When the AI provider rejects a request — invalid API key, exhausted quota or billing limit, model unavailable — the chat shows an error message instead of a reply. Administrators see the provider's detailed error while other visitors get a generic message, so log in as an admin and repeat the question to see the real cause. (Provider errors can include part of your API key, which is why visitors never see them.)
To check a key without asking a question, open Wyregraf → Options → Diagnostics and click Test connection beside the profile. A rejected key fails there straight away; a quota or billing problem usually passes, because the key itself is still valid — check the provider's dashboard for those.
The two common ones:
- Invalid or revoked API key — re-paste the key into the profile (embeddings and LLM profiles each hold their own key; check the one the failing feature uses).
- Quota or billing limits — check your usage and billing status on the provider's dashboard. A free-tier OpenAI key without billing enabled will fail once its trial allowance is gone.
The same provider errors will also fail an indexing run — if chat errors and a stuck queue appear together, fix the key or quota first.
