Comment by WillAdams | Hacker News Reader

Comment by WillAdams | Hacker News Reader

The big feature which I would like to see is a way to easily interact with the content of the local filesystem --- I have a prompt for re-naming scans based on parsing their content which I've been using in Copilot --- recent changes require that I:

- launch Copilot

- enter a prompt to get it into Copilot Pages mode

- click a button to actually get into that mode

- paste in the prompt

- drag in 20 files

- wait for them to upload

- click the button to process the prompt on the uploaded files

- quit Copilot, launch Copilot, delete the conversation, quit, launch Copilot and have it not start, which then allows repeating from the beginning

It would be much easier if I could just paste in the prompt specifying a folder full of files for it to run on, then clear that folder out for the next day's files and repeat.

Would that be something which your front-end could do? If not, is there one which could do that now? (apparently jan.ai has something like this on their roadmap for 0.8)

Hey, funny finding you comment as I've actually recently been developing a CLI app to improve LLM integration in my filesystem. Not sure what you are doing with your files, but maybe it could be useful for you too! You can check it out on my github (https://github.com/gael-vanderlee/whai) and hopefully that can help you with your use case. Look into roles in particular, as they allow you to save and reuse specific workflows.

I believe what you're describing is outside of the scope of Gerbil. Gerbil is not an LLM front-end, but Gerbil will run your LLM and seamlessly integrate (orchestrate) it with a custom front-end from the list in my original message. I believe this functionality will need to live in a custom front-end. I'm curious how jan.ai is planning on handling this. I'm guessing they're writing their own custom front-ends which is probably tightly integrated with their system.

Very hopeful of the multi-file stuff from jan.ai --- in the meanwhile, it's easier using Co-pilot for this than:

- isolating 50 files at a time

- dragging them into Adobe Acrobat

- closing each w/ ctrl w

- tapping enter to confirm saving

- typing the Invoice ID

- repeating until all 50 have been done, then remembering to quite Adobe Acrobat so as to re-launch it and repeat (can't leave it running, because there is a (reported) bug where after doing this several times, it stops saving)

- running a batch file made from a concatenated column in a spreadsheet to rename the files

The next question is when there will be an LLM front-end which can:

- open each file in a folder, parsing the content

- open each file in a PDF viewer

- fill in the entry fields of a Java application

- wait for the user to review both windows, if necessary, correct/update what was entered and save, then repeat for the next file

Ah well, job secure, even when that happens (though maybe hours would be cut back?) --- the big question is when LLMs will be reliable enough that human review is no longer viewed as worth the expense of a salary.

Reply on news.ycombinator.com