I was playing with it and was initially very impressed, but then progressively less so as I tried to add in some tool use and build some conversation flows.
It has a lot of fixed prompts that seem to work pretty well with GPT-4, but i've been more interested in llama, RWKV, and those minstral models and I haven't been able to get good autogen performance using those models.
The function calling system seems to be pretty hard tied to openai's api as well (you pretty much write the json the api expects as a python dict.)
Have you seen a blog comparing llamaindex with Microsoft approach? Would be very interesting to compare the results on a standard dataset with some standard metric (like ragas)
