[🐛 Bug]: Create Incidents With AI Function Fails with Connection Error(Reopened After Accidental Closure)

[kc945752357]

issue:#5428
This issue is reopened after being accidentally closed due to a misclick—the original problem persists: the "Create Incidents With AI" function consistently fails with a connection error when calling the /chat/completions interface, while all other AI features in the system work normally.

Thanks for the guided troubleshooting! Just to confirm—we’ve already verified the key environment variables are properly configured (screenshot attached for reference), but the issue persists.
The OPENAI_BASE_URL includes the /v1 suffix, and OPENAI_API_KEY, LITELLM_SALT_KEY, and LITELLM_VIRTUAL_KEY are consistent with other working AI modules across frontend and backend containers.
We’ll now focus on comparing the payload structure/size between the failing Create Incidents With AI function and working features, as well as testing the exact payload via curl. We’ll also check LiteLLM logs for payload-specific errors that might not show up in basic connectivity tests.
Let us know if there’s any additional context we should provide for further debugging!
（Note: Screenshot of the configuration file is attached, showing the correctly set OPENAI_BASE_URL, OPENAI_API_KEY,related variables.）

[dosubot]

Your environment variables and endpoint configuration look correct, and direct curl tests confirm connectivity, so the issue is likely with the payload or request handling for the "Create Incidents With AI" feature. Both the OpenAI and LiteLLM providers in Keep construct the payload for /chat/completions as a list of messages with roles and content, typically starting with a user role and prompt. If a structured output format is used, a system message with JSON schema instructions is inserted at the start of the messages array. The request is sent as a POST with JSON payload and appropriate headers, and any connection errors are surfaced as exceptions in the provider logic see implementation.

Since other AI features work, focus on comparing the exact payload (including message structure, size, and any system instructions) sent by the failing function versus working ones. Also, check LiteLLM logs for errors tied to specific payloads, as issues like oversized requests, malformed JSON, or missing required fields could trigger connection failures even when basic connectivity is fine reference. If you spot differences, try sending the failing payload via curl to reproduce the error outside Keep—this can help isolate whether the problem is in Keep's request construction or LiteLLM's handling.

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}