Troubleshooting

A website blocks the crawler

When a crawl cannot reach a site, the run page shows an alert that starts Cannot access website: and names the site, with a Fix access button. Only people who can add sources see the button; everyone else is asked to find someone who can.

The crawler identifies itself with the user agent RecordsLabsIngestion (https://recordslabs.ai/ingestion). Most blocks come from a firewall or bot protection such as Cloudflare, or from a page that needs a login.

Click Fix access. If the alert does not already know the blocked page, paste the page or PDF URL. Click Test and save access. Records Labs loads the page with your browser's request settings and, if that works, saves the access for this website.
If the website needs you to sign in, accept cookies, or pass a security check, open Website requires login or a security check? and click Open secure browser. Sign in, open the exact page, and click I can view it once the page is visible.
Click Try again to fetch the blocked pages again.
If Fix access didn't get through appears, the site owner has to allow us. Click Copy request to copy a short message that asks them to allow the user agent RecordsLabsIngestion through their firewall or bot protection. When they confirm, try again.
Need our server IP addresses or more help? Click Contact support under those steps. It sends your request to us from the app, no mail client needed, and we reply by email. To talk it through instead, click Chat with support next to it. The chat tells us which website was blocked.

Access you save applies only to that website. In the source's details, a blocked site offers Capture with browser first, which opens Browser capture so your own browser can capture the pages. A crawl stopped by a block keeps its refresh schedule and probes the site occasionally. The person who started the crawl (or your admins, if they have left) gets one in-app notification and one email each time the site starts blocking us.

A connector stopped syncing

Open Settings → Connectors. Turn on Needs attention to show only connections with a problem.
If the connection's authorization expired, click Reconnect on the row (it reads Connect when authorization is required). For providers that sign in with OAuth, this takes you straight to the provider's consent screen.
Open the latest sync run. Its tabs are All, Created, Updated, Failed, Skipped, and Ignored. Open Failed to see why each item failed, then click Retry failed. Attachments are listed as their own items, so a file that could not be imported shows on its own row (under Skipped) with its reason.
Still failing? Note the run time and the item that fails, and see How to reach support.

Connectors that send us change notifications start a sync shortly after a change; the rest sync on their schedule. See Connecting business systems.

Answers say the sources do not cover it

An agent declines when it cannot ground an answer in your sources. It is not guessing; that is the point. Check, in order:

Is the content in a source the agent can see? Open the agent and check its Knowledge step. A source outside the agent's scope is invisible to it.
Is it allowed on this channel? For a chat widget or Help Center, the Note or document must have External Channels set to Public. Content set to Signed-in customers is used only for visitors who are signed in, for example through your gated Help Center. Internal-only content never reaches customers.
Sensitivity. The agent's sensitivity ceiling and the asker's own permissions cap what can be used. See Access, sensitivity, and trust.
Did the source ingest? Open the source and look at its last run for failed or skipped pages.
Is it a policy decline? Some questions (pricing or promotions without a verified source, or things the agent is told not to discuss) are declined on purpose. Review the agent's instructions.

If the question was reasonable and the content exists, add a Note that answers it plainly. Notes are retrieved like any other source.

Answers are slow

Depth and Effort. On the agent's Behavior step, Depth (sources) (Auto, Fast, Standard, Thorough, Research) sets how much it reads, and Effort (Auto, Low, Medium, Max) sets how hard it thinks. Higher settings take longer. Try Standard or Fast for routine questions. In chat, the Effort control in the composer changes this for your own questions.
Council. A council runs every seated model and merges them. Quality goes up, and time and cost go up roughly with the number of seats. Use a council only where it matters.
Answer length. Long answers stream for longer. Set the agent's Answer length to Auto, Short, or Medium.

Open Help Desk → Channels, click Manage on the Chat widget card, and open the widget. On the Install step, expand Advanced — access, identity & privacy and find Allowed origins. Add the site that embeds the widget, one per line: a domain, *.domain, or a full origin. Your organization's own domains already count. On a site that is not allowed, the widget does not appear, and the browser console shows "Records Labs widget failed to load" with "Origin is not allowed for this widget."
Check that the widget is live (Public widget is on for the agent) and that its agent is active.
Check the embed snippet on your page matches the one on the Install step. Rotating the widget token breaks snippets already installed.
After changing settings, wait a minute and reload the page; changes take a moment to reach the widget.

See Website chat widget.

A publish is blocked by the safety review

Every publish is scanned first. A blocking finding stops the publish and lists what was found. Common causes:

A secret or credential in the content (an API key, a token, a password). Remove it. A high-entropy string that only looks like a token warns instead of blocking; confirm it is not a secret.
A password form that submits to another domain. A password field alone only warns: published pages are static and cannot process logins.
A script that reads cookies and sends them elsewhere, or a page that redirects to another domain automatically.
For published apps: an external script (<script src=…>) or an app bundle over 5 MB. Inline event handlers, javascript: links, and eval only warn, because they will not run under the app's security policy.

Fix the content and publish again. See Publishing Notes.