Loading...
Loading...
Go beyond simple vector search — hybrid search and re-ranking for better results
Query -> [Stage 1: Candidate Retrieval] -> Top 50 -> [Stage 2: Re-ranking] -> Top 5
Fast, cheap Slower, more accurateVector search (semantic): Finds meaning, handles synonyms
BM25 (keyword): Matches exact terms, good for codes/names
Combine both: score = alpha * bm25 + (1-alpha) * vector
A cross-encoder model scores (query, document) pairs directly — more accurate than embedding similarity.
Popular re-rankers: Cohere Rerank 3.5, BGE-reranker-v2
Query -> Embed -> Vector Search + BM25 -> Hybrid Fusion -> Top 50 -> Re-ranker -> Top 5 -> LLMDesign a retrieval strategy:
apps = {
'Legal document search': {'exact': True, 'semantic': True, 'rerank': True},
'E-commerce product search': {'exact': True, 'semantic': True, 'rerank': False},
'Chatbot memory': {'exact': False, 'semantic': True, 'rerank': False},
}