Expand description
Sparse (lexical) vectors for hybrid retrieval (zero dependencies). Sparse (lexical) vectors for hybrid retrieval.
A SparseVector is the learned-lexical counterpart to a dense embedding —
for example BGE-M3’s sparse head or SPLADE. It pairs vocabulary indices with
weights, and is what pgvector stores in a sparsevec column.
Combining a dense and a sparse branch is what makes hybrid retrieval work: the dense side matches paraphrases, the sparse side matches exact terms (identifiers, statute numbers, rare jargon) that dense vectors blur away.
Structs§
- Sparse
Vector - A sparse vector —
indices[i]carries weightvalues[i].