Comment by aetherspawn

Surprised models still output tools as text when for ages we’ve been able to constrain the output at the inference engine level and constrain the model what tools, parameters etc are available

Edit: found it, it’s called Grammar-Constrained Decoding (GCD)