JSON or XML Tags for LLM Output: The Format That Holds Under Pressure

A common issue with LLM output is that it's wrapped in text that's difficult to parse. To solve this, use XML tags as a delimiter to mark the start and end of the data, and JSON as the data format inside the tags. This allows you to use a regex to extract the data and then parse it as JSON. If you control the decoding, you can skip the wrapper and use a structured-output or JSON mode that constrains the generation to a schema you supply.