I’ve been following this thread and can confirm that I'm seeing similar behavior. The generic UnknownError is indeed a major roadblock for debugging, especially when working with complex responseConstraint implementations.
Regarding the Regex issue mentioned:
It seems the current parser for responseConstraint might be struggling with specific engine-level validations before even hitting the model. Even a catch-all regex like /^.*$/ failing suggests that the validation layer itself might be crashing or timing out before the GPU backend can process the prompt.
To help the team, I’ll perform a few more tests on my end with:
Different backend types (switching between GPU and CPU if possible).
Testing if the same failure occurs with very short prompts vs longer ones.
@Tom, as a developer with long-term experience in complex system architectures, I'd suggest that even if the root cause is a GPU memory/timeout issue, the API should ideally catch these specific failure modes and map them to more descriptive error codes (e.g., REGEXP_VALIDATION_FAILED or INSUFFICIENT_GPU_RESOURCES) rather than the catch-all UnknownError.
I'll collect my logs and attach them to a new bug report as suggested.
Best regards,
Felipe Aquino
Senior Software Developer