Recherche sémantique
Définition
Semantic search retrieves items by meaning rather than exact keywords. Query and documents are embedded; récupération returns the most similar vectors (par ex. cosine similarity or ANN search).
C'est the récupération backbone of RAG: see embeddings and vector databases for how vectors are produced and stored. Utilisez-le quand users express intent in natural language and you want “similar meaning” rather than literal keyword match. Combines well with keyword (hybrid search) when exact terms matter.
Comment ça fonctionne
The query (and optionally filters) is sent to an embedding model qui produit a vector. Vector search (par ex. k-NN or approximate k-NN over an index of document vectors) returns the ranked docs (or chunk IDs) with highest similarity (par ex. cosine or dot product). Embedding models are trained so that semantically similar text maps to nearby vectors; the same model is used for queries and documents. Indexing can be offline (batch) or incremental; scale and latency determine whether you need an approximate index (HNSW, IVF) and a dedicated vector database.
Cas d'utilisation
Semantic search is used whenever you need to find items by meaning rather than exact keywords (RAG, recommendations, dedup).
- RAG récupération: finding relevant chunks for a user query
- Recommendation and “similar item” search
- Duplicate or near-duplicate detection in document sets
Documentation externe
- Sentence-BERT — Dense récupération models
- LangChain – Vector stores