Knowledge base hygiene
A knowledge base (KB) is only as good as what’s in it. An agent that retrieves stale, duplicated, or off-topic content will answer confidently and wrongly. Keeping KBs clean is the cheapest way to improve answer quality. KBs are attached to agents in the builder — see Attaching knowledge bases.
Connect only what’s relevant
- Attach the fewest sources that cover the agent’s job. Extra sources add noise that competes with the right answer.
- Give each KB a clear title and description so retrieval and humans both know what it’s for.
- Don’t reuse one giant KB across unrelated agents. Scope KBs to a purpose.
Keep content current
- Remove outdated material. A superseded policy or price list is worse than no source at all.
- Refresh changing sources on a cadence. For data that updates regularly, consider a scheduled run to re-ingest or re-check it.
- De-duplicate. Multiple near-identical documents split retrieval and can return the weaker copy.
Structure for retrieval
Search works on chunks of text, so structure helps:
- Prefer clear headings and short, self-contained sections over long unbroken documents.
- Put the answer near the question it answers; avoid burying key facts in appendices.
- Strip boilerplate (navigation, repeated footers) that adds tokens without meaning.
Read vs. read-write access
When attaching a KB, choose the access level deliberately:
- Read — the agent can search and cite the KB. Use this for reference material.
- Read-write — the agent can also add to the KB. Use this only when the agent is meant to capture new knowledge, and review what it writes.
See Read vs read-write.
Verify retrieval
After changing a KB, test the agent with questions whose answers live in that KB and confirm it cites the right source. If it pulls the wrong passage, the fix is usually in the content (split, retitle, or remove), not the prompt.