search

simonw / llm

Access large language models from the command-line
https://llm.datasette.io
Apache License 2.0
3.93k stars 210 forks source link

Change default CLI command from `prompt` to `--help` #529

Open tutacat opened 1 month ago

tutacat commented 1 month ago

This is often not great user interface for a CLI. That feels kind of like if rm did not run in interactive mode by default on most computers. llm feels much more like git, or a package manager. Rather than be like a one-task tool, such as an interpreter like python, or a compiler.

This really comes to a head if you managed to misspell a sub-command for LLM because it will try to run that as an LLM inference, rather than show an error for a mistyped command. This is also the case if somehow you managed to uninstall a plug-in without remembering, because there are not registered keywords unless they are installed. This is also the case if you redirect a pipe into LLM with no arguments.

It feels much better having to distinctly write llm prompt for it to run generative text, because llm is used for so many other tools.

PR is here if you are interested (it would be a very simple change).

simonw commented 1 month ago

Posted in that PR:

It's important that echo "prompt" | llm continues to work, so I'm not going to accept this change. But... I do agree that having llm on its own hang is confusing, so I'll look into fixing this in a way that detects piped input instead.

tutacat commented 19 hours ago

PR just meant as a draft. The point is it is more robust and simple to change it to run help than implement detection. This is not difficult though, but requires changing the flow of the CLI more.

I think a way to do this is adding a new "detect" function as default to check for non-interactive stdin. Alternatively you could also allow interactive stdin with an info message, though that usecase is better served by the 'chat" TUI.