llama-index-vector-stores-mnestic

A LlamaIndex vector store backed by mnestic — an embedded graph + vector + full-text database (a maintained fork of CozoDB). Retrieval is hybrid: 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 llama-index-vector-stores-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 llama_index.core import VectorStoreIndex, StorageContext, Document
from llama_index.vector_stores.mnestic import MnesticVectorStore

vector_store = MnesticVectorStore(dim=1536, engine="sqlite", path="mydocs.db")
storage_context = StorageContext.from_defaults(vector_store=vector_store)

index = VectorStoreIndex.from_documents(
    [Document(text="the cat sat on the mat")],
    storage_context=storage_context,
)
nodes = index.as_retriever(similarity_top_k=4).retrieve("feline")

dim must match your embedding model's dimension. query runs hybrid search when the query carries text (the usual index path), otherwise vector-only.

Engine pass-through: graph legs and more

Extra keyword arguments flow from the retriever to the engine's hybrid query via vector_store_kwargs. That includes graph_legs — fuse graph proximity into the ranking alongside the dense and keyword legs, in the same single call:

retriever = index.as_retriever(
    similarity_top_k=4,
    vector_store_kwargs={"graph_legs": [{
        "edge_relation": "links", "from_col": "src", "to_col": "dst",
        "seeds": ["doc-1"], "max_hops": 2, "label": "graph",
    }]},
)

MnesticRetriever(..., search_kwargs={...}) forwards the same way, and so do extra_lists, vector_k, fts_k, rrf_k, and future engine keys — no adapter release needed when the engine grows new knobs.

License

Mozilla Public License 2.0.