Quanta wants to replace the pile of separate databases behind most retrieval-augmented generation systems with a single API.
Quanta is an open-source Python library detailed in a research paper published September 17, 2026. It merges three retrieval methods that developers typically wire together by hand: dense vector search over 4-bit quantized embeddings, BM25 keyword search, and knowledge-graph traversal. Rather than blending results by normalizing scores from each system onto a shared scale, Quanta uses weighted reciprocal rank fusion, arguing that score normalization is query-dependent and therefore unreliable. The graph component does not score relevance itself; it only widens the candidate pool, and those newly surfaced documents get re-ranked by the same dense retrieval logic under an identifier allowlist.
Most production RAG pipelines run a vector database, a full-text search engine, and sometimes a graph database side by side, stitched together with custom integration code that gets rebuilt for every project. Quanta's bet is that this glue code is a solved problem, not something every team should reinvent from scratch. If the rank-fusion approach generalizes beyond the paper's own tests, it could remove a real chunk of plumbing work from building retrieval systems.
One library, one paper, zero production war stories yet - the real test is whether it survives contact with messy, real-world document collections.