feat(noxa-68r.8): daemon binary + README

- noxa-rag-daemon: --config, --log-level, --version args (clap)
- Startup sequence: config load -> watch_dir create-if-missing -> embed probe
  -> vector store -> tokenizer -> Pipeline::new() -> signal handlers -> run()
- Embed dims probed dynamically (dummy embed call, no hardcode)
- tokenizer.json loaded from local_path (Rust tokenizers crate has no from_pretrained)
  - Clear error message with huggingface-cli download command
  - Accepts directory or direct file path
- Tracing to stderr (stdout may be piped)
- World-readable config warning on unix
- 10s force-exit timeout after CancellationToken cancel
- SIGTERM + Ctrl-C via tokio signal handlers
- README.md: CRITICAL TEI --pooling last-token warning, full config reference,
  quickstart, architecture diagram
- LEARNED: tokenizers Rust crate has no from_pretrained — local_path is required
This commit is contained in:
Jacob Magar 2026-04-12 07:23:33 -04:00
parent 9cde0d66ca
commit 5217b99601
2 changed files with 357 additions and 4 deletions

161
crates/noxa-rag/README.md Normal file
View file

@ -0,0 +1,161 @@
# noxa-rag
RAG pipeline for [noxa](https://github.com/jmagar/noxa) — watches noxa's output directory for `ExtractionResult` JSON files, chunks them, embeds via [HF TEI](https://github.com/huggingface/text-embeddings-inference), and upserts to [Qdrant](https://qdrant.tech/).
## System Requirements
- **Qdrant** running locally (gRPC port 6334)
- **HF TEI** with GPU (tested on RTX 4070)
- **CUDA** for TEI inference (CPU mode is possible but slow)
- **Rust 1.82+**
- **huggingface-cli** to download the tokenizer
## CRITICAL: TEI Launch Command
```bash
# CRITICAL: --pooling last-token is REQUIRED for Qwen3-0.6B
# Qwen3 is a decoder-only model. Mean pooling (TEI default) produces
# semantically incorrect embeddings. This flag is NOT optional.
docker run --gpus all -p 8080:80 \
ghcr.io/huggingface/text-embeddings-inference:latest \
--model-id Qwen/Qwen3-Embedding-0.6B \
--pooling last-token \
--max-batch-tokens 32768 \
--max-client-batch-size 128 \
--dtype float16
```
### Verify TEI is working
```bash
curl http://localhost:8080/health
# {"status":"ok"}
# Check embedding dimensions (must be 1024 for Qwen3-0.6B)
curl -s http://localhost:8080/embed \
-H "Content-Type: application/json" \
-d '{"inputs": ["test"], "normalize": true}' | python3 -c "import sys,json; v=json.load(sys.stdin)[0]; print(f'{len(v)} dims')"
# 1024 dims
```
## Quickstart
### 1. Download the tokenizer
The Rust `tokenizers` crate cannot download from HF Hub at runtime. Download once:
```bash
pip install huggingface_hub
huggingface-cli download Qwen/Qwen3-Embedding-0.6B tokenizer.json --local-dir ~/.cache/noxa-rag/tokenizer
```
### 2. Create config file
```toml
# noxa-rag.toml
[source]
type = "fs_watcher"
watch_dir = "/home/user/.noxa/output"
debounce_ms = 500
[embed_provider]
type = "tei"
url = "http://localhost:8080"
model = "Qwen/Qwen3-Embedding-0.6B"
# REQUIRED: path to directory containing tokenizer.json
local_path = "/home/user/.cache/noxa-rag/tokenizer"
[vector_store]
type = "qdrant"
