Tagged: vector search
2 posts
Three signals, one answer
Follow a single user question through a retrieval pipeline: from pronoun resolution to three parallel searches, rank fusion, and a final reranker pass. Discover why one similarity score is not enough.
Not every question needs the knowledge base
Reduce latency, avoid irrelevant answers, and save GPU costs by skipping knowledge base retrieval for queries that are not questions. How pattern matching and semantic gates filter trivial queries before they hit the vector store.