Vector search is superb at meaning and famously casual about exact strings — ask about "SKU-A1" and it may happily return the chunk about SKU-A2. Hybrid search runs a second, keyword-and-identifier search leg alongside the vector search and merges the two rankings, so exact names, codes, and error numbers can never be out-scored by a look-alike. It's configured per chat, on the chat's Context tab.
- Hybrid Search — the keyword/exact-identifier leg. Free (one extra database query per question) and on by default for every chat. Turn on Disable Hybrid Search only to compare against vector-only search.
- Entity Safety — softly prefers chunks whose recorded entities match the question and nudges down conflicting ones (the SKU-A2 chunk when you asked about SKU-A1). Never removes results — a penalized chunk still returns when nothing better exists. Works best on sources with entity extraction enabled.
- Query Planner — has the accuracy model split a compound question ("compare X with Y") into focused sub-searches and merge the results. One low-cost AI call per message; if the call fails, the chat falls back to the plain question automatically.
- Accuracy LLM Profile — a dedicated LLM profile used only for planning, reranking, and verification. Point it at a small, cheap model; your answer quality still comes from the chat's main profile. Leave it empty to use the chat's main profile for these checks too.
- Reranker Strategy — after the searches merge, a reranker re-orders the top candidates by true relevance to the question. Accuracy LLM uses the profile above; Voyage uses Voyage AI's dedicated reranker. Either costs one call per message, and any failure quietly preserves the merged order.
The Expected accuracy calls readout at the bottom of the section totals the per-message calls for your current switch positions.
Voyage setup #
Voyage AI is optional — Wyregraf remains fully functional with OpenAI and Anthropic alone. To use the Voyage reranker (or Voyage's contextualized embeddings), open an embeddings profile and fill in the Voyage fields:
The Voyage API key authenticates both features; Voyage reranker model defaults to rerank-2.5-lite. Back on the chat, pick that profile in the Voyage Profile dropdown and set Reranker Strategy to Voyage.
When to use what #
- Sites full of product codes, model numbers, or error codes: Hybrid Search is nearly mandatory, and Entity Safety stops near-miss citations.
- Visitors ask long, multi-part questions: add the Query Planner.
- Retrieval finds the right ten chunks but hands the model the wrong four first: add a reranker.
Changes here take effect immediately — no reindexing needed. Verify each step with the Retrieval Lab before enabling the next.
What to do next #
Once retrieval is finding the right evidence, make the answers prove it — Citations & Answer Verification.