# gRPC port 6334 (NOT 6333 which is REST)
url = "http://localhost:6334"
collection = "noxa_rag"
# api_key = "..." # or set NOXA_RAG_QDRANT_API_KEY env var
[chunker]
target_tokens = 512
overlap_tokens = 64
min_words = 50
max_chunks_per_page = 100
[pipeline]
embed_concurrency = 4
# Must be an absolute path (daemon may run with CWD = /)
failed_jobs_log = "/home/user/.noxa/noxa-rag-failed.jsonl"
```
### 3. Start Qdrant
```bash
docker run -p 6333:6333 -p 6334:6334 \
-v ~/.noxa/qdrant:/qdrant/storage \
qdrant/qdrant
```
### 4. Run the daemon
```bash
cargo build --release -p noxa-rag
./target/release/noxa-rag-daemon --config noxa-rag.toml
```
### 5. Index content with noxa
```bash
# Extract a page — the daemon will pick up the output file automatically
noxa https://docs.example.com --output ~/.noxa/output/
```
The daemon watches `watch_dir` for `.json` files. When noxa writes an `ExtractionResult` to that directory, the daemon detects it (within `debounce_ms` ms), chunks it, embeds it, and upserts to Qdrant.
## Configuration Reference
| Field | Default | Description |
|-------|---------|-------------|
| `source.watch_dir` | — | Directory to watch for `.json` files |
| `source.debounce_ms` | `500` | Debounce window for filesystem events (ms) |
| `embed_provider.url` | — | TEI server URL |
| `embed_provider.model` | — | Model name (used in logs) |
| `embed_provider.local_path` | **required** | Directory containing `tokenizer.json` |
| `vector_store.url` | — | Qdrant gRPC URL (port 6334) |
| `vector_store.collection` | — | Qdrant collection name |
| `vector_store.api_key` | `null` | Qdrant API key (or `NOXA_RAG_QDRANT_API_KEY` env var) |
| `chunker.target_tokens` | `512` | Target chunk size in tokens |
| `chunker.overlap_tokens` | `64` | Sliding window overlap tokens |
| `chunker.min_words` | `50` | Skip chunks shorter than this |
| `chunker.max_chunks_per_page` | `100` | Cap chunks per document |
| `pipeline.embed_concurrency` | `4` | Concurrent embed workers (must be > 0) |
| `pipeline.failed_jobs_log` | `null` | Absolute path for NDJSON error log |
## Architecture
```
noxa-cli (writes .json) → watch_dir
notify-debouncer-mini (500ms debounce)
bounded mpsc channel (256 capacity)
embed_concurrency worker tasks (default: 4)
┌─────────────────────────────────────┐
│ process_job() │
│ 1. Read file (TOCTOU-safe) │
│ 2. Parse ExtractionResult JSON │
│ 3. Validate URL scheme (http/https) │
│ 4. chunk() → Vec<Chunk>
│ 5. embed() → Vec<Vec<f32>> │
│ 6. UUID v5 point IDs │
│ 7. Per-URL mutex: delete + upsert │
└─────────────────────────────────────┘
Qdrant (gRPC)
```
## Notes
- **Vim/Emacs compatibility**: The daemon watches all filesystem events (not just Create/Modify). Atomic saves via rename are detected correctly.
- **Idempotent indexing**: Re-indexing the same URL deletes old chunks first (delete-before-upsert), so chunk count changes are handled correctly.
- **Point IDs**: UUID v5 deterministic — same URL + chunk index always produces the same Qdrant point ID.
- **Failed jobs**: Parse failures and oversized files (>50MB) are logged to `failed_jobs_log` as NDJSON and skipped (the daemon keeps running).

