I finally had Llama.cpp serve an LLM with tools enabled, it was just a matter of convincing it that it had tools… What is it with these models not trusting the user?
It was unable to get the date right since it was in a different language but that was easily fixed by instructing it to use date --iso-8601 to get a YYYY-MM-DD format, no need to translate.
By the way you can use ./bin/llama-server -m MODEL.gguf --tools exec_shell_command or, if that’s not enough, ./bin/llama-server -m MODEL.gguf --tools all. If you prefer to keep it in the Terminal just replace llama-server with llama-cli.
It insists that it can’t use python despite being able to run shell commands…
That’s better, note that only the month was wrong, these pictures were taken in the 3rd and 4th day perspectively.
If you prefer, use the tool that only gets the date/time: get_datetime.