Hybrid And Sparse Vectors

Agent upserts a sparse vector as {indices:[12, 901, 4503], values:[0.7, 0.3, 0.9]} under the declared sparse vector name.

Provide sparse vectors as parallel indices[] and values[] arrays of equal length under the sparse vector name declared in the collection's sparse_vectors config. indices are token/feature ids; values are weights. Confirm the collection declares the sparse vector before upserting.

Agent retrieves 500 candidates with a cheap matryoshka/low-dim vector in prefetch, then reranks the top set with a full-dimension vector in the outer query.

Use a wide cheap-vector prefetch (high limit) feeding a precise full-dimension rerank in the outer query — classic retrieve-wide-then-rerank. Tune the prefetch limit (candidate pool) vs final limit so recall is preserved while rerank cost stays bounded.

Agent does hybrid search: a prefetch[] pulls 100 candidates from the dense vector and another from the sparse vector, then fuses them in the outer query.

Express multi-stage retrieval with prefetch[]: each prefetch is a sub-query (dense or sparse, with its own using and limit) that produces candidates the outer query fuses or re-ranks. Set each prefetch limit wide enough to feed fusion. The prefetch limit governs candidate recall before fusion.