File upload

[GeoDerp]

Would love the ability to upload files for evaluation in the webserver.

I would love to help with this when I have time.

[rhatdan]

Not sure what you mean? We can currently push models to OCI, do you want to push models to Huggingface and Ollama?

[GeoDerp]

Sorry about the confusion. 😅
On the olama serve webserver. I would love the ability to input .txt and .pdf files into the prompt with the text.

[rhatdan]

You mean ramalama serve, but would love to see a PR to see what you would like.

[ericcurtin]

We can do things like:

cat some-textfile.txt | ramalama run "Some prompt:"

already.

Not sure if that's the kinda thing you are looking for.

.pdf not quite there yet...

[GeoDerp]

We can do things like:

cat some-textfile.txt | ramalama run "Some prompt:"

already.

Not sure if that's the kinda thing you are looking for.

.pdf not quite there yet...

Yeah totally!
But yes, the ability to parse multiple files / pdfs was the idea. ❤️

[amsnek]

The ability to upload files (pdf, txt, etc) directly via the webinterface (ramalama serve / llama.cpp) directly to summarize would be a very nice feature.

[ericcurtin]

Yes, we'd have to explore what's already out there in this area and try and integrate. It's unlikely we'd do all this from scratch :)

[markhpc]

New user here and very much would appreciate a capability like this as well. Basically an upload button next to the prompt ala chatgpt or Claude. Implementing the storage side of this as a sandboxed directory similar to chatgpt's "/mnt/data" where project files can be copied at the start of a session but the AI has read/write access for it's own use would be very useful. Ramalama could then use that as a staging area for uploaded files, but it would never touch the originals outside of the session scope.

FYI, I'm one of the core Ceph developers are have been interested in enabling a more permanent object storage solution for persistent memory as well, but that's probably outside the scope of this specific feature request.

[rhatdan]

We now have support for running the client with context information, not sure if this fully satisfies this requirement.
@ieaves Please comment.

[ieaves]

I believe so! Should be good to close.