How to structure a knowledge base that AI agents can actually use

An AI support agent reads your docs the same way a new hire reads them — except without the ability to ask questions. If the docs are scattered, stale, or contradictory, the agent's answers will be scattered, stale, or contradictory. No amount of prompt engineering fixes bad source material.

The five rules of an AI-ready knowledge base

1. One source of truth per topic

If refund policy lives in 3 places, retrieval pulls whichever one has the best embedding match — which may not be the most current. Pick one document per topic. Delete or redirect the others. Archival matters as much as authorship.

2. Short, self-contained sections

A well-chunked doc has sections of 150-400 words, each one standalone enough to answer a question without requiring context from surrounding sections. If a section starts with "As mentioned above," it's too coupled — rewrite.

3. Explicit metadata

Every doc should have: last-updated date, owner (a real name), applicable plans/product versions, and one-line summary. This lets the agent filter: "show me only docs updated in the last 90 days for the Pro plan."

4. Clear scope markers

At the top of each doc, write one sentence: "This document explains X for users on Y plan." It sounds redundant to humans. It's gold for retrieval because it sets an explicit match signal.

5. A regular prune cycle

Every quarter, archive docs that haven't been viewed. If nobody reads it and the agent never surfaces it, it's either redundant or wrong. Either way, it's adding noise to retrieval.

What to delete today

• Outdated changelogs that describe features that no longer exist.
• Screenshots with UI from 2+ versions ago.
• Help articles with the phrase "coming soon" (if it's still coming, move to a roadmap doc).
• Any doc last updated more than 18 months ago that hasn't been reviewed since.

Reindex on change, not on schedule

If your platform reindexes nightly, a doc edit at 9am won't surface to the agent until the next morning. That's a full business day of stale answers. Reindex on doc-change webhook instead. Signorian does this automatically; some other tools don't.