project · work in progress
HybridMind, local hybrid search over vectors, keywords, and a graph
hybridmind is a retrieval engine that runs on your machine and ranks with three signals at once. it is still a work in progress, and i use it for local memory experiments. i'm Akshat Kushwaha, an ai systems engineer in india.
what it combines
- dense: FAISS HNSW vector search over 4096-dim vectors.
- lexical: Okapi BM25 keyword search, with bm25s and PyStemmer.
- graph: typed proximity over a NetworkX directed multigraph.
- storage: SQLite in WAL mode.
how results are ranked
each signal produces its own ranked list. Reciprocal Rank Fusion (RRF) at k=60 merges the lists using positions rather than raw scores, so no score has to be calibrated against another. an optional cross-encoder can rerank the fused result.
what is in the box
snapshots are .mind.zip files with SHA256 manifests, and there is a Python SDK.
more writing on this: why we need hybridmind.