archon-search

A hybrid search server for your own machine — vector + full-text + reranking + routing across collections, with no external service to trust with your data by default.

• PyPI: https://pypi.org/project/archon-search/
• GitHub: https://github.com/user538295/archon-search

Point it at a directory, ask it a question, get back ranked chunks with source paths:

curl -X POST http://127.0.0.1:8765/search \
  -H "Authorization: Bearer $ARCHON_SEARCH_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"collection": "docs", "query": "how does the router work?"}'

{
  "results": [
    {
      "doc_id": "docs/routing.md",
      "chunk_id": "docs/routing.md#3",
      "text": "The router scores each collection's centroid against the query embedding...",
      "score": 0.84,
      "source_path": "/Users/you/project/docs/routing.md",
      "collection": "docs"
    }
  ],
  "acl_filtered": false
}

No cloud call happened to produce that by default — the index, the embeddings, and the reranker all run on the box you started archon-search on. (HyDE and RAG Fusion, both opt-in, are the exception: they call out to an LLM provider — see Features.)

What it's made of

• LanceDB as the local vector store — an embedded file format, not a service you have to run alongside it.
• fastembed for dense embeddings — no GPU required to get started.
• A cross-encoder reranker — first-stage recall from vector + FTS, second-stage precision from the reranker, RRF-fused.
• A multi-collection router — one query, many collections; the router scores each collection's centroid and picks which ones actually get searched instead of fanning out to all of them.
• FastAPI for the REST control plane, contracted via OpenAPI 3.x — GET /openapi.json is the source of truth, not this file.
• An MCP endpoint — the same tools your HTTP client uses, exposed to MCP clients (Claude Code, etc.) over the same auth.

It's one process. It persists everything under ~/.archon-search/. There's no separate vector DB to stand up, no message queue, no second auth system for the MCP side.

Features

Search that gets smarter about the query, not just the index:

• Vague or short queries under-retrieve with plain vector search. hyde=true on /search generates a hypothetical answer and embeds that instead of the raw query — closes the gap between "what you typed" and "what the answer looks like." (archon-search[hyde], needs an LLM provider: Anthropic — the default, included in the hyde extra — or Ollama (archon-search[ollama], plus a running Ollama server), OpenAI (archon-search[openai-provider]), claude_cli — uses Claude Code's own login, no extra install and no API key — or llama_cpp — a local llama-server instance, no extra install and no API key; use a small, direct-response instruct model, not a reasoning model, or the entire token budget goes to hidden chain-of-thought and HyDE/RAG Fusion silently stay disabled)
• One query rarely covers what a broad question needs. rag_fusion=true decomposes the query into sub-queries, searches all of them in parallel, and fuses the results with a second RRF pass. Mutually exclusive with HyDE — pick the one that fits the query shape. (archon-search[rag_fusion], same provider options and extras as HyDE above)
• Non-English corpora get silently penalized by English-tuned defaults. Per-chunk language is detected with fastText at ingest time, and filters.language lets you scope a query to it.
• Different collections need different embedding models (code vs. prose, different languages) — active_embedding_model is set per collection and enforced through ingest, search, and sync; a mismatch raises loud, not silent.

Retrieval that understands what it's indexing, not just text blobs:

• Code search on line-blind chunks returns half a function. Code files get tree-sitter-aware chunking that keeps function/class boundaries intact instead of cutting on line count.
• A hit with no location context means opening the file to find your answer. Headings, section paths, and page numbers get extracted and attached to each chunk at ingest.
• "Search everything" isn't the same as "search the right thing." filters on /search narrow by file type, source-path prefix or glob, and indexed_after/indexed_before — before the query ever reaches the reranker.
• A ranked list without a reason is a black box you can't debug. /explain returns the full pipeline trace — vector score, FTS score, rerank score, routing decision — per candidate, with per-stage timings.

Operations that don't need someone awake at 3am:

• API keys that never rotate are a liability, and rotating them without downtime is normally its own project. POST /keys/rotate swaps in a new key while the old one stays valid through a configurable grace period.
• A crashed backup script means you find out about data loss when you need the backup. A background loop exports and rotates backups on a schedule — no cron job to babysit.
• Schema changes usually mean a forced re-ingest. Migrations run through a versioned MigrationSpec registry with documented rollback rules — existing collections migrate in place.
• FTS index rebuilds that scale with corpus size make maintenance windows scale with corpus size too. optimize_fts() is incremental — it updates what changed, not what already existed.
• "Works on my machine" isn't a deployment story. CPU and NVIDIA GPU images are published to GHCR; the tiered install wizard (minimal / balanced / max) picks embedding + reranker models that fit your disk budget instead of downloading everything.

Installation

# pip
pip install archon-search
archon-search wizard

# uv (installs the CLI into an isolated managed environment)
uv tool install archon-search
archon-search wizard

archon-search wizard does the rest: pick a profile (minimal, balanced, or max), it pulls the matching embedding and reranker models and registers the server as a background service. Full profile comparison, flags, and disk-space requirements: Documentation/UserManual/10_installation.md.

