Meta SiRA: The Retrieval System That Beats Dense Embeddings With Zero Training
SIRA is a learned sparse retrieval approach: instead of a dense embedding, documents and queries expand into weighted term sets, keeping the exact-match precision and inverted-index efficiency of lexical search while learning which terms matter.
Where this fits in Mixpeek
Dense embeddings miss exact terms — SKUs, part numbers, names — that lexical search nails, and vice versa. Learned sparse retrieval is the sparse half of Mixpeek's hybrid retrieval: run a dense semantic leg and a learned-sparse lexical leg, then fuse them. SIRA is the modern version of the sparse leg that Mixpeek combines with dense vectors via RRF so a query gets both meaning and precise term matching.
About this research
Meta open-sourced SiRA, a sparse retrieval approach that outperforms trained dense retrievers on standard benchmarks without any training run. The video walks through why sparse lexical signals remain brutally competitive, what that means for embedding-first architectures, and when you actually need both in one pipeline.
Put the research to work
Mixpeek turns video, images, audio, and documents in your object storage into searchable, timestamped results through one API — the retrieval stack these papers describe.
Search your own data, free