Create Initial Version of Supervisor Agent for Kyma Companion

[Teneroy]

Description

The goal of this task is to create the initial version of the supervisor agent for Kyma Companion. This includes expanding the chat controller to accept necessary information, setting up a proper technique for the supervisor agent, and creating sample nodes for two agents and one tool for future development.

Subtasks

1. Expand Chat Controller:

• Modify the chat controller to accept session ID and cluster access information from the busload backend.
• Ensure the chat controller can properly handle and store this information for use by the supervisor agent.

2. Set Up Supervisor Agent Technique: Select and set up a suitable technique for the supervisor agent (for example Planning, Routing).

3. Create Sample Nodes:

• Develop dummy nodes for two agents that will be used by the supervisor agent. (Kyma agent, Kubernetes agent)
• Develop a dummy node for one tool that will be utilized in future development.

Acceptance Criteria

• [x] The chat controller is expanded to accept session ID and cluster access information from the busload backend.
• [x] The supervisor agent technique (for example Planning, Routing) is set up.
• [x] Dummy nodes for two agents are created and functional.
• [x] Planner should consider conversation history.
  • If possible, supervisor/planner should answer the common questions.

[Teneroy]

https://github.com/langchain-ai/langgraph/blob/main/examples/multi_agent/agent_supervisor.ipynb

[Teneroy]

• Make planner more intelligent by adding information from the state like conversation history and/or messages
• Improve Finalizer so it would be able to analyze whether responses from agents make sense, and fail accordingly when they are not
• Merge planner and common node for common question. In case there is no need to go with the plan, planner should be able to generate the response to user directly without going to the task executor

[muralov]

Further Improvements in the code:

• [x] Upgrade generative-ai-hub-sdk package
• [x] Simplify the routing according to the discussion
• [x] Make generate LangGraph diagram script run with python cli and document according to comment
• [x] Move environment variables constants to central config module
• [x] conversation_id is of string type (uuid), not int.
• [x] Streaming output should be as designed
• [x] Improve the error handling
• [x] Add unit tests for KymaGraph

[Teneroy]

Proper streaming format

[muralov]

Things that can be improved in the future:

• Improve the history handling
• Improve error handling (fallback logic)
• Retry logic is introduced for each node, for example if LLM invoke fails
• Subtasks may run concurrently to improve the speed

[muralov]

Current LangGraph workflow