Knowledge sources

Knowledge gives the agent retrievable source material. Adding a source extracts plain text, divides it into passages, creates embeddings, and associates those passages with the current site. At answer time, the agent retrieves relevant passages instead of receiving every source in every prompt.

Choose the right source type#

Use FAQ and products for predictable question/answer or catalog fields rather than burying them in a long document. See FAQs and products.

Add a source#

1. Open the agent's Knowledge page from the agent's sidebar and select Add source.
2. Select Paste text, From URL, or Upload file. YouTube, Podcast, and Blog / RSS are under More source types and need Operate or Scale; on Launch they show a lock.
3. Enter content or choose a file.
4. For From URL, choose Pages to crawl (1, 5, 10, 20, or 40). For fetched sources, optionally set Auto-refresh to Daily, Weekly, or Never.
5. Select Add to knowledge and wait for the result showing the number of indexed chunks.
6. Ask a question whose answer appears only in the new source.

Knowledge ingestion consumes credits because source passages are embedded. It is rejected when the workspace has no active access or usable credit balance.

Crawl public pages safely#

The crawler accepts public HTTP/HTTPS content and blocks unsafe private-network destinations. Start with one canonical page. Increase the page count only when the linked pages are part of the same useful information set; a larger crawl can add navigation, legal boilerplate, duplicated text, or stale archives that weaken retrieval.

The selected count is a maximum, not a promise. Robots behavior, authentication, client-only rendering, redirects, content type, response size, and link structure can reduce the number fetched. A page response is capped at 5 MB and must contain enough extractable text.

For important facts behind login, paste an approved summary or upload a sanitized export instead of attempting to crawl an authenticated page.

Use files and media#

Uploaded files must contain extractable text. OCR is not performed for a scanned image-only PDF; convert it to searchable text before upload. Spreadsheets are read sheet by sheet, and table rows keep their cells together (Coffee | 3000); other layout flattens into reading order, so keep tables simple. Remove a document's password before uploading it, and save macro-enabled Office files (.docm, .xlsm, .pptm) as ordinary .docx, .xlsx, or .pptx first. A file whose contents do not match its extension is refused.

YouTube ingestion prefers an available caption track. If captions are unavailable, transcription is provider- and network-dependent. Podcast audio must be publicly fetchable. Feed results depend on how much article or episode text the feed exposes. After any media import, inspect the source title/chunk count and test a specific fact instead of assuming the full recording was captured.

Plan refresh and replacement#

Daily and weekly schedules are available for URL, YouTube, podcast, and feed sources. The system re-fetches the saved source origin and replaces its retrievable passages when a scheduled crawl succeeds. The source list shows the refresh cadence and last refreshed time.

Paste and upload sources are snapshots. To correct or replace one, delete it and add the updated content. Deleting a source removes its passages from future retrieval; it does not edit historical messages that already quoted or summarized the old version.

Use one clear owner for pricing, policies, and schedules. If two active sources disagree, retrieval can surface either statement. Remove superseded versions rather than adding “new” copies beside them.

Write content the agent can retrieve#

• State the subject in each section: “Pro plan cancellation policy” is more retrievable than “More details.”
• Put the answer in plain language near the relevant heading.
• Include units, currency, timezone, effective date, and exceptions where they matter.
• Use stable canonical URLs and descriptive file names.
• Separate public information from internal instructions or private records.
• Keep persona behavior in Persona and instructions, not inside factual source material.

Test grounded answers#

Create a small acceptance set before publishing:

1. A direct question whose answer is explicit in the source.
2. A paraphrase that uses different words.
3. A boundary question the source does not answer.
4. A recently changed fact.
5. A language variant for each configured agent language.

The correct result is not always a confident answer. When the source lacks the fact, the agent should follow its fallback/escalation instructions rather than inventing one. Check Conversations after launch for repeated unanswered questions and turn validated answers into a source or structured FAQ.

Troubleshooting#

Never upload credentials, private keys, unrestricted customer exports, health records, or any content the public agent must never reveal.