In plain words
Old-style search matches words. Type "cheap flights" and it looks for pages containing "cheap" and "flights". A page about "low-cost airfares" might be missed, even though it is exactly what you wanted.
Semantic search matches meaning instead. An AI model turns every document, and your question, into a list of numbers that captures what the text is about. Texts about similar things get similar numbers. The search then returns whatever sits closest to your question.
It is why you can type a question in your own words and still find the right help page, photo or email. It also powers the lookup step inside many AI chatbots that answer from documents.
Everyday example
A shopkeeper who gets what you mean
At a kirana store you ask for "something for a sore throat". A shelf label search would find nothing, because no packet says those words. A good shopkeeper hands you honey and lozenges, because they know what you mean. Semantic search is the shopkeeper: it understands the need, then finds things that fit it.
Where the example breaks: The shopkeeper understands why you want it. Semantic search only measures how close two texts are in meaning, so it can return something on the right topic that does not actually answer your question.
How it actually works
Semantic search has two stages: turn text into numbers once, then compare numbers for every search.
Embed the question
Your query goes through the same model, so it lands in the same space. "Leave policy" ends up near "annual holiday entitlement" even with no shared words.
Find the nearest
The system measures closeness, usually with cosine similarity, and returns the nearest items. With millions of documents, a vector database finds near neighbours fast without checking every one.
Mix in keywords
Meaning search can miss exact names, codes and part numbers. Many systems combine it with classic keyword ranking. Anthropic's tests found "Embeddings+BM25 is better than embeddings on their own" (BM25 is a keyword method).
- Older idea
- Latent semantic indexing, developed at Bellcore in the late 1980s, already aimed to match documents by meaning rather than exact words
- Big moment
- October 2019: Google said its BERT language model would help it understand "one in 10 searches in the U.S. in English"
- Speed trick
- Sentence-BERT (2019) cut finding the most similar pair among 10,000 sentences "from 65 hours with BERT / RoBERTa to about 5 seconds"
- Used in
- Help centres, photo and email search, shopping search, and the retrieval step of RAG chatbots
Questions people ask
What is the difference between semantic search and keyword search?
Keyword search finds documents that contain the words you typed. Semantic search finds documents whose meaning is close to your question, even with different wording. Keyword search is better for exact names and codes; semantic search is better for questions in everyday language. Many systems use both.
Is semantic search the same as vector search?
Nearly. Vector search is the technique: compare lists of numbers and return the closest. Semantic search is the goal: find things by meaning. Most semantic search today is done with vector search over embeddings.
Does Google use semantic search?
Yes, as one part of its system. In October 2019 Google said it was using BERT, a language model, to understand the meaning of queries, citing one in ten English searches in the U.S. Its search also uses many other signals.
Do I need a vector database for semantic search?
Not for small collections. A few thousand items can be compared directly. A vector database helps when you have millions of items and need answers in milliseconds.