FAQ · Typical use cases
How do you accept enterprise knowledge Q&A (RAG)?
RAG acceptance relies on a golden Q&A set: predefined frequent questions and standard answers. The system must answer accurately, cite correctly, and refuse appropriately. Hit accuracy targets before launch; keep sampling and feedback loops after.
Key points
- Build 50–200 golden Q&A items covering core policies, products, and processes
- Metrics: accuracy, citation correctness, refusal rate, hallucination rate
- Sampling: random questions + edge cases (expired docs, no permission)
- User feedback: flag wrong answers → fix knowledge base → re-evaluate
- Joint UAT sign-off with business owner
How to do it
- Business provides Top question list
- Annotate standard answer source for each question
- Run automated/semi-automated evaluation until targets are met
- Weekly wrong-answer review in the first month after launch