Create API interface for compiled graph

[TanGentleman]

I like the inspiration from the agent-service-toolkit repo: https://github.com/JoshuaC215/agent-service-toolkit/

Specifically:
Service
Agent (client)
Simple streamlit app

[TanGentleman]

Using State Machine from tan_graph.py:

To add /upload to our api

• User will pass auth details and a request
• From auth, we obtain the correct agent state, and add a task: Upload, with the Fetch Action populated
• This allows for different flows depending on request body and AgentState, like if the user can be interrupted.

Sample flow:

• Backend receives auth + upload request
• Construct an Upload Action (w/ populated ActionArgs)
• Create a new Task and populate backend DB
• RETURN success — added to queue
• Based on the user's settings / request type, this can also be immediately executed

After graph is compiled, run with the user's thread_id:

• Assume state_dict["task_dict"] is populated with an unstarted task (unstarted)
• Start Node (load) -> Agent Node (decision) -> Task Manager (initialize state dict and update task status to in-progress) -> Agent Node (decision) -> Action Node (executes Upload Action and modifies state_dict from ActionResult -> Agent Node (decision) -> Task Manager (updates backend DB with completed task, updates status of task_dict) -> Agent Node (decision) -> END

This allows for:

• Realtime updates to user with optimized async/parallelism
• Custom backend logic for UX flow by use case, like smoothly handling a backlog of upload jobs

The Actions abstraction allows for high compatibility with HITL and Chat. This meshes well with more complex RAG steps.