Add LodeDB vector database adapter

[Davidobot]

Summary

Adds cognee-community-vector-adapter-lodedb, registering LodeDB as a cognee vector_db_provider. LodeDB is a local-first, on-disk, in-process vector store (no server, no account, no API key): an exact brute-force scan by default, optional IVF-style ANN for large corpora, hybrid (vector + BM25) retrieval, and O(changed) delta persistence for fast incremental writes.

How it works

cognee owns embedding (its EmbeddingEngine), so the adapter uses LodeDB's vector-in path:

• One LodeDB index per cognee collection under vector_db_url (a local directory, created if absent).
• The serialized DataPoint payload is kept in LodeDB's raw-text sidecar, so retrieve and include_payload searches return it.
• belongs_to_set membership is stored as scalar metadata presence keys, so node_name (NodeSet) filtering pushes into LodeDB's metadata planner rather than post-filtering in Python.
• cognee ranks by cosine distance (lower is better), so results report 1 - cosine_similarity.

Implements the full VectorDBInterface (has_collection, create_collection, create_data_points, search, batch_search, retrieve, delete_data_points, create_vector_index, index_data_points, prune, embed_data) plus remove_belongs_to_set_tags and upsert_raw_vectors.

The adapter implementation ships in the lodedb package (lodedb.local.integrations.cognee, added in lodedb 1.3.1) where it is maintained and tested against LodeDB's API; this package registers it via use_vector_adapter("lodedb", ...).

Usage

import cognee
from cognee_community_vector_adapter_lodedb import register  # noqa: F401

cognee.config.set_vector_db_config(
    {"vector_db_provider": "lodedb", "vector_db_url": "./.lodedb"}
)

Or via env (.env.example): VECTOR_DB_PROVIDER=lodedb, VECTOR_DB_URL=./.lodedb.

Contents

• cognee_community_vector_adapter_lodedb/ — register module + adapter re-export
• pyproject.toml (lodedb>=1.3.1, cognee>=1.1.0,<2), README.md, example.py, .env.example
• tests/test_lodedb.py — offline tests (deterministic stub embedding, no LLM or network): provider registration, create/search/retrieve roundtrip, and node_name (belongs_to_set) filtering.

Notes

I held off on adding a test_lodedb.yml workflow and a committed lockfile, since CI wiring depends on the org's LLM/embedding secrets and your preferred lock tooling. Happy to add a per-package workflow following the milvus/qdrant pattern (the included tests are offline and need no secrets) and a poetry.lock if you'd like.