How can we move forward with replacing the out-of-date man pages?

[Potherca]

I think that the majority of the issues here are caused by the fact that the man pages are out of date.

@idank You've stated in other issues that you'd like to eventually change the input from man pages to something else.

Is there a plan somewhere? Is this something you'd be willing to accept support with?

I'd love to get things to a place where it would be easier to keep the underlying data updated!

On a side-note, I'm crazy enough to go through all the issues and provide you with a list of issues that are caused by out-of-date man pages, if you'd want it.

[idank]

Hey there,

Thank you for your interest.

The entire pipeline that imports and analyzes man pages needs to be revamped. It uses an old tool to convert the raw manpage to html, and then uses a pretty naive classifier to extract the sections with command arguments. This last part might be entirely replaceable with an LLM for better results, one would hope.

And then if you wanted to keep them fresh, a mechanism would need to be in place to periodically do this for newer distro releases.

I'm curious though, with LLMs and their cli/ide integrations, haven't they replaced the need to use a service like Explainshell?

[Potherca]

One of the main advantages of the ExplainShell website over AI, is that it provides a simple way to share a URL that explains a command (for instance, to link to from documentation). AI would not replace that IMHO.

Also, having a static dataset that is periodically updated seems more efficient than needing CPU cycles every time someone wants a command explained to them...

Although an LLM might be a good tool to help with triaging GH issues, to clean things up a bit. 🤔

[idank]

Those are nice to haves, sure. But I'm more curious if the data that LLMs have these days cover people's need, and whether they're capable of dissecting complicated syntax.

[Potherca]

I thought I'd just ask ChatGPT, according to its response (with references linked), the largest issue is that LLMs are not reliable parsers and might mix up OS (or version) specific params.

It also mentions hallucinations (which I think could be worked around using error correction prompts), and not having the latest version (which could also be worked around by explicitly specifying source data). Especially if an MCP server was made available that could be used by the AI as data source.

Might be worth running an experiment for...