Semantic search is search that matches on meaning rather than exact words. A query for “my card was declined” can return an article titled “Payment failures: common causes”, even though the two share no keywords. It works by converting text into vector embeddings and finding the closest ones, instead of looking for matching terms.
Key takeaways
- Keyword search matches words; semantic search matches intent.
- It handles synonyms, paraphrases and natural-language questions.
- It struggles with exact identifiers, codes and rare product names.
- Hybrid search combines both and usually beats either alone.
- It is the retrieval layer behind most RAG systems.
Semantic search explained in simple terms
A librarian who only does keyword search hands you books containing your exact words. A librarian who understands your question brings you the right book even if you described it loosely. Semantic search is the second librarian, built from embeddings.
How it works
- Every document or passage is embedded once and stored in a vector database.
- The user’s query is embedded with the same model.
- Vector search finds the nearest passages.
- Results are ranked by similarity, optionally refined by a reranker.
Semantic vs keyword search
| Query | Keyword search | Semantic search |
|---|---|---|
| “can’t log in” | Misses “password reset” articles | Finds them |
| “SKU-48213” | Finds the exact product | May return similar products instead |
| “cheapest way to ship to Pune” | Needs matching words | Understands intent |
| “GST rate for software services” | Works if terms match | Handles paraphrases |
| Rare surname or code | Reliable | Often unreliable |
The pattern is consistent: semantic search wins on natural language, keyword search wins on exact strings.
A worked example
A help centre has an article titled “Resetting your account password”. A user searches “forgot my login details”.
- Keyword search: no overlap on “forgot”, “login” or “details”, so the article may rank poorly or not appear.
- Semantic search: the query embedding sits close to the article embedding, because both concern account access, so it ranks first.
Where semantic search is used
- Site and help-centre search.
- Enterprise search across wikis and drives.
- E-commerce product discovery.
- Support ticket routing and deduplication.
- Retrieval in RAG assistants.
- Internal linking suggestions for content sites.
Practical guidance
- Always use the same embedding model for documents and queries.
- Chunk long documents sensibly; see document chunking.
- Add metadata filters for language, date and permissions.
- Combine with keyword search for identifiers and exact phrases.
- Evaluate with real queries and known correct results, not intuition.
Limitations
- Weak on exact codes, part numbers and rare names.
- Cannot express strict boolean logic naturally.
- Quality depends heavily on the embedding model and the languages it covers.
- Similar content is not always the content that answers the question.