How to Build a Bounded JSON Repair Loop for Unreliable LLM Output in Python

Large language models do not reliably return valid JSON, and failures can fall into three distinct categories: syntax errors, wrong structure, and semantic violations. A developer tutorial on DEV Community demonstrates how to classify these failures using only Python's standard library, with a validator function that checks each layer in sequence. The guide recommends a bounded repair loop capped at a small number of attempts, where the model is sent its original output along with the precise error and required schema — without expanding conversation history. Logging attempt counts, error classes, and final outcomes is advised, while sensitive prompt data should be excluded from logs by default. The core takeaway is that treating JSON validity as a measurable, testable rate — rather than an assumption — leads to more reliable LLM integrations.