Learned Facts & Knowledge Gaps

Learned Facts & Knowledge Gaps

Your agent pays attention to what visitors ask and how it responds. Over time, it extracts two things:

• Learned facts — Q&A patterns where your agent gave consistent, confident answers. These become draft entries in your knowledge base.
• Knowledge gaps — topics visitors asked about repeatedly that your agent couldn't answer well.

Reviewing What Your Agent Has Learned

In admin chat:

"Show me what my agent has learned"

This calls show_learned, which lists draft Q&A facts extracted from past conversations. Each entry shows the question, the answer, and how many conversations it appeared in.

You can approve entries that look accurate:

"Approve that one" or "Yes, approve the first three"

Approved facts are added to your knowledge base immediately.

Knowledge Gaps

"Show knowledge gaps"

This calls show_gaps and lists topics visitors asked about that your agent couldn't confidently answer — sorted by how often they came up. These are your clearest signal for what to add next.

For each gap, you can:

• Add a knowledge entry directly: "Add an entry about [topic]: [the answer]"
• Point to a URL: "Add my [topic] docs page: https://..."
• Write a Q&A: "The answer to 'what is X' is: ..."

How Extraction Works

After every 10 conversations (or when the conversation log hits 10K chars), your agent reflects on what it's been talking about. During this reflection, it scans the log for consistent Q&A patterns:

• Questions asked 3+ times with a consistent answer → draft learned fact
• Questions where the agent repeatedly said "I'm not sure" or gave vague answers → gap

This happens automatically in the background — no action needed on your part.

Tips

• Check gaps after your first 50 conversations — by then, patterns are clear
• A gap appearing 5+ times is a strong signal to add that content
• Don't approve learned facts that are slightly wrong — edit first, then approve