Wyregraf's accuracy features are a set of optional upgrades to how the assistant finds and cites your content: smarter indexing on each source, a second keyword-based search leg on each chat, and checks that keep the final answer honest to the evidence. Every one of them ships switched off — your existing sources and chats keep working exactly as before until you opt in.
They come in three layers:
- Index-time enrichment (per source) — rewrite ambiguous chunks, extract entities, and preserve document structure so each stored chunk stands on its own. Set up in Content Enrichment & Chunking.
- Query-time retrieval (per chat) — combine the vector search with exact keyword and identifier matching, plan multi-part questions, and rerank the candidates. Set up in Hybrid Search & Reranking.
- Answer-side checks (per chat) — require the model to cite the supplied evidence and optionally verify every claim before the visitor sees it. Set up in Citations & Answer Verification.
What each feature costs #
Some features are free at run time; others spend an extra AI call. As a quick map:
- Free — structure-aware chunking, hybrid search, entity safety.
- One AI call per chunk, at indexing time — content enrichment (blurbs, rewriting, entities, facts share a single combined call). Summaries and Questions each cost the same way and are on by default.
- One low-cost AI call per visitor message — the query planner, and the reranker (LLM or Voyage strategy).
- Per answer — citation integrity rides the normal answer request; strict verification adds one support check and at most one regeneration.
The Estimated indexing calls readout on each source's Retrieval Accuracy tab and the Expected accuracy calls readout on each chat show these numbers for your actual configuration before you commit.
A sensible rollout order #
- Measure first. Write a handful of benchmark questions in the Retrieval Lab and save a baseline, so you can see whether each change actually helps your content.
- Keep Hybrid Search on (the default) — it's free and fixes the most common miss: exact names, SKUs, and error codes that vector similarity alone can skip.
- Enrich your most important sources — coreference rewriting and entity extraction make individual chunks self-contained, which pays off most on reference-style content. Reindex the source afterward.
- Add citations and verification last — they trade a little latency and cost for answers that are provably grounded in your content.
Re-run your Retrieval Lab cases after each step and compare against the baseline.
Keeping an eye on it #
Wyregraf → Options → Diagnostics gains a Retrieval accuracy group: schema/backfill progress, keyword-index availability, sources flagged for reindex, enrichment failures, and a running count of accuracy AI calls.
What to do next #
Start with Content Enrichment & Chunking for your sources, or jump straight to Hybrid Search & Reranking for the free chat-side win.
