Dear support team.. I'm running. asurvey and collecting feedback scores. I have a few examles where the customer did not complete the call but the AI passed a wrong / made up score to the SO. callid:019e67ea-f7a9-711a-84f1-3a5ef618f09f Had thi spreviously when I was using OpenAI models.. significantly improved when I changed to Antropic (sonnet 4.x) for evaluation. Now even Anthropic is failing. This feedback is important for customers.. any help is appreciated. Impacted call IDs: 019e67ea-f7a9-711a-84f1-3a5ef618f09f 019e67ea-f7a9-711a-84f1-3a5ef618f09f 019e67df-f938-7cca-9fe8-83e416941b22 https://cdn.discordapp.com/attachments/1511822543386837214/1511822543810728046/Screenshot_2026-06-04_at_12.02.32_AM.png?ex=6a21d9f6&is=6a208876&hm=989a76d7bdd121983ceaef7dbcd1c7ac6567ce0e4d7af486af1086567fe62b52&

Hey Nisar, Thanks for providing the call IDs. I'm currently looking into this issue. once I identify the root cause, I'll get back to you with an update. Kind regards, Akash

Composer advised I tell you it's a production issue.. here is what composer wants you to know: Hello Vapi Support Team, We’re seeing consistent 500 Internal Server Error responses when trying to re-run Structured Outputs via the API endpoint: Endpoint: POST https://api.vapi.ai/structured-output/run Mode: previewEnabled: true (preview-only; no write) This appears to be a platform issue (not schema validation), and it’s a production risk because it blocks reruns/backfills. Repro details Call IDs (both fail): 019e67ea-f7a9-711a-84f1-3a5ef618f09f 019e67ea-8338-711a-84eb-2560e9c77b44 Structured Output IDs (both fail): 5f0a3d27-6bd7-4d4d-b705-e95eb234feb9 (Feedback Score) 508ed0fc-af75-4870-a44c-8437c16f3186 (feedbackProvided) Example request body json Copy { "structuredOutputId" "5f0a3d27-6bd7-4d4d-b705-e95eb234feb9", "callIds" ["019e67ea-f7a9-711a-84f1-3a5ef618f09f"], "previewEnabled" true } What we observe Response: 500 Internal server error Happens consistently across multiple calls and structured outputs We attempted to find related logs via Vapi dashboard/log tooling, but no useful API/call logs were surfaced for the failure. Could you please check internal logs/traces for POST /structured-output/run around the time of these requests and advise on the root cause and fix? Thanks, Nisar

Hi Nisar, I checked the corrected call logs. In 019e67ea-8338-711a-84eb-2560e9c77b44, the call ends without a rating and Feedback Score is correctly null. In the other impacted calls, the customer did not explicitly give a score, but the structured output still inferred one. So this looks like a structured output extraction issue, not a call transport issue. The score field is currently too permissive and needs to be gated so it only fills when the customer explicitly states a rating. Possible Fix: - Make

Feedback Score

depend only on an explicit numeric answer from the customer. - If the customer never says a score, set

Feedback Score

to

null

. - Add a strict rule in the extractor prompt: do not infer or guess a score from context. - Keep

feedbackProvided

or call-complete logic separate from the score field. - Test one finished call and one incomplete call to confirm the extractor returns

null

when no score is spoken. Best regards, Akash