We should allow advanced usage such as:
- all kinds of chaining (CoT, ToT, etc.)
- scripting (clean output, clean status, signals handling, etc.)
- re-compiling model with different parameters ?
Please suggest any advanced usage you may think of
Chaining should be as simple as:
$ llm -q -m modelname "prompt" | llm -m modelname2
# output of modelname is used as prompt for modelname2
# -q is used to remove pretty-printing and only output the model's output
$ llm --output prompt,output -m modelname "prompt" | llm -m modelname2
# output both prompt and output to pipe