Learning path · Retrieval & Ranking · 43
Semantic Search
Finding documents by meaning similarity between query and corpus embeddings rather than exact keyword match.
Why it matters
- Captures paraphrases and conceptual questions keywords miss.
- Core retrieval stage before RAG generation.
- Fails on rare proper nouns without hybrid lexical backup.
Key ideas
- Query embedding
- Top-K retrieval
- Similarity thresholds
Semantic search embeds the user question, pulls nearest neighbours from the vector index, and passes hits to downstream rerankers or the LLM. Tune top-K: too low misses evidence; too high adds noise and tokens. Log queries with zero hits—they signal glossary gaps or stale indexes. Always pair with ACL filters so embeddings never leak cross-tenant data. Publish zero-hit queries to content owners monthly; they reveal glossary gaps faster than tuning embedding models alone. Ship only after eval gates pass on representative production failures.
Updated 2026-08-09 · Full learning path