bool:
+    """Check if Qdrant backend is configured via environment."""
+    return bool(os.environ.get("QDRANT_URL"))
+
+
+def get_store(
+    path: Path | str | None = None,
+    user_id: str = "default",
+    read_consistency_interval: timedelta | None = None,
+) -> "MemoryStore | QdrantStore":
+    """Factory: return MemoryStore (LanceDB) or QdrantStore based on env.
+
+    When QDRANT_URL is set, returns QdrantStore.
+    Otherwise returns MemoryStore (LanceDB) — the legacy/local path.
+    """
+    if _use_qdrant():
+        from iai_mcp.qdrant_store import QdrantStore
+        return QdrantStore(
+            path=path,
+            user_id=user_id,
+            read_consistency_interval=read_consistency_interval,
+        )
+    return MemoryStore(
+        path=path,
+        user_id=user_id,
+        read_consistency_interval=read_consistency_interval,
+    )