AI glossary · Advanced · Language models · Data & evaluation
Hybrid search
Hybrid search is a technique that runs two or more different search methods together and then merges the results into a single ranked list. It commonly combines 'keyword search,' which matches words exactly, and 'semantic search,' which finds documents with similar meaning.
About this term
- Level
- Advanced · Specialist terms that need the inner workings
- Category
- Models & generation
- Fields
- Language models, Data & evaluation
- Added
- 2026-10-07
What it is
Hybrid search runs several search methods together on a single query and merges the results from each method into one ranking. The most common combination is keyword search (lexical search) and semantic search (vector search). Keyword search is good at matching exact words and phrases, while semantic search is strong at grasping the intent behind a question. Combining the two aims for both precision and broad coverage. Other combinations are also possible, such as text and images, or location information and semantic search.
How it works
Keyword search usually uses a scoring method called BM25. BM25 scores a document by how often the words in the query appear in it and how rare those words are across the whole document collection, and it also adjusts for document length. Semantic search turns text into embeddings, which are lists of numbers, and then finds documents with similar meaning. The keyword-side representation, in which most values are 0, is sometimes called a 'sparse vector', and the semantic-side representation, in which most values are filled in, a 'dense vector'. The results of the two methods are merged by a dedicated 'fusion' method into a single ranked list.
How it developed
BM25 is a long-established keyword scoring method based on probabilistic retrieval theory developed in the 1970s and 80s, and it served as the default for search for a long time. In 2020, the paper 'Dense Passage Retrieval for Open-Domain Question Answering' showed that a simple setup that turns questions and passages into separate embeddings could find relevant passages better than a strong BM25 system, establishing that semantic search was practical. Then in 2021, the BEIR paper built a benchmark comparing many retrieval methods across 18 diverse datasets. It showed that in unfamiliar domains BM25 is still a robust baseline, while dense and sparse retrieval models often generalize poorly. Results like these, showing that neither side always wins, strengthened the case for hybrid search, which uses both methods together.
Why it matters
Hybrid search lets each retrieval method cover the other's weaknesses. Keyword search is fast, its results are easy to explain, and it handles rare words such as product codes and technical terms well, but it easily misses matches when the wording changes. Semantic search understands the meaning even when a question is phrased differently, but it can be weak when exact word matches matter. It has become more important for services such as AI agents and RAG, which need to find highly relevant material for a single question, because retrieval quality feeds directly into answer quality.
Things to know
The two retrieval methods score results on different scales, so the way the results are merged and the weight given to each method must be tuned to the data. Because retrieval runs twice, the system becomes more complex and costs can rise. When higher accuracy is needed, the merged results are sometimes re-sorted once more with a reranker; the BEIR study also pointed out that this kind of reranking performs well on average but is computationally expensive.
Example
Search engines and vector DBs such as Elasticsearch and Weaviate offer hybrid search as a feature. For example, Weaviate describes its approach as combining BM25 (or BM25F, which can weight each document field differently) with vector search. In a chatbot that searches internal company documents to produce answers, keyword search helps when a user asks for an exact document number, and semantic search helps when the question is vague, like 'how to take vacation leave', and the results are delivered as a single list.