AI runs can fail before a model call, during provider execution, or while saving the result. The visible error identifies which layer to check.
Quota and rate limits
Evane applies space limits for concurrent runs, hourly runs, and monthly tokens. Provider services can impose their own rate or credit limits as well. When a retry time is shown, wait for it; clearing browser data does not reset server-side usage.
Model availability
Confirm the selected model is enabled for the configured provider and deployment. A model name can be valid in the UI but unavailable in the current Azure/OpenAI/Anthropic account or region. An Admin should use the provider probe or model settings rather than exposing credentials in chat.
A reply stops mid-answer
A dropped connection to the model (a proxy reset, a gateway error) is retried automatically a few times before it ever reaches you. If the retries do not succeed, Evane ends the run and names the specific cause instead of leaving an empty bubble: a dropped connection, an internal error, a conversation too long for the model to read at once, the step limit for one reply, the space's usage limit, or a model that is not available. Send your message again for a dropped connection or an internal error; the other causes need a shorter request, a new chat, or an administrator's help.
Empty or incomplete answers
A run can exhaust its output budget during reasoning and return little visible text. Evane should report truncation rather than treating that as a valid empty result. Retry with a narrower request, or ask an Admin to review model-aware token limits.
Preserve evidence
Keep the run ID, selected model, time, error name, and whether the run appears in conversation history. Never paste API keys or provider tokens into a support request.
