LLMs Struggle with Complex Structured Outputs; Developer Shares 'Researcher-Formatter' Split Architecture

DiligentBend9150 · reddit · 2026-08-16

A developer building with LangGraph and Pydantic structured outputs found LLMs often get stuck or hallucinate with complex schemas. He adopted a 'researcher-formatter' architecture: the researcher node produces prose, the formatter node converts it to the required schema, separating responsibilities for easier debugging and retries. He asks about drawbacks, schema design best practices, and considers using an in-memory graph instead of large nested Pydantic schemas.

Original post →

More from coding & agent

coding & agent channel →