Hybrid search is a retrieval method that runs two kinds of search on the same query, one that matches exact words and one that matches meaning, and then merges the results into a single ranked list. It exists because each method misses things the other catches. Many AI search systems and retrieval tools use some form of it to pick the passages a language model reads before writing an answer.
What are the two halves of hybrid search?
The first half is keyword search, also called lexical search. It looks for documents that contain the words in the query and scores them by how often those words appear and how rare they are across the whole collection. A common scoring formula for this is BM25, which gives more weight to uncommon terms and stops rewarding a word after it has appeared a few times.
The second half is vector search. A model turns the query and every stored passage into an embedding, a list of numbers that encodes meaning. The system returns the passages whose embeddings sit closest to the query's. This is the mechanism behind semantic search, and it can match "how to stop a dripping tap" with a passage about fixing a leaking faucet even though they share almost no words.
| Keyword search | Vector search | |
|---|---|---|
| Matches on | Exact words and their frequency | Closeness of meaning |
| Strong with | Product names, model numbers, error codes, rare terms | Paraphrases, synonyms, vague or conversational questions |
| Weak with | Different wording for the same idea | Exact identifiers, new or uncommon terms the embedding model handles poorly |
| Index type | Inverted index of words | Vector index of embeddings |
Why combine keyword and vector search?
Each method fails in predictable ways. Vector search can return a passage about a similar product when the user typed one specific model number, because the embeddings for two related products sit close together. A keyword search, on the other hand, can miss a perfectly relevant passage because it uses "cost" where the query says "price".
Running both and merging the lists covers more of these cases. A passage that ranks well in both lists is usually a strong match, and a passage that ranks well in only one still has a chance to appear. In practice, hybrid retrieval tends to return a better candidate set than either method alone. That matters because everything later in the pipeline can only work with what retrieval returned.
How does hybrid search merge the results?
The two searches produce scores on different scales. A BM25 score can be any positive number, while vector similarity usually falls in a narrow range. The system needs a way to combine them, and two approaches are common.
Reciprocal rank fusion
Reciprocal rank fusion ignores the raw scores and uses only positions. Each passage gets a value of one divided by a constant plus its rank in each list, and the values from both lists are added together. Depending on the constant, a passage ranked third in one list and fifth in the other ends up above a passage ranked first in one list and absent from the other. Because the method is simple and needs no tuning of score scales, it is widely used.
Weighted score combination
The other approach normalizes both sets of scores to the same range, then adds them with weights. A system might give more weight to keyword scores for a shopping search full of product codes and more weight to vector scores for open questions. This gives finer control, though the weights have to be chosen through testing.
Some systems use a third form called sparse plus dense retrieval, where a learned model produces a sparse vector of weighted terms alongside the dense embedding. It behaves like keyword search with some understanding of related words, and its results are merged with the dense results in the same way.
Where does hybrid search fit in an AI answer?
Hybrid search is a retrieval step. In retrieval-augmented generation, the system first retrieves candidate passages, then passes them to a model that writes the answer. Hybrid search usually produces the first candidate list, often a few dozen or a few hundred passages.
After fusion, many systems run a reranker. This is a model that reads the query and each candidate passage together and gives a more precise relevance score. It is slower than the first retrieval step, so it only runs on the short list. The top passages after reranking are the ones placed in the language model's input. How AI search chooses which sources to cite describes the later stages in more detail.
The exact setup varies by product and is rarely published. Some AI search tools query a web search engine that already blends many signals, then apply their own passage ranking, while others maintain their own index. Across them, the general pattern of combining lexical and semantic matching and then reranking is common.
What does hybrid search mean for a website?
Because hybrid retrieval looks at both words and meaning, a page benefits from serving both halves. The keyword side rewards the literal terms people type, such as product names, brand names, model numbers, standard industry terms and the phrasing customers use. If a page describes a product only in internal jargon, the keyword half may not find it for the obvious query.
On the vector side, what counts is whether a passage's meaning is clear on its own. A paragraph that states what something is, who it is for or how it works produces an embedding close to the questions it answers. A passage that depends on the paragraphs above it to make sense produces a vaguer embedding and loses to clearer text elsewhere.
In practice, this means:
- Use the exact names of products, features and categories in the text where they are described, in addition to descriptive language.
- Write sections that answer one question each, with the answer near the start.
- Keep tables, specifications and definitions in plain text so both indexes can read them.
Retrieval decides which passages a model can quote at all, so changes in wording can show up as changes in how often a brand is cited. You can only see those shifts by running the same prompts repeatedly over time and recording which sources appear, which is what AI visibility measurement does.
