feat: agent

[atarashansky]

PR Description

Summary

This PR introduces a new "Agent Component" to the frontend and corresponding backend support for managing AI-assisted workflows within the application. The implementation enables users to interact with an agent for automated actions, visualizations, and insights, specifically tailored to single-cell data analysis and visualization tasks.

Changes in This PR

1. Frontend Enhancements:

• New AgentComponent for user interaction:
  • Implemented a draggable and resizable UI component using react-draggable and re-resizable.
  • Integrated ReactMarkdown for rendering markdown-based responses.
  • Enabled chat history management and message editing features.
• Redux Action:
  • Added a Redux action for reverting application state to a specific action count.
• App Modifications:
  • Modified App to display the agent component conditionally based on a feature flag.
• Dependencies:
  • Updated package dependencies:
    • Added react-markdown, re-resizable, react-draggable, and zod.

2. Backend Integration:

• AI Agent Endpoint:
  • Introduced an AI agent endpoint (/agent/step) in the Flask API:
    • Handles multi-step agent workflows using the langchain library.
    • Supports dynamic tool invocation for dataset-specific operations.
• Tool System:
  • Added a modular tool system in server.common.tools:
    • Tools include subsetting, categorical selection, coloring, histogram selection, and more.
• OpenAI Integration:
  • Implemented a structured OpenAI API integration for agent response generation.

3. New Features and Functionality:

• Undoable State Enhancements:
  • Added a new revert-to-action action type for undoable state management.
• Agent-Specific Redux Actions:
  • Introduced actions for categorical and continuous data selections.
• Enhanced Metadata Handling:
  • Improved schema parsing for agent operations.

4. Technical Improvements:

• Refactored and extended existing Redux reducers to accommodate agent-specific states.

5. Developer Dependencies:

• Added langchain, langchain-openai, and related libraries for structured agent workflows.
• Updated pydantic to version 2.x for enhanced model validation.

Usage

• Users can interact with the agent via the UI to perform tasks like subsetting, coloring, or visualizing data points.
• Developers can extend the toolset or modify agent behavior by adding new tools or updating the system prompt.

Testing

• Manual testing for agent workflows, tool integration, and UI interactions.
• Verified compatibility with existing app features and ensured no regressions.

Next Steps

• Add unit tests for backend API integration and tools.
• Enhance UI accessibility and error handling for agent interactions.

Known Issues

• Limited test coverage for new agent workflows.
• Further optimization needed for large dataset interactions.
• Summarization step sometimes includes already-summarized actions in the summary.