Skip to main content
Vector search compares the embedding of a visitor’s question with stored source chunks. It retrieves semantically related material even when the wording differs.

​
Query flow

RAG_TOP_K defaults to 5 in the agent configuration. FAQ prechecks request only the top result and apply a 0.85 threshold. Normal knowledge retrieval can use multiple chunks to assemble context.

​
Relevance levers

​
Diagnose poor results

Confirm the source is indexed, vectors exist for the document, the query uses the same embedding dimension, and organization identifiers match in both record and payload.
Inspect duplicate content, headings lost during extraction, overly broad chunks, stale vectors, and missing catalog/type filters.
Separate retrieval quality from generation quality. Inspect returned payload text first, then the system prompt, tool output formatting, and selected chat model.
A vector similarity score is not a factual confidence score. The answer pipeline still needs source attribution, policy constraints, and tools for live account data.
Last modified on July 17, 2026