FathomDB

Python SDK for FathomDB — a local-first, embedded retrieval engine for application and agent workloads, built on SQLite.

A FathomDB database is a single file on your disk. There is no server, no daemon and no network call: you open a path and you have hybrid full-text + vector search over your own data, in your own process.

Status: pre-1.0, beta

The 0.8.x line is under active development. The public surface can change between minor releases, and scale/stability guarantees are staged for 0.9.x and later.

Install

pip install fathomdb

Requires Python 3.10 or newer. Wheels are abi3, so one wheel per platform serves every supported interpreter. The package is a PyO3 binding to the FathomDB Rust engine and has no Python dependencies of its own.

Quick start

from fathomdb import Engine

engine = Engine.open("./app.sqlite")

engine.write([
    {
        "kind": "doc",
        "body": "the quick brown fox",
        # Provenance is MANDATORY on every canonical row: source_id is the
        # axis erase_source() erases on, so a row without one is unerasable.
        "source_id": "import-2026-07",
    }
])

result = engine.search("brown fox")
for hit in result.results:
    print(hit.id, hit.kind)

engine.close()

What is on the surface

• Engine lifecycle — Engine.open, engine.close, engine.drain, engine.counters(), engine.open_report(), and a logging-subscriber seam.
• Writes — engine.write(batch) of canonical nodes ({"kind", "body", "source_id", "logical_id"?, "state"?, "reason"?, "valid_from"?, "valid_until"?}) and edges ({"edge": {"kind", "from", "to", "source_id"}}).
• Search — engine.search(query, ...) fuses FTS5 lexical retrieval with dense vector retrieval, with optional cross-encoder reranking, JSON filters, and an opt-in explain sidecar. A soft_fallback field reports when a branch could not contribute, rather than degrading silently.
• Reads — the fathomdb.read module: read.get, read.get_many, read.list, read.collection, read.mutations, read.projections, read.crossed_boundary_since. Each takes a ReadView selecting existence state and world-time validity.
• Graph — fathomdb.graph.neighbors and fathomdb.graph.search_expand for bounded BFS expansion from search hits.
• Record lifecycle — engine.transition, engine.purge, engine.erase_source.
• Projections — engine.configure_projections, declaring which fields are filterable, searchable or rankable. The engine owns every derived index; you never maintain one by hand.
• Typed errors — a single-rooted hierarchy under fathomdb.errors.EngineError with typed payload attributes, so you branch on DatabaseLockedError or EmbedderNotConfiguredError instead of parsing message strings.

Embeddings

Vector search needs vectors. Engine.open(path, use_default_embedder=True) opts into the engine's pinned default embedder; on first use the weights are downloaded and cached under ~/.cache/fathomdb/embedders/. The default is False — a plain Engine.open performs no network access at all, and vector writes then raise EmbedderNotConfiguredError rather than quietly doing nothing.

Erasing data

Every canonical row carries a source_id, and that is deliberate: it is what makes the row erasable. engine.erase_source(source_id) removes every row written under that provenance together with its full-text, vector and secondary-index shadows, and finishes the erasure at rest. engine.purge(logical_id) does the same for a single governed record. Both are idempotent, so an interrupted erasure obligation can simply be retried.

Other SDKs

The same engine is available as a Rust crate (fathomdb) and a Node.js package (npm install fathomdb).

Building from source

From a checkout of the repository:

pip install -e src/python/

License

MIT. See the LICENSE file shipped in this distribution.

Source, issues and full documentation: https://github.com/fathomadb/fathomdb