Analyse this document.

[iplayfast]

Thinking of an enhancement. With llava, you could ask what a picture is about and give the file location.
I wonder if it would be useful or worthwhile to analyse a document by giving it the file location.

Downsides, no rag so info can't be easily stored.
Upsides, would be super useful and can use as a reference, using the document itself as the storage medium while its running, ie load the vectors into memory that links to file locations.

Use case:
rewrite the code located at ./oldcode.py and save the new version at ./newcode.py change it to do ....

[iplayfast]

maybe using a /set parameter to insure safefy.
eg
/set parameter outDir "/home/ollama/generatedcode/"
anytime a file is written, the outDir is prepended it the directory.

[technovangelist]

Hi, thanks for the issue.

If the file fits in the context of the model, then you can do something like this today:

ollama run llama2 "can you rewrite this fish function more efficiently: $(cat md.fish)"

We won't write to the filesystem, so it would be up to you to redirect output to a file. But usually you should verify before writing.

[iplayfast]

Hi, thanks for this great project. I've never enjoyed AI more!
Yes I've been doing exactly as you've said, and I don't know how this would work using llama.cpp just felt that if you can do it for images, you should be able to do it for text.
and if I specify a file to write to, it should be able to write to the file.

[jmorganca]

Hi @iplayfast, will merge this with #834