Start with the free ones, and measure as you go:
- Save a baseline. Put a handful of real visitor questions into the Retrieval Lab and run them with everything off, so later changes have something to beat.
- Hybrid Search on the chat (on by default) — free, immediate, and fixes the most common failure: exact product names, SKUs, and error codes losing to look-alike content.
- Entity Safety on the chat, plus entity extraction on your key sources — stops the near-miss citation (the Model Plus chunk answering a Model Pro question). The source needs a reindex after enabling extraction.
- Coreference rewriting and contextual blurbs on reference-heavy sources — chunks become self-contained, which helps both matching and answer quality. One AI call per chunk at reindex.
- Query Planner and a reranker if visitors ask long or multi-part questions — one low-cost call per message each.
- Citation Integrity, then Strict Answer Verification last, on chats where a wrong answer is worse than no answer.
Re-run your Lab cases after each step; keep what improves your numbers, switch off what doesn't. The full tour is in Improving Answer Accuracy.
