← AI Terminology

Semantic Search

Semantic search retrieves documents by meaning using embeddings and similarity, rather than only exact keyword matches.

It powers modern RAG, enterprise search, and recommendation-like retrieval.
Why It Matters in AI
Users ask with different words than documents use. Semantic search finds relevant content despite paraphrase — transformative for knowledge bases, but weak alone on exact IDs and rare terms (hence hybrid search).
Key Points
Aspect Description
Eval nDCG, recall@k on labeled queries
Models Sentence transformers, proprietary embedding APIs
Related BM25, hybrid search, rerankers
Pipeline Embed query → ANN over doc embeddings → top-k
Strength Paraphrase and conceptual match
Weakness Exact SKUs, names, legalese keywords
Simple Analogy
Finding a café when you search “somewhere cozy to read” even if no listing uses the word cozy — matching intent, not strings.
Common Usage Examples
  • SentenceTransformer.encode + cosine top-k
  • Azure/OpenAI/Cohere embeddings search
  • Intranet RAG search boxes
  • Compare to Elasticsearch BM25-only
Summary
In short: Semantic search finds content by meaning via embeddings — the retrieval style that makes RAG work across paraphrases.