Hybrid Search
A retrieval approach that combines keyword (lexical) search with vector (semantic) search and merges the two result lists into one ranking.
Definition
Hybrid search runs two retrievers on the same query. A lexical engine (typically BM25) scores documents on matching words, which is excellent for names, codes and rare terms. A vector engine compares embeddings, which is excellent for paraphrase and intent. The two ranked lists are then fused, most often with reciprocal rank fusion or a weighted score, and the merged list is returned or passed to a reranker.
The approach exists because each method fails in a different way. Keyword search misses "money back" when the document says "refund"; vector search can bury "SKU 88-1204" among loosely related product text. Running both covers each other's blind spots.
Example: a support agent searches "error E4021 on checkout". Keyword search pins the exact error code; vector search adds articles about checkout failures that never mention the code. Fusion puts the article that satisfies both at the top.
Why it matters for your business
- Noticeably better retrieval for mixed content: policies, part numbers, names and natural-language questions.
- Reduces the "the answer was in the documents but the assistant did not find it" class of RAG failure.
- Usually available in the database you already run, so it adds little infrastructure.
- Fusion weights and filters need tuning against real queries, which is where evals come in.
How Teksmobile uses it
Hybrid retrieval is our default for enterprise RAG and for customer-facing search. Our hybrid search service tunes lexical, vector and fusion settings against a query set drawn from your logs, usually on pgvector or Milvus.
Related terms
Part of the Teksmobile AI Glossary. For longer reads, see our guides.
Frequently asked questions
What is reciprocal rank fusion?
A simple way to merge ranked lists: each document scores points based on its position in every list, and the totals decide the final order. It needs no score calibration, which is why it is the common default.
Is hybrid search slower?
Two queries run instead of one, often in parallel, so latency rises modestly. For most applications the gain in relevance outweighs a few extra milliseconds.
Which databases support hybrid search?
PostgreSQL (pgvector with full-text search), Elasticsearch and OpenSearch, Milvus, Qdrant and Weaviate all support some form of it, with differences in how the fusion is done.