Troubleshooting
The problems that come up most, each with its likely cause and what to do.
The chat button does not appear
Open your browser's developer tools (F12) and look at the Console tab.
| What you find | Cause | What to do |
|---|---|---|
[ChatterLab] missing data-agent | The tag has no agent id | Copy the snippet again from Deploy |
Nothing about ChatterLab, and w.js is not under Network | The tag is not on this page | Check where your site builder puts "footer code". Some only add it after you publish. |
w.js is blocked, with a message about Content Security Policy | Your site restricts which scripts may load | Allow our address in your policy. Install lists where. |
w.js loads and there is still no button | Something of yours covers the corner, or an ad blocker hides chat widgets | Try a private window without extensions. Move the button with data-side="left". |
The button appears, but the chat stays empty or refuses to load
Nearly always Allowed domains. Once the list has an entry, the chat only opens on the sites listed, and the match is exact: https://example.com does not cover https://www.example.com. Add the exact address from your browser's address bar, without the path, or *.example.com for every subdomain, and wait half a minute. Allowed domains.
The same cause explains "it works on the live site and not on staging", and the other way round.
Visitors see an error
The widget says one of three things, and which one narrows it down.
"I can't answer right now. Please try again later." Something on your side stops the agent, and the message deliberately does not say what, since that is not your visitor's business. In order of likelihood:
- The credits are used up. Look under Settings, Usage. What to do.
- The agent is paused, because the workspace has more agents than the plan includes. The agent's page says so at the top.
- The agent's model is no longer in your plan, after a downgrade to Free. Choose a 1-credit model under Settings.
"That was a lot of messages at once." The visitor passed the limit of 20 messages per 60 seconds. It passes by itself, and nothing needs fixing.
"Something went wrong." Everything else: a disturbance at the model provider, a connection that dropped, an address the widget may not be embedded on. Try again a little later, and write to us if it lasts.
The Playground says plainly when the credits are used up, so ask your question there first.
The agent says "I don't know" about something that is in my sources
- Check that the source says Ready, not Pending, Processing or Failed.
- Open the conversation in Activity and look at the sources under the answer. If yours is not among them, it was not found: see step 3 of Fixing weak answers.
- A source on Failed teaches the agent nothing. A scanned PDF is the usual case: it holds pictures of text, not text. Files.
- If you raised Minimum similarity, put it back to 0.25.
The agent gives an old answer
The agent knows each page as it was when it was read. If the page has changed since, crawl the site again or add the page again as an Individual link: it is updated, not added twice. If the old information is on a page that has gone from your site, or in a file, delete that source, because a crawl never removes anything. Limits and freshness.
The agent says something that is not in my sources
Look the answer up on the Overview under Worth adding to your sources. If it carries Partly backed or Not backed, follow step 4 of Fixing weak answers: lower the temperature, tighten the instructions, fill the gap in your sources.
A crawl found few pages, or none
Websites has a table of causes. The three that matter most: the site builds its content with JavaScript, robots.txt or a firewall stops the crawler, or the pages live on another host such as www. or shop..
A crawl seems stuck
A crawl works in short turns, and picks itself up after an interruption. If the Sources page shows Crawl paused, press Resume crawl. Nothing already read is lost. It stops by itself at the end of the site or at your storage limit, and Stop crawl stops it sooner, keeping the pages it has read.
A source ends on "Failed"
The Sources page says why, under the source. A file that could not be read usually has a password on it: remove the protection and upload it again. No text was found means a scan: run it through OCR first, or paste the content as a text source. Too long applies to a question-and-answer pair: split it into shorter ones. A source that was started several times and never finished is too large to process in one go: add it in smaller parts. If it says the fault was on our side, press Retrain agent after a few minutes, and write to us if it fails again.
A source is refused because of storage
The agent has reached its plan's storage. Delete sources it does not need, exclude paths when crawling, or move up a plan. Limits and freshness.
My logo is refused or looks wrong
- Over 10 MB, or not a PNG, JPG, WebP, GIF or SVG: convert it first.
- Tiny or blurry in the chat: the header space is small and square. Use your mark instead of your full wordmark, and check the true-size preview in the crop window. Appearance.
Changes to the look do not show on my site
Press Save first: the preview shows unsaved changes, your site does not. After that, browsers keep the button's settings for up to five minutes. A hard reload (Ctrl+F5) shows the new version at once.
Monthly reports do not arrive
See the end of The monthly report.
The API returns an error
Every error has a code. API overview lists them all. The two most common: 400 invalid_request with a field in details, which is usually a misspelt or unknown field since bodies are strict, and 403 forbidden, which means the key's scope does not cover the call.
Still stuck?
Write to management@sri-group.eu. Include the agent's name, what you did, what you expected and what happened. For an API problem, add the x-request-id of the response.