View file

@ -1,5 +1,197 @@
// noxa-rag-daemon — implemented in noxa-68r.8
fn main() {
eprintln!("[noxa-rag] daemon stub — not yet implemented (noxa-68r.8)");
std::process::exit(1);
/// noxa-rag-daemon — watches an output directory for ExtractionResult JSON files
/// and indexes them via TEI + Qdrant.
///
/// Usage:
/// noxa-rag-daemon [--config <PATH>] [--log-level <LEVEL>] [--version]
use std::path::PathBuf;
use std::sync::Arc;
use std::time::Duration;
use clap::Parser;
use tokio_util::sync::CancellationToken;
use tracing_subscriber::EnvFilter;
use noxa_rag::{
build_embed_provider, build_vector_store, load_config,
config::{EmbedProviderConfig, SourceConfig},
pipeline::Pipeline,
};
#[derive(Parser)]
#[command(name = "noxa-rag-daemon", about = "noxa RAG indexing daemon")]
struct Args {
/// Config file path
#[arg(long, default_value = "noxa-rag.toml")]
config: PathBuf,
/// Log level (overrides RUST_LOG)
#[arg(long, default_value = "info")]
log_level: String,
/// Print version and exit
#[arg(long)]
version: bool,
}
#[tokio::main]
async fn main() {
let args = Args::parse();
if args.version {
println!("noxa-rag-daemon {}", env!("CARGO_PKG_VERSION"));
std::process::exit(0);
}
// Init tracing to stderr (stdout may be piped).
tracing_subscriber::fmt()
.with_env_filter(
EnvFilter::try_from_default_env()
.unwrap_or_else(|_| EnvFilter::new(&args.log_level)),
)
.with_writer(std::io::stderr)
.init();
if let Err(e) = run(args).await {
eprintln!("[noxa-rag] fatal: {e}");
std::process::exit(1);
}
}
async fn run(args: Args) -> Result<(), Box<dyn std::error::Error>> {
let config_path = &args.config;
// Warn if config file is world-readable (may contain api_key).
#[cfg(unix)]
{
use std::os::unix::fs::PermissionsExt;
if let Ok(meta) = std::fs::metadata(config_path) {
let mode = meta.permissions().mode();
if mode & 0o004 != 0 {
eprintln!(
"[noxa-rag] WARNING: config file is world-readable (mode {:o}). \
Consider: chmod 600 {}",
mode,
config_path.display()
);
}
}
}
// Load config — fail fast with clear error.
let config = load_config(config_path).map_err(|e| {
format!("failed to load config from {}: {e}", config_path.display())
})?;
// Ensure watch_dir exists (create if missing — convenience for first-run).
let watch_dir = match &config.source {
SourceConfig::FsWatcher { watch_dir, .. } => watch_dir.clone(),
};
if !watch_dir.exists() {
std::fs::create_dir_all(&watch_dir).map_err(|e| {
format!(
"watch_dir does not exist and could not be created ({}): {e}",
watch_dir.display()
)
})?;
eprintln!("[noxa-rag] created watch_dir: {}", watch_dir.display());
}
// Build embed provider — startup probe (exits 1 if TEI unavailable).
let embed = build_embed_provider(&config)
.await
.map_err(|e| format!("embed provider startup failed: {e}"))?;
// Probe embed dims dynamically by calling embed with a dummy text.
// This avoids hardcoding and reuses the already-built provider.
let probe_vecs = embed
.embed(&["probe".to_string()])
.await
.map_err(|e| format!("embed dims probe failed: {e}"))?;
let embed_dims = probe_vecs
.first()
.map(|v| v.len())
.ok_or("embed probe returned empty vector")?;
// Build vector store — collection create/validate.
let store = build_vector_store(&config, embed_dims)
.await
.map_err(|e| format!("vector store startup failed: {e}"))?;
// Load tokenizer.
let tokenizer_model = match &config.embed_provider {
EmbedProviderConfig::Tei { model, local_path, .. } => {
(model.clone(), local_path.clone())
}
_ => ("Qwen/Qwen3-Embedding-0.6B".to_string(), None),
};
// Rust tokenizers crate has no from_pretrained — local_path is required.
// Download tokenizer.json from HF Hub before running:
// huggingface-cli download Qwen/Qwen3-Embedding-0.6B tokenizer.json --local-dir ./
let tokenizer = {
let path = tokenizer_model.1.ok_or_else(|| {
format!(
"embed_provider.local_path is required — the Rust tokenizers crate cannot \
download from HF Hub. Set local_path to the directory containing tokenizer.json.\n\
Download: huggingface-cli download {} tokenizer.json --local-dir <dir>",
tokenizer_model.0
)
})?;
// If given a directory, look for tokenizer.json inside it.
let tokenizer_file = if path.is_dir() {
path.join("tokenizer.json")
} else {
path.clone()
};
tokenizers::Tokenizer::from_file(&tokenizer_file)
.map_err(|e| format!("failed to load tokenizer from {}: {e}", tokenizer_file.display()))?
};
eprintln!("[noxa-rag] tokenizer: {} — loaded", tokenizer_model.0);
let shutdown = CancellationToken::new();
let pipeline = Pipeline::new(
config,
embed,
store,
Arc::new(tokenizer),
shutdown.clone(),
);
// Signal handling: Ctrl-C + SIGTERM -> cancel.
let shutdown_signal = shutdown.clone();
tokio::spawn(async move {
#[cfg(unix)]
{
use tokio::signal::unix::{signal, SignalKind};
let mut sigterm = signal(SignalKind::terminate())
.expect("failed to register SIGTERM handler");
tokio::select! {
_ = tokio::signal::ctrl_c() => {}
_ = sigterm.recv() => {}
}
}
#[cfg(not(unix))]
{
let _ = tokio::signal::ctrl_c().await;
}
eprintln!("[noxa-rag] shutdown signal received");
shutdown_signal.cancel();
});
eprintln!("[noxa-rag] daemon started");
// Run pipeline with 10s force-exit timeout after cancellation.
match tokio::time::timeout(Duration::from_secs(10), pipeline.run()).await {
Ok(Ok(())) => {}
Ok(Err(e)) => eprintln!("[noxa-rag] pipeline error: {e}"),
Err(_elapsed) => {
eprintln!("[noxa-rag] WARNING: pipeline did not shut down within 10s, forcing exit");
std::process::exit(0);
}
}
eprintln!("[noxa-rag] daemon stopped");
Ok(())
}