Reranking is a two-stage retrieval process that scores an initial batch of retrieved documents using a cross-encoder model to surface the most relevant passage to the top.
Cross-encoder reranking is a secondary relevance scoring pass that jointly evaluates the query and retrieved candidate chunks to reorder them with high semantic precision before LLM synthesis.
Initial vector retrieval is fast (bi-encoder), computing similarities independently. However, bi-encoders can miss subtle semantic relationships.
A reranker takes the top-20 retrieved candidates and scores each candidate against the query simultaneously, pushing the true answer to position #1.
SiteMind applies Reciprocal Rank Fusion (RRF) and relevance scoring across dense vector, full-text tsvector, and trigram candidates to ensure optimal chunk ranking before generation.
Estimate support savings, token counts, or test prompt injection security guardrails with our zero-cost sandboxes.
Related Technical Concepts
Hybrid Search (Dense + Sparse Retrieval)
Hybrid Search merges dense vector search with sparse keyword matching to deliver superior accuracy across both conceptual and exact-term queries.
Reciprocal Rank Fusion (RRF)
Reciprocal Rank Fusion (RRF) is an algorithmic scoring method that merges and ranks search results from multiple independent retrieval pipelines.
Vector Database & pgvector
A vector database is specialized storage optimized to index high-dimensional embeddings and execute sub-second approximate nearest neighbor (ANN) searches.
Crawl your domain in 2 minutes, get instant grounded answers, and see your deflection rates.