Skip to main content

Edit this page

Hybrid Search

Hybrid search combines a lexical signal (full-text @@ matching, ranked by BM25) with a vector signal (ANN distance). Because a single inverted index can cover both a text column and a vector column, the two can be combined in one query. There are two common strategies.

The examples use a catalog table indexed on a text column (name), a verbatim column (category) and a vector column (emb) in one inverted index.

Filtered ANN​

When one signal is a hard filter and the other ranks, put the filter in WHERE and the vector distance in ORDER BY. This restricts the candidate set, then ranks the survivors by similarity. The filter can be a full-text predicate:

Query
SELECT id, nameFROM catalog_idxWHERE name @@ ts_phrase('shoes')ORDER BY emb <-> [1.0, 0.0, 0.0]::FLOAT[3]LIMIT 2;
Result
 id | name----+-------------------  1 | red running shoes  3 | red trail shoes

…or a structured / verbatim predicate:

Query
SELECT id, nameFROM catalog_idxWHERE category @@ 'footwear'ORDER BY emb <-> [1.0, 0.0, 0.0]::FLOAT[3]LIMIT 2;
Result
 id | name----+-------------------  1 | red running shoes  3 | red trail shoes

Filtered ANN is the right choice when the filter is meaningful (a category, a permission, a required keyword) and you only want similarity to order the matches.

Score fusion​

When both signals should rank results — neither is a strict filter — fuse two independently-ranked result lists. How you fuse depends on whether the scores are comparable:

  • Weighted sum (α·s₁ + β·s₂) works when the scores share a scale. Lexical and vector scores usually do not: a BM25 magnitude and a cosine/L2 distance live on different scales, so a raw weighted sum is dominated by whichever scale is larger.
  • Reciprocal Rank Fusion (RRF) sidesteps the scale problem by combining ranks instead of scores: each document's contribution is 1 / (k + rank) summed across branches. This is the standard choice for lexical + vector fusion.

The query below runs a BM25 branch and a vector branch, ranks each independently and fuses them with RRF. The vector distance is ascending (nearer is better), so its branch orders by dist; the BM25 branch orders by score descending:

Query
WITH fused AS (    SELECT id, ROW_NUMBER() OVER (ORDER BY s DESC) AS rank FROM (        SELECT id, BM25(catalog_idx.tableoid) AS s        FROM catalog_idx        WHERE name @@ ts_phrase('red')        ORDER BY s DESC LIMIT 100    ) lex    UNION ALL    SELECT id, ROW_NUMBER() OVER (ORDER BY dist) AS rank FROM (        SELECT id, emb <-> [1.0, 0.0, 0.0]::FLOAT[3] AS dist        FROM catalog_idx        ORDER BY dist LIMIT 100    ) vec)SELECT idFROM fusedGROUP BY idORDER BY SUM(1.0 / (60 + rank)) DESC, idLIMIT 3;
Result
 id----  1  3  5

Each branch caps its contribution with a per-branch LIMIT (the window size), assigns ranks with ROW_NUMBER() and the outer query sums 1 / (k + rank) per id. See Reciprocal Rank Fusion for the full treatment of k, window size and tuning.

See also​