Prefer a checkout over a package? Clone and sync:

git clone https://github.com/user538295/archon-search.git
cd archon-search
uv sync --dev

Uninstall

Stop and unregister the service first, while the CLI is still on disk:

archon-search uninstall

Add --delete-db if you also want the search database gone — that's irreversible, it removes every indexed chunk:

archon-search uninstall --delete-db

Then remove the package itself:

# pip
pip uninstall archon-search

# uv tool
uv tool uninstall archon-search

# checkout / dev install — delete the cloned directory

Neither step touches your data. uninstall only stops and unregisters the OS service; removing the package only removes the CLI binary. If you want a clean wipe, these paths are still on disk and need manual deletion:

Or skip the table and remove all of it in one shot:

rm -rf ~/.archon-search/

That does not remove the embedding/reranker model weights fastembed downloads — those are cached in fastembed's own default cache (outside ~/.archon-search/), not under this directory. Clear that separately for a truly complete wipe.

Quick start

Start the server in the foreground:

archon-search serve

That runs the FastAPI app on the configured host/port — serve defaults to 0.0.0.0:8765 (all interfaces, not just loopback), unlike the library default of 127.0.0.1:8765 used when embedding SearchConfig directly. Binding all interfaces means the server is reachable from other machines on the network the moment it starts; put a reverse proxy in front of it (see Docker) before exposing it past loopback. It blocks until you stop it (Ctrl-C). For a background service managed by launchd/systemd, use archon-search wizard instead — see Installation.

Once it's up:

• GET /health — unauthenticated liveness probe
• GET /ready — unauthenticated readiness probe
• GET /docs — interactive Swagger UI
• GET /openapi.json — machine-readable OpenAPI schema

Then hit /search as shown at the top of this file.

Running with Docker

Two images, CPU (:latest) and NVIDIA GPU (:gpu). Both run the foreground archon-search serve subcommand, bind to 0.0.0.0:8765, persist everything under /data, and write logs to stderr so docker logs actually shows something.

Kick the tires — ephemeral, the key regenerates on every start, nothing persists:

docker run --rm -p 8765:8765 ghcr.io/user538295/archon-search:latest

Run it for real — pin the key, pin the volume:

docker run -d \
  --name archon-search \
  -e ARCHON_SEARCH_API_KEY=$ARCHON_SEARCH_API_KEY \
  -v archon-search-data:/data \
  -p 8765:8765 \
  ghcr.io/user538295/archon-search:latest

Skip the volume or the env var and every restart mints a new key — every token you handed out stops working, silently. Mount a volume so the key persists at /data/.search.env, or pass ARCHON_SEARCH_API_KEY explicitly. Pick one; don't rely on neither.

Want a dev/test/prod stack with isolated volumes instead of one container? docker-compose.yml and .env.example have it. Full operator guide — compose stack, image variants, env-var reference, persistence layout: Documentation/UserManual/140_running_with_docker.md.

LanceDB is single-writer. Mount the same data volume into two running containers and the on-disk state is undefined — don't do that.

The container speaks plaintext HTTP, nothing else. Put a reverse proxy (nginx, Caddy, Traefik) in front of it before you expose it past loopback.

Authentication

Every endpoint requires a Bearer token in the Authorization header, except a small unauthenticated set: GET /health, GET /ready, GET /docs, GET /openapi.json, GET /redoc. One route has a second, narrower exemption — GET /graph/{collection}/view (the HTML graph viewer) also accepts the token as a ?token= query parameter instead of the header, since it's meant to be opened directly in a browser; the query param is still validated against the same key set (archon_search/server/middleware_auth.py).

First start auto-generates a key and writes it to ~/.archon-search/.search.env at 600. Running in Docker, CI, or across multiple hosts? Set ARCHON_SEARCH_API_KEY — it wins over the file every time. Need the key read from somewhere else? Set ARCHON_SEARCH_KEY_FILE. Need the whole runtime tree — index, logs, key file, jobs file, the fastText language-ID model cache, ingest history — under one root instead of ~/.archon-search/? Set ARCHON_SEARCH_DATA_DIR (the Docker image already does this for you, pointing it at /data). This does not relocate fastembed's embedding/reranker weight cache, which fastembed manages in its own default cache outside this tree — see Uninstall.

Configuration

Everything server-side lives in one file: ~/.archon-search/archon-search.toml.

• [database] — db_path, embedding_model, chunk_size, top_k_return, model paths, per-collection embedder pool sizing (embedder_cache_size, default 3; eager_load_embedders, default false, pre-warms embedding_model plus every distinct per-collection model and the reranker cross-encoder at startup instead of lazily)
• [search] — multi-collection fan-out bounds (max_fanout, fanout_timeout_seconds)
• [routing] — routing_shortlist_size, routing_confidence_threshold, routing strategy
• [collections] — pinned_collections, static collection definitions, watcher settings
• [telemetry] — opt-in local query logging, covered below

The full annotated reference — every key, every default — is archon-search.toml.example. This section is the map, not the territory.

REST API

GET /openapi.json is the contract — endpoint shapes, request/response types, error codes, all of it. This README is not. GET /docs serves the same thing as an interactive explorer.

