Performing "Redo" for a prompt with an image doesn't send the image

[neodon]

Expected

1. Add an image with a prompt.
2. Model addresses the image with the prompt appropriately.
3. Click "Redo" on the prompt box.
4. Model again addresses the image and the prompt.

Actual

Clicking "Redo" only sends the text and not the image.

Example

First prompt with image

You can see it sends 789 tokens, so we can assume most of those are for the image. The model response is appropriate for the meme I gave it.

Clicking "Redo" on the prompt

This time it only sends 24 tokens, so the image was not sent. This model is openrouter/polaris-alpha, which seems to make up random stuff if not given an image to describe. 😅

[mythz]

That's because there is no backend storage and files are too large to keep in browser storage so after it's sent the first time it's discarded and no longer available.

We may be able to do some sha256 file lookups to support this, I'll leave this open.

[neodon]

That makes sense. I think modern browsers on desktop support at least 1GB per origin, but other devices would probably be much more constrained. It could also introduce headaches for managing that storage as it grows.

I was playing around with the codebase and have a branch that demonstrates storing attachments in local storage. It's mostly AI-generated since I'm still learning, but I verified it works for images and audio.

https://github.com/neodon/servicestack-llms/tree/redo-includes-attachments

[mythz]

Yeah browser storage will work for a while, but when screenshots can take up to 10MB each and would need to be base64 encoded to store in IndexedDB (taking up even more space) I don't think that's going to be a viable long-term solution.

I'd rather store an optimized version on disk at something like ~/.llms/cache so it could be manually cleared out to save space.