I cooked, idk what to say

[thmsmlr]

I'm not going to lie, I coded locally for seven days, and this is the result.

I'm going to try my best to describe the changes, but, c'est la vie.

• I improved streaming reliability. Previously, sometimes the chunks weren’t full lines, which would break the parsing. That issue is now fixed.
• I refactored the internal API between Instructor and the adapters so that it is the adapter's responsibility to parse the JSON. Instructor seamlessly works with that. This change makes it considerably easier to implement any OpenAI-like adapter. For non-OpenAI style adapters like Google Gemini, some work is required to convert the OpenAI message format into the Gemini API format. However, you do not need to map the responses.
• I added the Anthropic and Gemini adapters. I also changed the Llama CPP adapter to use its built-in OpenAI compliant server.
• I also refactored the tests a bit to make it easier to validate new adapters as you write them. I hope to eventually write some stronger evaluations, but this should work for now.