Function

Huggingface models do not have a function calling api

Instead, we have to rely on the LLM to write structured JSON.

We support both openais function calling api, and normal json outputs that we parse, for all major features

Are there some strategies for non openai models to ensure consistent output when using the LLM for data cleaning?

Mostly prompt engineering. Overall open source model are not very reliable in my experience for this haha

We do our best to parse the output, even when it's a little messy.

Usually you'll get best results with larger models (30B+)