LangChain Structured Output & HITL

To get structured output from LLM responses, you define typed schemas using Pydantic or Zod to validate and extract data. This transforms unstructured text into typed objects, which is essential for reliable form filling and API integrations.

Human-in-the-loop (HITL) works by pausing agent execution before critical actions through middleware. It allows humans to review, approve, reject, or edit tool calls, ensuring oversight for tasks like sending emails or modifying sensitive data.

You parse unstructured text into typed data by defining a schema that maps the required fields. The LLM generates content matching this structure, which is then validated into typed objects for immediate use in data extraction workflows.

Structured output fails with missing field descriptions because LLMs need explicit context to map text to schema fields accurately. Adding clear descriptions to your Pydantic or Zod schemas ensures the model understands what data to extract.

You can implement approval workflows for critical actions in LangChain agents by integrating HITL middleware. This pauses execution, allowing human reviewers to approve, reject, or edit tool calls before actions like sending emails are executed.

Incorrect response access often stems from Pydantic version compatibility issues or improper model integration. Ensuring your schema definitions align with your specific Pydantic version and correctly accessing the validated response object resolves this.