Concept Fable

Hybrid search

Conclusion first:Hybrid Search is a retrieval method that combines keyword/sparse retrieval (such as BM25) with vector/semantic retrieval. Keyword retrieval is good at exact matching of proper nouns, numbers, codes, and regulatory clauses; vector retrieval is good at finding content that differs literally but is similar in meaning. Enterprise RAG often needs both, then improves recall quality through merging, deduplication, weighted fusion, and reranker reordering.

Hybrid Search · Retrieval method

Read the story first. There is no rush to give a definition here; let the question grow on its own first.

Fable

The two book seekers finally learned to go out together

There were two clerks in the library. One was good at finding books by characters: as soon as the reader said the title, author or a distinctive phrase, he could quickly turn to it. The other was good at hearing intent: even if the reader spoke vaguely, he could guess what problem they wanted to solve.

At first, the curator had the character-index clerk stand duty alone. When someone asked for "Record of the South Levee Flood," he found it immediately; but when a farmer asked "what should I do about water always pooling by the field," he searched every title and couldn't find a few good books.

The curator also put the intent-listening clerk on duty. The farmer was satisfied, but the scholar was frantic. The scholar clearly wanted a certain old text with a rare place name, but the clerk recommended a water conservancy book with a similar meaning—just not that one.

The first fix was to have the two take turns receiving visitors. As a result, readers had to first guess which category their question belonged to, and if they guessed wrong they still wasted the trip. Only then did the curator realize the problem was not who was stronger, but that different clues each had their blind spots.

Later, the two archivists sat at the same table. After the reader stated a request, one searched for exact matches by wording, while the other searched for similar materials by meaning, and then the two candidate sets were combined and ranked. Rare words would not be drowned out, and vague intent would not be ignored.

Sometimes the two people's opinions conflicted, so the curator looked at the scenario: to find the original text of a legal provision, literal weight was high; to find a solution, intent weight was high; when both the original text and background were needed, the two clues were allowed to complement each other.

The curator also set up a review for the merged results. If a book hit literally and a book hit by intent appeared at the same time, the overlap was checked first; if they were completely separate, the ranking was adjusted according to the reader's purpose rather than simply added together. The curator also found that the two clerks' results could not simply be split half and half. When readers came with a number, literal hits almost decided the outcome; when readers only described a problem, similar meaning mattered more. Good merging requires understanding the context, otherwise adding the two clues together just creates new noise. Later, the library posted wrong cases to train newcomers. Which errors came from looking only at words, and which came from only guessing intent, could be told at a glance. The curator said: the more clues there are, the more you need to know which clue is more trustworthy at that moment.

The library therefore had far fewer complaints. People understood that good searching does not require betting on a single method. Exact words are like house numbers, meaning is like neighborhoods; using house numbers and neighborhoods together makes it easier to bring people to the book they truly need.

Reveal

This story is about: hybrid search

Hybrid Search is a retrieval method that combines keyword/sparse retrieval (such as BM25) with vector/semantic retrieval. Keyword retrieval is good at exact matching of proper nouns, numbers, codes, and regulatory clauses; vector retrieval is good at finding content that differs literally but is similar in meaning. Enterprise RAG often needs both, then improves recall quality through merging, deduplication, weighted fusion, and reranker reordering.

What matters is that it is not just a term, but a structural problem that repeatedly appears in real AI / Agent systems. Understanding it lets you judge when to add a model, when to change a process, and when to add governance.

Metaphor mapping

The sharp-eyed book finder: keyword search, BM25, inverted index
The book finder with a good sense: vector retrieval, embedding, semantic search
Explicit terms like BM25: the advantage scenario for exact literal matching
The old way of finding materials by keyword weight: synonymous expressions cause keyword search to miss relevant results
Contract Clause 7.2: vector retrieval may not be good at exact number and symbol matching
Two people set out at the same time: hybrid search uses both sparse retrieval and dense retrieval
Merge, deduplicate, requeue: result fusion, dedup, score normalization, and reranking
Chinese new regulations and English old notices appear at the same time: hybrid search suits enterprise knowledge bases where cross-language, proper nouns, and synonymous expressions are mixed together

Andy Ren Xu's judgment

The practical value of this concept is that it helps you break "AI features that look smart" into deliverable, verifiable, and governable units of work.

BM25sparse retrievaldense retrievalembeddingsrerankingsemantic search

Related distinctions

This concept is easily confused with adjacent concepts; it is recommended to continue to the comparison page.

Hybrid search and dense retrieval Hybrid search and sparse retrieval