Hybrid search
Hybrid search concept
Hybrid search combines fulltext search and vector search into a single ranked result. Instead of choosing one signal, a hybrid query scores each document by both how well its text matches the query (lexical relevance) and how close its embedding is to the query vector (semantic similarity), and then fuses the two rankings into one.
The two signals are complementary:
- Fulltext search (BM25) is precise for exact terms, names, codes, and rare keywords, but misses documents that use different wording.
- Vector search (nearest neighbor over embeddings) captures semantic similarity and paraphrases, but can rank a topically related document above one that contains the exact term the user typed.
Fusing them yields results that are both lexically and semantically relevant. This is a common building block for search over knowledge bases and for the retrieval stage of Retrieval-Augmented Generation (RAG).
Hybrid search in YDB
In YDB, a hybrid query runs over a table that has both:
- a fulltext_relevance index over the text column, and
- a vector_kmeans_tree index over the embedding column.
The query expresses the fusion in the ORDER BY clause with the HybridRank function, which takes one scoring expression per branch — a FullTextScore for the text branch and a Knn distance or similarity for the vector branch:
$queryText = "machine learning";
$queryVector = Knn::ToBinaryStringFloat([0.1, 0.2, 0.3, 0.4]);
SELECT id, title
FROM documents
ORDER BY HybridRank(
FullTextScore(text, $queryText),
Knn::CosineDistance(embedding, $queryVector))
LIMIT 10;
Each branch runs against its own index and returns its own list of best-matching documents. YDB then merges these lists into a single ranking — in one of two ways:
- Reciprocal Rank Fusion (RRF) — the default. Each branch contributes a term based on a document's rank within that branch, so the absolute magnitudes of the scores (which are not comparable across branches) do not matter.
- Linear — a weighted sum of the (optionally min-max normalized) per-branch scores.
Per-branch weights, the RRF constant, and the candidate-pool sizes are all configurable. For the full query syntax and parameters, see Hybrid search (HybridRank).
Learn more: