Documentation · Replicate Labs

The product, page by page. Every page is also served as Markdown, so the same words reach the models a team works in.

Reference

Troubleshooting

The failures that produce no error message, in the order they actually happen, with the check that identifies each.

Most problems on the platform fail quietly rather than loudly. This page is organised by symptom, and every entry is something that will not show you an error.

#A call will not start

Microphone permission. The most common cause by a distance. The browser blocks access, the call never properly starts, and nothing says so.

Check the permission indicator in the address bar and allow the microphone for the site, then reload. If your organisation manages browser policy centrally, permission may be denied at that level and needs IT rather than you.

Also worth checking: whether another application is holding the microphone, and whether you are on a browser your organisation has locked down for media.

#A button does nothing when clicked

The skill it invokes is not loaded for you. A button that runs a skill fails silently if the skill is not available to your user.

You cannot fix this from the rep surface. Report it with the button label, the coach you were on, and your role. Those three facts are what an administrator needs; without them the diagnosis takes days. See Coach configuration.

#The buttons are wrong, or a button has disappeared

You are on a different coach than you think. The platform remembers your last coach selection between sessions, so you can return days later on a coach you switched to once.

Check the coach selector at the top left first. Buttons belong to a coach and to your active role, so a manager looking at the rep coach correctly sees rep buttons.

#The Training Lab looks wrong or empty

Check which coach is active first. The coach selector is at the top left, and the platform remembers your last selection between sessions, so you can return days later on a coach you switched to once. The scenarios and personas listed are the ones configured for that coach and account.

If the list is genuinely empty, nothing has been published to you yet, which is a configuration question rather than a fault. See Building roleplays.

#The coach does not know something about our business

Distinguish two cases, because they go to different people.

  • Wrong about your deal. Correct it in the conversation. It carries forward. This is yours to fix.
  • Wrong about your products, pricing or competitors. A knowledge base gap. This belongs with your administrator.

#A document was uploaded but the coach cannot use it

Four causes, in order of likelihood.

  1. It has not finished indexing. Upload succeeding is not the same as being ready. Wait for the ready status.
  2. It is a scanned PDF. No text layer means nothing to index. It will look uploaded and contribute nothing.
  3. It is an image-heavy deck. Extract the text to Markdown and upload that instead.
  4. You tested it badly. Asking "do you know about our pricing?" gets an accommodating yes. Ask something only answerable from the document and check the specifics.

#Answers contradict each other between sessions

Two documents in the knowledge base disagree. Re-uploading does not replace, it adds, so an old price list and a new one can both be retrievable.

Find the superseded document and delete it. This is the hardest problem on this page to diagnose from the symptom, so check it early whenever answers are inconsistent rather than merely wrong.

#A manager's team view is empty

The team has no members. An empty team reports nothing and receives nothing published to it, and neither state produces an error. Check membership before anything else.

#A rep cannot be added

The seat limit. Nothing else will explain it. An administrator can see the limit on the account.

#Roleplay scores dropped across the team

Before concluding capability has fallen, rule out three things.

  1. The scorecard was edited. Grades before and after are against different standards and are not comparable.
  2. A new roleplay was deployed. A different scenario grades differently.
  3. People ran short sessions. A four minute conversation produces a score computed from four minutes of evidence.

#Still stuck

Include these when you report anything: what you clicked, which coach was active, your role, and what you expected instead. Contact us.

Last updated 3 September 2026

View as Markdown