I think the function calling can just be a boolean: supports_tools. Inside the ai_backend's we need to make sure that a really long conversation starts skipping older messages so that the submitted payload never exceeds the context size. Maybe this attribute is simply context_window, that's what OpenAI and Anthropic both seems to call it.
I think the function calling can just be a boolean:
supports_tools
. Inside the ai_backend's we need to make sure that a really long conversation starts skipping older messages so that the submitted payload never exceeds the context size. Maybe this attribute is simplycontext_window
, that's what OpenAI and Anthropic both seems to call it.