Fix AI Output Format Bug

I’ve deployed an AI feature on Replit, but the model’s responses jump between the right structure and a completely different, unexpected format. The logic itself is sound—the issue sits somewhere between the inference call and the response-handling layer.

Here’s what I need:
• Pinpoint why the output format becomes inconsistent.
• Adjust the relevant code or API usage so every response follows the expected schema.
• Add a lightweight test (unit or simple script) that confirms the fix and guards against regression.

The project is already live, so I’ll grant you Replit access as soon as we begin. Keep changes tightly scoped and documented; once the tests pass on my end, we’re done.