How do I handle failed LLM JSON parsing using the Instructor library?

All answers to this question.

The beauty of this library is that it handles the "re-asking" logic for you. When you define a Pydantic model and pass it to the response_model parameter, the library validates the LLM's response against that schema. If validation fails—say a required string is an integer instead—it sends the error message back to the LLM and asks it to correct the mistake. You can set the max_retries parameter to control how many times it tries. In my experience, even complex nested objects usually get corrected within one or two retries, which is much better than writing custom regex or try-except blocks.

   Answered 2025-02-13 by Kimberly Roberts

Does this retry logic also work for custom Pydantic validators, or is it strictly for basic type checking like strings and integers?

   Answered 2025-02-14 by Jeffrey Davis

• Jeffrey, it works perfectly with custom validators! If your @field_validator raises a ValueError, the library captures that specific error string and passes it back to the model as feedback. This allows the LLM to understand exactly why its previous answer was rejected, making the second attempt much more likely to succeed.

     Commented 2025-02-15 by Brenda Miller

Using this has reduced our "JSON Hell" bugs to almost zero. It’s a must-have for production-grade AI.

   Answered 2025-02-16 by Michael Garcia

• Totally agree, Michael. The fact that it’s just a lightweight patch over the existing SDK makes it so easy to adopt without refactoring everything.

     Commented 2025-02-17 by Kimberly Roberts