How to Improve Your Bot's Answers
Your bot is only as good as its content and configuration. Here’s a practical guide to diagnosing and fixing answer quality issues.
Step 1: Use the Debug panel
Section titled “Step 1: Use the Debug panel”The Debug panel in the Playground is your most important tool. For every response, it shows:
- Source chunks — the actual content pieces the bot used. Are they relevant?
- Relevance scores — how well each chunk matched the query. Low scores mean weak matches.
- Enhanced query — how the bot interpreted the question. Is it reasonable?
If the right content is retrieved but the answer is wrong: Your system prompt needs work (Step 2).
If the wrong content is retrieved: Your RAG threshold or content needs work (Steps 3—4).
If no content is retrieved: You have a knowledge gap (Step 5).
Step 2: Refine your system prompt
Section titled “Step 2: Refine your system prompt”The system prompt (base prompt) tells the bot who it is, how to behave, and what to focus on. A good prompt makes a huge difference:
Be specific:
You are a technical support assistant for [Product Name]. Help users troubleshoot issues and find documentation. Be concise and link to source pages.Set boundaries:
Only answer questions about [Product]. For billing questions, direct users to [email protected].Control tone:
Be friendly and professional. Use simple language. Avoid jargon unless the user uses it first.Step 3: Adjust the RAG threshold
Section titled “Step 3: Adjust the RAG threshold”The RAG Score Threshold controls how relevant content must be before the bot uses it:
- Too high (e.g., 0.7): The bot says “I don’t know” too often because it’s being too picky.
- Too low (e.g., 0.1): The bot uses marginally relevant content, leading to off-topic answers.
- Default (unset): Leave the slider at its default position and the engine auto-tunes the threshold (around 0.30, depending on your content type) — a good starting point for most use cases.
How to tune it:
- Ask questions in the Playground.
- Check the Debug panel for relevance scores.
- If good content is being filtered out (scores below threshold), lower the threshold.
- If bad content is being included, raise it.
Step 4: Improve your content
Section titled “Step 4: Improve your content”Sometimes the issue isn’t the bot — it’s the content:
- Enable Reader Mode if pages include navigation or sidebar text that confuses retrieval.
- Add Q&A entries for frequently asked questions that aren’t well-covered.
- Re-crawl after updating your website to pick up content changes.
- Check for empty or truncated pages in your crawl report.
Step 5: Fill knowledge gaps
Section titled “Step 5: Fill knowledge gaps”Go to Analytics and check the Knowledge Gaps section. If users are asking about topics not in your content:
- Add the missing content to your website and re-crawl.
- Or add Q&A entries directly for quick coverage.
See Find and Fix Knowledge Gaps for details.
Step 6: Choose the right model
Section titled “Step 6: Choose the right model”If you’ve done all the above and answers still aren’t good enough, try a more capable model:
- GPT-5 Mini handles most content well.
- GPT-5.2 or Claude Sonnet 4.6 are better for complex technical content.
See Choosing the Right AI Model.
Quick checklist
Section titled “Quick checklist”- Check the Debug panel for retrieval quality
- Write a specific, clear system prompt
- Set the RAG threshold based on your content (leave at default to auto-tune, ~0.30)
- Enable Reader Mode if content is noisy
- Add Q&A entries for top unanswered questions
- Check Analytics for knowledge gaps
- Consider a more capable model if needed
Related
Section titled “Related”- Bot Settings Reference — all tunable settings
- Use the Bot Playground — debug panel details
- Find and Fix Knowledge Gaps — fill missing content
- Choosing the Right AI Model — model comparison