FAQ · Data & knowledge
What if the knowledge base gives wrong answers?
Wrong answers usually mean data is not standardized, chunking is poor, golden Q&A acceptance is missing, or nobody owns document versions. Start with data inventory and cleanup, then build a golden set for ongoing evaluation — not by swapping models repeatedly.
Key points
- Inventory document sources: are policies, products, and records current, and do they have owners?
- Clean duplicate, expired, or low-quality scanned content; standardize naming and metadata
- Build a golden Q&A set (50–200 items) as a pre-launch gate
- Tune chunking and retrieval; require answers with citation sources
- Assign a knowledge base maintainer and update cadence (weekly/monthly)
How to do it
- List Top 20 frequent questions and where standard answers come from
- Clean and upload core documents; tag classification level and expiry
- Run golden-set evaluation; promote only after accuracy targets are met
- Close the feedback loop: flag wrong answers → fix → re-evaluate