Definition
Semantic search retrieves results based on meaning rather than literal string matching. A query like "how to fine-tune a model" can match documents that say "training tips for neural networks" even without shared keywords.
It typically relies on embedding models to encode queries and documents into vectors, then ranks by cosine similarity or related metrics.
In simple terms
Keyword search is like looking up an exact product SKU. Semantic search is like describing what you want to a knowledgeable shop assistant who understands synonyms, context, and intent.
Where you see it
- Google Search uses semantic signals alongside traditional ranking.
- Notion and enterprise wikis embed pages for AI Q&A.
- Legal and medical databases find conceptually related cases and papers.
- Somali content platforms can match Af-Soomaali queries to relevant passages across dialects.
How it works
1.Embed the corpus
Index all searchable documents as vectors in a vector store.
2.Embed the query
Convert the user's question into the same vector space.
3.Retrieve neighbors
Return top-k chunks whose vectors are closest to the query vector.
4.Optional reranking
A cross-encoder or LLM reranker refines order for higher precision.
Why it matters
- Users rarely know the exact words in your documentation — semantic search meets them where they are.
- It is the retrieval half of most RAG systems.
Often confused
Semantic search always beats keyword search.
Exact matches (SKUs, error codes, IDs) still need keyword or hybrid retrieval.