LLMs return text, but applications need data. Parsing JSON from model output is fragile, and one missing field or extra comma can break everything after it.

This agent checks every response against a Pydantic schema. If the check fails, it retries and tells the model what was wrong. It also logs the failure, so bad data never slips through quietly.

Only valid, typed objects reach the rest of your code. No manual JSON parsing and no try/catch at every call.

When validation fails, the model is asked again with the error included, which makes the second try much more likely to work.

Failures are logged with the raw model output, the validation errors and the retry history. This helps when prompts drift or a schema needs tuning.

Agents, tool calls and RAG pipelines all depend on data in a known shape, so this is the base they build on.