Added doc indexing, moved processing to ray actors

[3coins]

Summary

This PR moves the chat prediction code to distributed tasks using ray actors, this will allow non-blocking processing of chat requests. To process the replies, a ReplyProcessor has been added that that will distribute the replies from the chat provider. This PR also includes the ability to load the specified directories into a FAISS index, which can be used to ask questions pertaining to the loaded documents. Users can type /index <dir-name> in the chat to start loading the directories. They can then use the /fs <question> or /filesystem <question> to ask a question, which will switch to a RAG chain to get relevant context from the indexed documents.

Demo

Tasks

• Moved chat provider prediction to ray actors, allows non-blocking distributed tasks
• Added a reply processor, single processor to distribute replies from actors, this was originally in the web socket
• Ability to index directories, users can index directories from chat by sending /read <dir-name>
• Ability to chat with indexed directories, users can chat with indexed directories by sending /fs <query>
• Save index to file, so users can ask questions about directories indexed in previous sessions
• Add a menu item for indexing directories from the file browser
• Add a data loader for faster directory indexing, explore dask or ray for this
• Add loading for python, markdown and R files
• Add sources in the reply
• Some visual indication of what directories are indexed
• Mechanism to reply to user if directory is already indexed, and take input from user if they want to update the index

[ellisonbg]

This is really amazing work! I left some specific feedback inline, but overall the design is really solid. It is great to see how we can use ray like this to build a more robust, distributed backend for usage cases like this. Super cool!