Structured output : Pydantic-AI vs Instructor.

Hello everyone,

Pydantc-AI is shaping up to be a beautiful framework for agentic development yet I feel there should be some clarification on how it handles structured output, specifically compared to the amazing Instructor library. From what I understand the difference is as follows:

### Instructor

Uses OpenAI's function calling API for structured outputs
Enforces schema *at the API level*
Automatic retries for invalid responses
Real-time streaming validation
Potentially fewer retries and more reliable results

### Pydantic-AI

Uses tool calling and Pydantic for schema verification
*Post-response* validation
More flexible, supports most LLM providers
Offers additional features like dependency injection and cost tracking
May require more retries for invalid outputs
Suitable for complex AI applications with broader model support

The main trade-off seems to be between Instructor's more reliable output structuring (especially with OpenAI) and Pydantic-AI's greater flexibility and additional features for diverse AI applications.

What do you think, is Pydantic-AI's method of post-response retry/validation sufficient compared to Instructor's or should we wait until more LLM providers provide structured outputs in their API ?

Cheers 💫

2 comments