Skip to content

Validating & improving knowledge

Open in ChatGPT Open in Claude

Path: Agent Studio → Knowledge

Adding a source isn’t the finish line. You still need to confirm the agent actually answers correctly from it, and keep that content accurate as your product changes. Run this as a continuous loop.

1. Validate before users do — in the Playground

Section titled “1. Validate before users do — in the Playground”

Open the Playground and ask the questions your users actually ask. Check each answer against what the source contains:

  • Right answer, right source — the content is covered and retrievable. Done.
  • Wrong or “I don’t know” — the content is missing or stale. Add or re-sync a source.
  • Right answer exists but the agent didn’t find it — a retrieval miss, not a coverage gap. Fix it with Pointers.

This is the fastest way to catch a problem before it reaches a real user.

The Playground tells you what you thought to test; Conversations tells you what users actually hit. In the filter bar, open Conversation Signals and select the Agent Knowledge Gap tag to see exactly the chats where the agent lacked the content to answer.

Each one is a concrete to-do:

  1. Read the conversation to understand what the user wanted.
  2. Add or update the source that should have answered it — usually a new internal article or a re-synced one.
  3. Re-test the question in the Playground to confirm it’s fixed.

This conversation-driven loop is how your Knowledge Base gets better over time instead of drifting out of date.

A wrong or incomplete answer is most often an out-of-date source. Review sync status regularly:

  • Crawlers and connectors re-sync on their own — confirm the last sync date is recent and the article count looks right.
  • Files, single pages, and internal articles only change when you change them. Refresh them when the underlying product or policy changes.

4. When the right content isn’t retrieved

Section titled “4. When the right content isn’t retrieved”

If the article exists but the agent reaches for the wrong one — or doesn’t surface it because the user’s wording differs from your terminology — that’s a retrieval problem, not a coverage problem. Those rare cases are exactly what Pointers are for.