Skip to main content
Your agent answers from the sources you give it. They all live in one Knowledge base panel on the Your Agents page, in four tabs.

​
The four source types

Files

PDF, Word (.doc, .docx) or plain text (.txt). Up to 4 MB per file.

Text

Paste anything the agent should know — a policy, an FAQ, a page of notes. No file needed.

Websites

Add a URL and the pages get read for you when you publish.

Q&A

A question customers actually ask, and the answer you want given.

​
Files

4 MB is a hard per-file ceiling, on every plan including Elite. It is a platform limit on request size, not a plan feature — a larger file is rejected before it reaches AgntLabs at all. Split large documents up.
This is separate from your storage allowance, which is the total across your whole account and every agent in it:
Five focused documents usually beat one fifty-page manual. Retrieval works on chunks, and tight documents produce tighter chunks.

​
Websites

Paste a publicly reachable URL. Pages are fetched when you publish, not when you add the URL. A site is read as it exists at publish time. To pick up changes, publish again.
AgntLabs will not crawl the same URL twice for one agent — if the URL is already there, add it once and republish.

​
Q&A pairs

The most precise control you have. Write the question the way a customer would ask it, and the answer exactly as you want it given. Q&A pairs are the right tool when a specific question needs a specific answer — pricing, a policy, a caveat you always want mentioned.

​
Publishing

Adding a source does not make it usable. Publishing does. Press Publish and AgntLabs reads each pending source, splits it into chunks, embeds those chunks and stores them for this agent. It also applies edits and deletions you have made to sources that were already live.
Edits to a live source are staged, not immediate. Change a Q&A answer and the old answer keeps being used until you publish. This is deliberate — it lets you make several changes and apply them together.
If some sources fail, the rest still go through. Failed ones stay in the queue and can be retried.

​
Reading the summary

The closed Knowledge base row reports three numbers:
  • Total — every source in the list, including failed ones
  • Unpublished — added but not yet embedded; your agent cannot use these
  • Failed — could not be read at all
A failed website counts towards the total, so a row reading “3 sources” can describe an agent that can answer from two. Open the panel to see which.

​
Fixing what your agent cannot answer

When your agent is asked something on-topic that it cannot answer, that is recorded as a knowledge gap. Gaps show up in two places:
  • Home — a Needs attention card with the top failures
  • Analytics → Performance — the full ranked list, with the reason:
    • No source covers this — nothing relevant was retrieved. Add a source.
    • Answer incomplete — a source came back but did not carry the answer. Edit it.
Clicking a gap opens the Knowledge base panel on the Q&A tab with the question already filled in, so you can write the answer immediately.
This is the single most useful loop in the product: read the gaps, answer them as Q&A pairs, publish. Your agent gets measurably better every time.