Breaking changes to REST or MCP land in BREAKING.md, not buried in a changelog entry.

MCP tools

Your MCP client gets the same server your HTTP client does — same auth, same data, no separate integration to build. 16 tools always register; 4 more join when a key store is configured, 20 total (archon_search/server/mcp.py):

• search — hybrid vector + FTS search; returns {"results": [...], "acl_filtered": bool}
• search_with_context — same as search with adjacent-chunk context
• explain — per-stage retrieval/reranking trace plus routing decision (mirrors POST /explain)
• ingest_file — index a single file into a collection
• ingest_directory — recursively index a directory
• list_collections — list collection names
• get_collections_meta — metadata for all collections
• get_collection_meta — metadata for one collection
• list_documents — list documents in a collection
• delete_document — remove a document by doc_id
• update_collection — change a collection's embedding model (mirrors PATCH /collections/{name})
• export_collection / import_collection — archive a collection out or restore it
• get_graph — entity-graph summary for one collection (nodes, edges, top entities by salience)
• get_graph_cross_collection — merged entity graph across 2+ collections
• graph_impact — blast-radius caller/callee analysis for a code symbol
• create_key / list_keys / revoke_key / rotate_key — API key lifecycle, mirroring the REST /keys endpoints (registered only when key management is configured)

Telemetry (opt-in)

Off by default (enabled = false), and it stays off until you flip it. Flip it, and every search, search_with_context, search_multi, POST /route, and /explain call appends one JSONL line to a daily file under ~/.archon-search/search-logs/. It never leaves the machine — no export, no phone-home, no exceptions.

Enabling

# ~/.archon-search/archon-search.toml
[telemetry]
enabled = true
retention_days = 30          # files older than this are deleted at startup and every 24h
log_dir = "~/.archon-search/search-logs"
hash_doc_ids = false         # set true to HMAC-SHA256 result_doc_ids before writing to JSONL

What is logged

query_id (random UUID), timestamp (UTC), endpoint, latency_ms, status, plus whatever's specific to the call — collection/result_count/result_doc_ids for retrieval, collections/decomposer_invoked for routing. Errors add one more field, error_kind, from a closed set: empty_query | slot_out_of_range | timeout | internal_error | validation_error | other.

What is never logged

The raw query string, never. Not a policy — a structural guarantee: the factory methods that build telemetry entries have no query parameter to pass one in. Exception messages don't make it in either; only the coarse error_kind string does.

The catch: doc_ids leak paths

doc_ids may reveal filesystem paths: result_doc_ids comes straight from the file path on disk — /Users/<name>/Documents/<project>/<file>.md. Turn telemetry on and those paths, username included, sit in your log files. Set hash_doc_ids = true and every doc_id gets HMAC-SHA256'd before it's written, opaque to anyone without the salt file at ~/.archon-search/.telemetry-salt. Decide if that's worth doing before you turn telemetry on, not after.

export_enabled does nothing yet

It's reserved for a future release. Set it to true today and the config loader logs a warning and quietly coerces it back to false (archon_search/config.py). Nothing gets transmitted in v1 — there's no code path that would.

Telemetry read-back API

Both endpoints return {"enabled": false} when telemetry is disabled.

GET /telemetry/stats

Aggregated query statistics over an optional time window.

Response shape summary:

{
  "schema_version": 1,
  "enabled": true,
  "total_queries": 42,
  "success_rate": 0.95,
  "latency_ms": {"p50": 120, "p95": 380},
  "by_endpoint": {"search": 30, "route": 12},
  "by_collection": {"docs": 25, "code": 17},
  "error_breakdown": {"timeout": 2, "internal_error": 0}
}

success_rate is null when no queries exist in the window.

GET /telemetry/entries

Paginated raw log entries.

Response includes entries, next_offset, and total_in_window. Clients should continue calling with the returned next_offset until entries is empty (equivalently, until next_offset >= total_in_window).

Evaluation harness

"Retrieval quality didn't regress" is a claim, not a feeling — tests/eval/ is what backs it: a synthetic corpus, query/label fixtures, deterministic eval backends, committed thresholds, a measured baseline. It's the sanctioned gate for any change touching retrieval, reranking, routing, or latency.

uv run pytest -m eval --thresholds-path tests/eval/thresholds.toml tests/eval/test_eval_suite.py

The backends are deterministic — corpus-aware but label-blind, so metrics hold steady across runs without loading real model weights. Latency p50/p95 is a regression guard, not a production SLA — it tells you if this change got slower than the last one, not what to expect in prod.

Current baseline (recall@k, MRR, nDCG@k, reranker lift, routing accuracy, latency percentiles): tests/eval/baselines/baseline.md, machine-readable twin at tests/eval/baselines/baseline.json. Changing a threshold or a fixture? Read tests/eval/README.md first — it's the maintenance guide and the waiver policy, not optional reading.

Development

git clone https://github.com/user538295/archon-search.git
cd archon-search
uv sync --dev
uv run pytest

License

MIT — see LICENSE.