← Back to concept explainers

Concept Fable

Semantic search

Conclusion first:Semantic search retrieves results based on meaning rather than exact keyword matching. It embeds both queries and documents into the same vector space and finds the semantically closest content by computing vector distance. Compared with traditional keyword search, semantic search can handle synonyms, near-synonymous expressions, cross-language queries and fuzzy descriptions, giving more complete recall. In enterprise RAG, semantic search is usually combined with keyword search (sparse retrieval such as BM25) into hybrid search, making up for pure vector retrieval's shortcomings in exact terminology, numeric ranges and structured filtering. Implementation requires choosing an embedding model, determining vector dimensions and indexes, configuring similarity metrics, and continuously monitoring recall drift.

Semantic 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 old curator locked the keyword drawer into the warehouse

The city's public library has an eighty-year history. The borrowing rules have never changed from the beginning: to find a book, first check the index cards, spell the characters in the title or author correctly, then go to the shelf to find it.

Readers began to grow in number, and the way they asked changed too. Some wanted to know "what plant can cure a cough," some looked for "how to keep a small team from arguing," and others only remembered "an old book with a blue cover about a fishing village story."

The administrator rearranged the cards by theme and added more keywords. Readers still sighed: cough and cough suppressant did not match, quarrel and team friction did not match, and fishing village and seaside town did not match either.

The director then had someone make a thick keyword cross-reference table. Next to each phrasing was written which card to look up. The table grew thicker and thicker, and the administrator spent more time consulting the table each day than searching the books.

A new librarian did not continue adding words. She turned each book's content, table of contents and reader notes into a string of meaning coordinates, where each book's position represented roughly what it was about. Readers' questions were also placed on the same map, and books with nearby distances were found.

When the first version came out, someone used "herbs for a sore throat" to find the sore-throat entry in the medical texts, even though those words were not on the index card. Only then did the curator see that what was being matched was meaning, not literal characters.

Later the library used the old and new methods together. Meaning coordinates handled discovering associations, while year, category and collection location were still constrained by precise conditions. The new librarian did not discard precise clues either. If readers knew the title, title lookup was still fastest; if they only described symptoms, scenarios and goals, the meaning map came into play. The two search methods served different entry points. She also set boundaries for the map. Folk remedies and formal medical texts could be similar, but could not be mixed into the same level of credibility; the fishing village in a novel and a navigation manual might have similar scenes, but different uses. When searching for meaning, category, source and use scenario still had to be preserved. Readers also gradually learned to ask questions in natural language. The library served human intent, and no longer required people to first become index cards. The curator said that once questioning became freer, the collection truly opened to ordinary people. But she still reminded everyone that similar does not mean the same, and after finding it, you still have to read the text.

The old curator sent the largest keyword mapping table to the basement: retrieval should not require the questioner to first learn the institution's jargon.

Reveal

This story is about: semantic search

Semantic search retrieves results based on meaning rather than exact keyword matching. It embeds both queries and documents into the same vector space and finds the semantically closest content by computing vector distance. Compared with traditional keyword search, semantic search can handle synonyms, near-synonymous expressions, cross-language queries and fuzzy descriptions, giving more complete recall. In enterprise RAG, semantic search is usually combined with keyword search (sparse retrieval such as BM25) into hybrid search, making up for pure vector retrieval's shortcomings in exact terminology, numeric ranges and structured filtering. Implementation requires choosing an embedding model, determining vector dimensions and indexes, configuring similarity metrics, and continuously monitoring recall drift.

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

  • Index card: traditional keyword/term retrieval, requiring literal matches
  • 'Cough' and 'cough suppressant' don't match: keyword matching can't handle synonyms and semantic variation
  • Keyword mapping table: synonym dictionaries or rule mappings, maintenance costs explode with expression diversity
  • SQL LIKE Query: exact pattern matching, capturing only surface strings rather than intent
  • Turn it into a set of numbers: embedding text into vector space
  • The book with the closest number: vector similarity calculation, based on semantics rather than literal wording
  • Two old and new methods stacked together: hybrid search, a combination of semantic retrieval + keywords/filters
  • No need to learn the system language first: semantic search lowers the user's query precision requirements

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.

embeddingsvector databasedense retrievalhybrid searchsimilarity searchANN search

Related distinctions

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

京ICP备2026014272号