langchain-mnestic

A LangChain VectorStore backed by mnestic — an embedded graph + vector + full-text database (a maintained fork of CozoDB). Retrieval is hybrid by default: dense (HNSW) + keyword (full-text) fused with Reciprocal Rank Fusion, in one call.

mnestic is a maintained fork of CozoDB; it is not the official CozoDB. Original design credit belongs to Ziyang Hu and the Cozo Project Authors.

pip install langchain-mnestic

Use 0.2.1 or newer. It validates relation, column, index, distance, and dtype identifiers and coerces numeric HNSW settings before constructing CozoScript; earlier releases interpolated those configuration values directly.

from langchain_mnestic import MnesticVectorStore
from langchain_openai import OpenAIEmbeddings

store = MnesticVectorStore.from_texts(
    ["the cat sat on the mat", "a dog ran in the park"],
    embedding=OpenAIEmbeddings(),
    metadatas=[{"src": "a"}, {"src": "b"}],
    engine="sqlite", path="mydocs.db",   # or engine="mem" for ephemeral
)

docs = store.similarity_search("feline", k=2)
retriever = store.as_retriever(search_kwargs={"k": 4})

Scores returned by similarity_search_with_score are RRF fused scores — higher is better (a relevance score, not a distance).

Engine pass-through: graph legs and more

Any extra keyword argument on the search methods goes straight to the engine's hybrid query. That includes graph_legs — fuse graph proximity into the ranking alongside the dense and keyword legs, in the same single call:

db = store._store.db  # or pass your own CozoDbPy via MnesticVectorStore(db=...)
db.run_script(":create links {src: String, dst: String}", {}, False)
db.run_script("?[src, dst] <- [['doc-1', 'doc-9']] :put links {src, dst}", {}, False)

docs = store.similarity_search(
    "feline", k=4,
    graph_legs=[{
        "edge_relation": "links", "from_col": "src", "to_col": "dst",
        "seeds": ["doc-1"], "max_hops": 2, "label": "graph",
    }],
)

extra_lists, vector_k, fts_k, rrf_k, and future engine keys pass through the same way — no adapter release needed when the engine grows new knobs.

Mem0

The store works as a Mem0 vector-store backend through Mem0's provider="langchain" (it implements the by-vector search, scored search, get_by_ids, add_embeddings, and scalar metadata filter calls Mem0's wrapper makes — exercised in tests/test_mem0.py):

config = {"vector_store": {"provider": "langchain",
                           "config": {"client": store, "collection_name": "mem0"}}}

License

Mozilla Public License 2.0.