Using Pydantic to Get Reliable JSON Output from an LLM in Python

A small shop wants customer emails turned into structured order data for its spreadsheet, but language model outputs can break in several ways. Common failures include stray text around the JSON, wrong types such as writing a quantity as a word, and invented details like order numbers absent from the email. The guide separates format errors, which parsers and schemas catch, from factual errors, which require checking the output against the source text. It recommends a layered approach: clear prompts, schema validation with Pydantic, retries that feed errors back to the model, source checks, and escalation to a human after repeated failures. It also notes that fixed-template inputs may be better served by forms or regular expressions than by a language model.