turbo-graph
A fork of the turbovec vector search library that adds a graph-memory layer, letting a single query filter retrieval results by tenant, tag, source and time alongside vector similarity.
turbo-graph extends turbovec's TurboQuant vector search core with graph memory, tag, source and time indexed views, cached SlotMask compilation, graph-aware reranking, and explain and cache telemetry, exposed to Python through a GraphMemoryIndex class. It targets retrieval queries that combine several constraints at once, such as tenant, graph neighborhood, tag, source, time and BM25 candidates, instead of rebuilding that filtering logic per request.
It keeps the same TurboQuant encoding and search core, file formats, and core Python index APIs as turbovec, adding the graph layer as a new public surface on top. A bundled example replay script and an HTML panel let a developer inspect the constrained graph view, stale or duplicate candidate ids, and cache telemetry for a query. It is written in Rust.