Open WAS-PlaiLabs opened 2 months ago
We could technically export the code to be rerun elsewhere but the issue is that due to the interactive notion of Agents, we would have to do every execution in a sandbox. Not impossible of course but a bit involved.
We are aware it's not perfect and we do have it in the roadmap.
Could you point to the specific resources in Llama-index and autogen that could serve as inspiration? We would be happy to explore them, and thanks for pointing these out!
So in Autogen they have a flag param for the LLM config use_docker
which would then use docker if running and configured correctly. You can provide a custom image name in this flag apparently as well (say you needed specific tools installed for the LLMs).
https://microsoft.github.io/autogen/blog/2024/01/23/Code-execution-in-docker/
I'm stuck on my phone and not getting the same results I did before with source code hits so sorry for vague-ness.
This would be great to have.
OpenDevin has an action loop for coding that runs completely Dockerized. Might be worth borrowing some of the methods there! https://github.com/OpenDevin/OpenDevin
That looks pretty cool. Though would it be viable for a service?
On Mon, Jun 3, 2024, 11:25 PM Duke Jones @.***> wrote:
OpenDevin has an action loop for coding that runs completely Dockerized. Might be worth borrowing some of the methods there! https://github.com/OpenDevin/OpenDevin
— Reply to this email directly, view it on GitHub https://github.com/lavague-ai/LaVague/issues/264#issuecomment-2146707215, or unsubscribe https://github.com/notifications/unsubscribe-auth/BGIZLDMBJXYL3KQOA6FTTF3ZFVMT7AVCNFSM6AAAAABILPSIO6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDCNBWG4YDOMRRGU . You are receiving this because you authored the thread.Message ID: @.***>
there was some speculation that GPT Code Interpreter was implemented by spinning up hosted docker sandboxes. could do with some kind of kubernetes for bin-packing containers, plug it into a hosted dev environment offering like replit, gitpod, etc. If somebody wants to fund doing this & run the company, I have a few friends at the intersection of devops & ai, and also doing an ai (gpu) hardware cloud platform..... they'd probably be stoked to work on something like that
Could you provide internal methods to utilize docker, like autogen and other LLM tools that execute code? Figuring out how to wrap LaVague externally doesn't make me confident in helping secure execution of code.
API Keys are currently provided only through environ, which is not the best method for parallel services (where users use their own keys). Please provide a method to input the OpenAI API key as a param, like in Llama-Index itself.