Preserve newlines and other formatting about dalai

[rany2]

Fantastic, thank you. That's a fast pace!

[bernatvadell]

fixed in PR #39

[rany2]

I don't see anywhere in your change where you escape shell input. I'd much prefer if the prompt was in a file instead

[bernatvadell]

after the change in textarea and the responses encapsulate them in "

" I am not having problems with the escaped characters:

[rany2]

@bernatvadell There is no way your change made a difference. Try with a single-quote to get it to obviously not work. If you end up closing the quote it does work, sure but try to put anything special like $, `, ", <, etc... it does NOT work and gets passed as is to bash.

[rany2]

If you need a description, $(ls) ended up just running and the output of ls got sent to the model. This is not what should happen under any circumstances...

[rany2]

IMHO the most proper way to get around this is to put the prompt in some temp file and have the model read from it (using LLAMA's -f). I recommend against trying to escape because there will always be edge cases you will miss and it might be a hit or miss across different shells, OSes, etc... putting it in a file is a universal fix so to speak

[bernatvadell]

I think we should use the "args" to be able to correctly escape the input.

[rany2]

@bernatvadell Yes that would be ideal, no idea why it is spawning in a shell...

[bernatvadell]

I imagine it would be possible to import the llama.cpp project as a library, I haven't looked into it, and right now, as far as I know, it's published as a console program.

[bernatvadell]

Finally, I have needed to escape the characters, the ones that I have verified that give problems are these:

This seems to work correctly:

The change is in the TypeScript PR #44

[rany2]

Why do you escape when you could just call exec directly? You don't need to exec it in a shell ..

[bernatvadell]

Why do you escape when you could just call exec directly? You don't need to exec it in a shell ..

Originally it was like this, I'm going to try what you say.

[rany2]

Anyway this proposal makes me a bit uneasy because in different shells you need to escape different things. For example if user was using fish, then (ls) would execute also. Best solution IMO is to just put the prompt in a file or use whatever the programming language offers to execute a program without relying on a shell (while allowing you to define arguments as an array of strings of course)

Edit: also if this was a Windows user then none of your escaping would be relevant.

[bernatvadell]

Atm we're forcing to use powershell or bash:

[rany2]

Is executing it in a shell still needed? I think you could call node's spawn directly now.

[bernatvadell]

it works quite well.

Now directly spawns the process, the arguments do not need to be sanitized manually.

the prompt is stored in a temporary file.

You can change default directory to store temporal prompts:

   this.tempPromptsPath =
      process.env.TEMP_PROMPTS_PATH ?? path.resolve(os.tmpdir(), "prompt-tmp");