hybrid-retrieval

A local daemon plus a Claude Code UserPromptSubmit hook that retrieves the files most relevant to your prompt and injects them into context before the model sees it.

Dense (sqlite-vec) + sparse (FTS5 BM25) + AST import expansion, fused with RRF, reranked by a LightGBM model trained on your repo's own git history. Everything runs locally.

Design: docs/SPEC.md
How it is proven to save tokens and time: docs/BENCHMARK.md

Status

Pre-alpha. See the milestone table in docs/SPEC.md §10.

Try it

uv run hybrid-retrieval doctor            # check the environment
uv run hybrid-retrieval index  /path/repo # build the sparse index (seconds)
uv run hybrid-retrieval embed  /path/repo # build embeddings (slow, GPU-bound)
uv run hybrid-retrieval query  --repo /path/repo "add jitter to retry backoff"
uv run hybrid-retrieval serve             # resident daemon
uv run hybrid-retrieval dashboard         # web UI on http://127.0.0.1:5111

Dashboard

hybrid-retrieval dashboard serves four tabs:

Overview — index contents, embedder, and the reranker's held-out metrics.
Pipeline — type a prompt and watch it move through every stage. The request goes through the daemon, so what you see is exactly what the hook does, including the skip rule and the trace it writes. Each stage card shows whether it ran, how long it took, and what it produced.
Traces — every recent request, including skipped ones, with the gate decision and token cost.
Config — every setting, with defaults shown and only changed values written to disk. Saves to the repo's .retrieval/config.toml or globally.

It binds to localhost and has no authentication: it exposes repo contents and writes config files, so do not put it on a network interface. Restart the daemon after saving config, since it holds its settings in memory.

Installing the hook

Add to .claude/settings.json in any repo you want it active for:

{
  "hooks": {
    "UserPromptSubmit": [
      { "hooks": [{ "type": "command", "command": "python3 /ABS/PATH/TO/hook/shim.py" }] }
    ]
  }
}

Invoke it through python3 rather than relying on the file's executable bit. A hook that cannot exec fails silently: you get no context and no error, which is indistinguishable from the hook simply deciding to skip. Confirm it is live with hybrid-retrieval trace — every request writes a row, including skipped ones.

The shim autostarts the daemon on first use and fails open: if anything is wrong it injects nothing and exits 0, so a broken index can never break a session.

Development

uv sync --all-groups     # install, including dev tools
uv run pytest            # tests
uv run ruff check .      # lint

AIConductor