Conversation from Discord:
Aziz
Hi SantiagoDePL I was curious about Video generation models. Following OpenAI's API specification, its a request and poll method. Where you call the endpoint /videos with a prompt and other variables, then you get back an ID and poll an endpoint until it says "completed".
Have you thought about how you would implement that for GoModel? I imagine it would need to have connections to a storage - or multiple options - type/platform. Could be file server, or buckets, etc.
Then also there's the time to live for the video itself. I think openai allows the videos to live for 1 hour before they're deleted. I didn't see this in the roadmap when I checked last.
SantiagoDePL
I still need to think about it - that’s why there is no details about it on the roadmap page.
Definitely database is a wrong place to store files. Today we log audio files there (when appropriate env variables is set to true). It will be changed.
I can imagine that in first implementation you will be able to set a path for file storage. For example ‘/mnt/files’ . In this case you will be able to mount any storage on Claude/Infra layer there (AWS EFS and S3 would work this way when configured properly).
Later perhaps I will add a native support for S3 and similar cloud services. (It might be done based on the plugins system I recently developed)
WDYT about this?
Of course it will be as much configurable as possible (for example a different retention period for different media types etc.)
Aziz
I think that sounds like a good plan. I think one point to consider is that GoModel needs to:
Store the ID and maybe poll for the user if the user is using a non-self-hosted model (from MiniMax, Gemini, OpenAI, etc.)
Edit: vLLM exposes /content and polling so this is not an issue
Conversation from Discord: