techniques
Hybrid search
Ensemble retrieval that combines dense embeddings + sparse BM25 + authority signals — what every modern engine actually uses.
Definition
Hybrid search is the retrieval architecture that combines dense vector similarity (semantic) with sparse keyword matching (BM25) and metadata weighting (authority, recency). Reciprocal Rank Fusion (RRF) is the standard combiner. Both Perplexity and Anthropic's web tool publicly describe using hybrid retrieval. Pure-dense retrieval misses exact-term queries (product names, error codes); pure-sparse misses paraphrases.
Example
Query: 'Stripe 402 error' — sparse BM25 finds pages with 'Stripe 402' verbatim, dense finds pages discussing 'payment-failed responses', RRF picks the union.How to optimize
Use buyer vocabulary (helps dense) AND exact-term repetition (helps sparse). Cover product names, error codes, model numbers verbatim somewhere on the page even if you reword everywhere else.