Troubleshooting
Resolve sign-in, source, administration, AI Agent, and MCP connection issues.
Get Answers and sign-in
| What you see | What to try |
|---|---|
| No relevant information found | Check the active agent and ask a focused question using source terminology. |
| No source chips | This response may not contain source files. Check More details and the agent’s capabilities. |
| Agent selector disabled | Your administrator may manage the selection. Check assigned access with them. |
| Missing Nuggets or agent header | These can be switched off for the active agent. |
| Microphone disabled | Browser speech recognition may be unavailable. Type instead. |
| File will not open | Reopen it after checking your session; ask support to check that the file still exists and is accessible. |
| Video/audio starts at the beginning | The result may lack a usable timestamp, or seeking may have failed while loading. Use the player or summary navigation. |
| Recording will not play | Check controls/volume and browser codec support. |
| SCORM opens a start screen | Allow launch to complete; some packages require Start/Resume. Exact targets depend on course/index compatibility. |
| Summary or Nugget is still processing | Allow generation to finish; report persistent failures with the source name and question. |
| Connected tool error | Record the displayed tool/service name and ask support to check the connection. |
| Magic Link invalid or expired | Request a new link; the old link may have been consumed or expired. |
Administration and data
| What you see | What to try |
|---|---|
| An administration option is missing | Available controls depend on your role and deployment. Ask your administrator to confirm your assigned access. |
| An optional import source is missing | SharePoint, Google Drive, GCS, and custom sources depend on instance configuration. Check which import sources your organisation has enabled. |
| An import fails or appears stuck | Check Active imports or Completed / Failed for the job’s details and error. Confirm the source connection and destination dataset before retrying; allow large uploads to finish transferring. |
| An imported file is absent from answers | Check that processing completed, that the file is published in File Management, and that the selected agent uses that dataset. |
| A user cannot see an AI Agent | Check the agent’s active state and assigned user access in Manage AI Agents. Signing in does not automatically grant access to every agent. |
| Feedback or logs seem empty | Check the selected agent, date range, and filters. Use feedback and logs to locate the relevant event. |
Super admin and custom AI Agents
| What you see | What to try |
|---|---|
| MCP Tool Store is missing | It is optional and depends on deployment and account access. Confirm availability before trying to configure a server. |
| MCP Registry search has no useful results | Check whether you used Local filter or Search MCP Registry. Try a provider or service name; only compatible remote servers can be imported. See MCP Tool Store. |
| OAuth is connected but an agent cannot use a tool | Check server configuration, Active state, enabled tools, and the agent’s MCP Tools selection. Save the assignment with Save MCP Tools. |
| A tool returns a provider permission error | The connected provider account may lack access to the requested resource. Review the error and account permissions; reconnect OAuth if authorisation has expired. |
| Chat history or reports are empty | Select the correct agent and date range, clear restrictive filters, and retrieve chat logs. Reports need conversations to analyse. See View Chat History and reports. |
| A report is not ready | Check its status in View Reports. Avoid generating duplicates while processing; use View Logs to investigate a persistent failure. |
| An embedded agent will not load on a website | Check the widget’s agent configuration and Allowed origins against the website address. See embedded AI Agents. |
| A custom CTA or header button does not appear | Check that the relevant control is enabled and saved, with its label and destination set. Keyword-triggered CTAs and the Embedded Chat Demo Button are configured separately. See embedded AI Agents. |
Connected assistants and MCP
| What you see | What to try |
|---|---|
| Connection fails or returns unauthorised | Confirm the HTTPS MCP access URL and the key created for the intended user and agent. Reauthorise if needed; a Slack/Teams app key is not an MCP key. See MCP access. |
| A connected assistant cannot find an expected document | Check the dataset bound to the key. Content in a different agent or Talvi instance is not automatically available to this connection. |
| A source viewer expires or will not load | Request the source again or use Reload. Source links expire; some clients offer native content instead of an interactive player. |
| A long request times out | Check whether the operation already completed before retrying, especially for Nugget creation. Check rate limits, then review the client’s timeout and Talvi logs. |
Rate limits and timeouts
If responses are slow, requests time out, or you see rate limit exceeded or 429 Too Many Requests, check whether usage is reaching a configured limit. A timeout alone does not confirm rate limiting.
- Ask an administrator to open Read dashboard metrics. On Talvi AI, check the AI agent request cards for today, the last hour, and the last minute. For a connected assistant, also check Talvi AI MCP where available.
- Compare usage with each card’s capacity. A short burst can reach a minute limit even when daily usage is low. These are current counters: selecting Last month does not show the limit usage at the time of an earlier failure.
- Review the relevant logs around the failure time. Look for AI agent rate limit exceeded, MCP rate limit exceeded, or a provider’s own rate-limit response. Note the affected agent, operation, and time. Dashboard totals may not reveal a limit applied to an individual caller or by an external provider.
- Wait for the stated retry interval before trying again. Integrations should honour Retry-After or retry_after_seconds when supplied and space out retries. Avoid repeatedly clicking Send or launching the same generation job.
- If the limit is reached regularly, ask your administrator to review request volume, overlapping jobs, and the capacity configured for your deployment. Stagger bulk work where possible.
Talvi AI requests and MCP requests have separate checks; an MCP operation that invokes the LLM can encounter either. Embedded bots and external providers can also have their own limits. Increasing a client timeout does not remove a rate limit.
If there is no rate-limit evidence, investigate other causes such as a long-running generation task, an unavailable connected service, or a client/network timeout. Check whether the operation completed before retrying, especially when creating Nuggets or reports. Include the error and approximate time in your support report, but never include access keys.
Send a useful report
Use Share feedback with the agent name, question, source name, and expected result. Include a relevant screenshot if useful, without credentials or unrelated personal data. Avoid repeatedly submitting a slow request, as that can create duplicate work.
For more detail, see sources, file types, SCORM, or Account.
