fathomdb 0.4.5

Local datastore for persistent AI agents — graph, vector, and full-text search on SQLite

fathomdb

Local datastore for persistent AI agents — graph, vector, and full-text search on SQLite.

Installation

pip install fathomdb

Quick Start

from fathomdb import Engine, WriteRequestBuilder

# Opt into the Phase 12.5 read-time embedder so search() fires its
# vector branch on natural-language queries. The embedder="builtin"
# shape requires a fathomdb build with --features default-embedder;
# when that feature is off it silently falls back to text-only search.
engine = Engine.open("my_agent.db", embedder="builtin", vector_dimension=384)

# Write data
builder = WriteRequestBuilder("ingest")
node = builder.add_node(
    kind="memory",
    properties={"text": "hello world"},
    source_ref="quickstart",
)
engine.write(builder.build())

# Unified search — the recommended retrieval entry point.
rows = engine.nodes("memory").search("hello", 10).execute()
for hit in rows.hits:
    print(hit.node.logical_id, hit.score, hit.modality.value, hit.snippet)

engine.close()

Features

• Graph backbone with nodes, edges, and temporal tracking
• Unified search() entry point -- one call runs a strict-then-relaxed text pipeline plus an optional vector branch and returns ranked SearchHit rows over both document chunks and structured property projections
• Read-time query embedder (Phase 12.5): opt in with Engine.open(path, embedder="builtin", vector_dimension=384) to let search() fire its vector branch on natural-language queries. See the querying guide for the full configuration surface and degradation semantics. The "builtin" embedder requires fathomdb to be built with the default-embedder Cargo feature; when the feature is off, the engine logs a warning and silently falls back to text-only search
• Vector similarity search via sqlite-vec (advanced override for callers that want to supply a vector literal directly)
• FTS property schema management -- register JSON property paths per node kind, including recursive-mode paths that populate a sidecar position map and unlock per-hit match attribution
• Provenance tracking on every write
• Single-writer / multi-reader with WAL

Unified search

from fathomdb import Engine, FtsPropertyPathMode, FtsPropertyPathSpec

engine = Engine.open("/tmp/fathom.db")

# search() is the primary retrieval entry point. The engine owns the
# strict-then-relaxed policy and returns SearchRows, not QueryRows.
rows = engine.nodes("Goal").search("ship quarterly docs", 10).execute()
for hit in rows.hits:
    print(hit.node.logical_id, hit.score, hit.modality.value,
          hit.source.value, hit.snippet)
print(rows.strict_hit_count, rows.relaxed_hit_count, rows.vector_hit_count)

# Recursive property FTS + opt-in match attribution.
engine.admin.register_fts_property_schema_with_entries(
    "KnowledgeItem",
    entries=[
        FtsPropertyPathSpec(path="$.payload", mode=FtsPropertyPathMode.RECURSIVE),
    ],
)
attributed = (
    engine.nodes("KnowledgeItem")
    .search("quarterly docs", 10)
    .with_match_attribution()
    .execute()
)
for hit in attributed.hits:
    if hit.attribution:
        print(hit.attribution.matched_paths)

# Advanced overrides (pin modality or supply both shapes verbatim):
#   engine.nodes("Goal").text_search("ship quarterly docs", 10).execute()
#   engine.fallback_search("quarterly docs", "quarterly OR docs", 10).execute()
# See docs/guides/querying.md for when each is the right tool.

For the full retrieval pipeline, supported query grammar, and property-FTS schema registration semantics, see the guides under docs/guides/.

Documentation

See the GitHub repository for full documentation.

License

MIT