Skip to content

VERCEL_LOGS.md

Latest commit

 

History

History
60 lines (35 loc) · 3.31 KB

VERCEL_LOGS.md

File metadata and controls

60 lines (35 loc) · 3.31 KB

Viewing Vercel logs and fixing “AI not responding”

How to see runtime logs on Vercel

The build output you see in the dashboard is build logs. To see why the AI isn’t responding you need runtime (function) logs:

  1. Open Vercel Dashboard → your Relay project.
  2. Go to the Logs tab (or Deployments → click a deployment → Functions / Logs).
  3. Use Runtime Logs (or Function Logs) and filter by time or by path (e.g. /api/briefing).
  4. Trigger the flow that uses the AI (e.g. open Briefing, or generate a draft in Actions), then refresh the logs.

You’ll see console.error and console.warn from your API routes and server code there. Look for lines like:

  • Briefing API error: ...
  • OpenAI request failed (401): ... or similar
  • OpenAI briefing prioritization failed: ...

Why the “AI model” might not be responding

1. OPENAI_API_KEY not set on Vercel (most common)

If OPENAI_API_KEY is missing in the Vercel project, the app uses deterministic fallback (no real AI). Briefing and drafts still work, but without model-generated text.

Fix: Project → Settings → Environment Variables → add OPENAI_API_KEY with your OpenAI API key, then redeploy.

2. Wrong or unsupported model name

The app uses the OpenAI Responses API and reads the model from env:

  • OPENAI_REASONING_MODEL (default in code: gpt-4o-mini)
  • OPENAI_HEAVY_REASONING_MODEL (default: gpt-4o-mini)

If your OpenAI account or region doesn’t support these, the API can return an error. Check runtime logs for messages like OpenAI request failed (400): ... or model not found.

Fix: Set env to a model your account supports (e.g. gpt-4o-mini, gpt-4o). See OpenAI API docs for current model IDs.

3. Response format mismatch (usage goes up but you see fallback)

If your usage goes up but the app still shows fallback text, the API is returning 200 but the response shape may not match what we parse. In Runtime Logs look for:

  • [OpenAI] no text output in response; top-level keys: ...

The "top-level keys" line shows what the API actually returned. Share that with your dev or open an issue so the parser can be updated. The code already tries several shapes (output, output_items, choices); if your model returns a different structure, we need to add it.

4. Timeouts or rate limits

Vercel serverless functions have a timeout (e.g. 10s on Hobby). If the model is slow or you hit rate limits, the request can fail. Errors will appear in runtime logs.

Fix: Upgrade plan for longer timeouts if needed; fix rate limits or retry with backoff.

5. Briefing / draft errors

If the briefing API or the draft flow throws, the route logs Briefing API error: or the OpenAI service logs the failure. Use the Logs tab as above to see the exact error.

Checklist

  • OPENAI_API_KEY is set in Vercel (Settings → Environment Variables) for the right environment (Production/Preview).
  • Redeploy after changing env vars (new deploys pick up new variables).
  • Runtime Logs (not only build logs) are checked after reproducing the issue.
  • Optional: set OPENAI_REASONING_MODEL / OPENAI_HEAVY_REASONING_MODEL if you need a specific model.