QaG

Quirks and Gotchas

Despite being similar in many ways, all LLMs have their quirks and gotchas. We tried to work around some of them. For the others we provided tools so that you can work around them. Some examples we learned so far:

• Gemini and ChatGPT require all types in JSON Schema to be explicitly defined Solution: never forget to put type: object in objects, despite most engines being permissive about it
• ChatGPT requires all JSON Schema fields to be explicitly required Solution: never forget to put all the fields in the required array. Pretty dumb, I know.