zeke/tool-calling

Teaching workspace: how LLMs invoke tools, from autoregressive decoding to constrained decoding and why some models are better at it