Attach documents and webpages so the agent answers from your own policies, pricing, and procedures instead of guessing.
A knowledge base is the set of documents an agent can read while answering. Without one, the agent knows your product only through the system prompt and whatever your client tools return. With one, it can look up the details that live in a PDF or a help center page and quote them back accurately.
Knowledge is configured per project in the dashboard and applies to every conversation immediately — there is nothing to deploy.
Knowledge works best for stable reference material that a person would otherwise have to look up:
It is the wrong home for anything live or per-user. Order status, account balances, and current inventory belong in client tools, which read your authenticated APIs at the moment the question is asked.
Understanding the retrieval model explains most of the advice further down this page.
Every enabled document's name is listed in the agent's system prompt, so the agent knows what the knowledge base covers before it searches. When a question looks like it depends on your own material, the agent calls a searchKnowledge tool with a natural-language query. The question is then matched against your documents by meaning rather than by keyword, and the closest passages come back with a short summary.
Matching by meaning is why a customer asking about "canceling my yearly subscription" reaches a document that only ever says "annual plan termination." The words do not have to line up.
There is nothing to tune: no chunk sizes, no index settings, no relevance thresholds. Each document is prepared for search when you add it, so it becomes searchable the moment it reaches Ready.
Two consequences are worth knowing:
2026 Refund and Cancellation Policy tells the agent when to search. export-final-v3 tells it nothing. Names are the cheapest accuracy win available.An answer that used knowledge carries its sources underneath it. Each citation shows the document name and, when the matched passage sits under a heading, that heading too — so a reader sees Refund Policy 2026 · Annual plan termination rather than only a filename. Hovering a citation reveals the passage the answer was drawn from, which is usually enough to confirm it without opening the document.
The heading comes from the extracted markdown, so structure in the original file is what makes citations precise. A document that extracted without headings still cites correctly; it just points at the file rather than the section.
Open Add data source and pick Import documents to upload files, or Learn from a webpage to fetch a public URL.
The document is converted to markdown and indexed for search in the background, appearing as Processing until both finish. The table refreshes on its own; most documents settle within a few seconds.
Once it reads Ready, open the row's View content action and skim what was actually extracted. This is the text the agent will search — checking it now avoids debugging a bad answer later.
Turn on the Search knowledge toolkit under Tools. Documents alone do not give the agent the ability to search them.
Uploads are limited to 10 MB each and must be one of these formats:
| Format | Extensions |
|---|---|
.pdf | |
| Word | .doc, .docx |
| Excel | .xls, .xlsx |
| PowerPoint | .ppt, .pptx |
| Text | .txt |
| Markdown | .md |
| HTML | .html |
| CSV | .csv |
| JSON | .json |
Anything else is rejected on upload. Scanned PDFs with no text layer are accepted but extract nothing useful, so they fail with a "no readable content" error.
Paste a public URL and the page is fetched and reduced to its main content — navigation, footers, and sidebars are dropped. The URL must be http or https, and pages behind authentication cannot be reached.
Webpage documents do not re-fetch on a schedule. Use the row's Refresh from source action when the page changes.
Only add pages you have the right to use. Fetched content is stored with your project.
| Status | Meaning |
|---|---|
| Processing | Extraction and indexing are running. The document is not searchable yet. |
| Ready | The document is indexed and the agent can search it. |
| Failed | Extraction or indexing did not succeed. Hover the status for the reason. |
| Disabled | The document is indexed, but you turned it off. It is not searched. |
Disabled is a switch you control, not a failure. Use it to retire a document temporarily without losing it — an expired policy you may need to restore, or a page you want to exclude while you verify a replacement.
Each row exposes the actions you need to keep the base trustworthy:
| Limit | Value |
|---|---|
| Documents per project | 50 |
| File size per upload | 10 MB |
| Extracted text kept per document | 500,000 characters |
| Document name length | 120 characters |
| Documents named in the system prompt | 25 most recent |
Text beyond 500,000 characters is truncated rather than rejected, so a very long document is stored only up to that point.
The agent says it has no knowledge base. The Search knowledge toolkit is off, or no document is both Ready and enabled. All three conditions are required before the tool exists at all.
The agent ignores a document. Check its name. If the name does not suggest the topic, the agent may never search it. Rename it and ask again.
A document failed. Open the status tooltip. The usual causes are a scanned PDF with no text layer, an empty file, or a page that returned no readable content. Re-export the source as text and upload again.
A document says it needs re-adding. Documents added before knowledge search moved to semantic matching were never indexed. Delete the row and add the source again.
Answers cite outdated content. Webpage documents are only as current as their last sync. Use Refresh from source, then confirm with View content.