Blog

Hybrid search without a second index

20 Jan 2026

Sparse indexes are the right tool for some lexical models. They are overhead when all you need is BM25-style weight on text you already keep in the payload.

1.4 scores those terms on the candidate set the dense walk already touches, then fuses. Exact-token queries stop losing to a vaguely similar paragraph.

If you truly need a learned sparse model with its own vocabulary, keep a named sparse vector. The engine still accepts it.