Workflow Provider

Architecture

FastAPI + Uvicorn webserver with a three-agent system (BuilderAgent, SearchAgent, and TaskIdentifierAgent) backed by ChromaDB vector storage for RAG-based workflow retrieval.

Components

• BuilderAgent: Creates and edits workflows using structured LLM output via ToolStrategy(Workflow)
• SearchAgent: Retrieves relevant workflows from vector store using semantic similarity
• TaskIdentifierAgent: Uses agent structured-output calls for intent classification, deadline extraction, context detection, and task construction
• AnalyzerAgent: Analyzes traces from Confident AI to extract patterns and update knowledge files (user preferences, task patterns, workflow trends)
• ChromaVectorStore: Manages two ChromaDB collections (manual_workflows, generated_workflows) with OpenAI embeddings

API Endpoints

• POST /create_workflow accepts a CreateWorkflowRequest (task, optional rejected_workflows, optional user_feedback, optional thread_id) and returns a structured Workflow
• POST /edit_workflow accepts an EditWorkflowRequest (task, proposed_workflow, feedback, optional thread_id) and returns an updated Workflow
• POST /edit_task accepts an EditTaskRequest (task, user_feedback, optional thread_id) and returns an EditTaskResponse with status: "edited", the edited Task, and context items
• POST /search_workflows accepts a SearchWorkflowsRequest (task, optional thread_id) and returns relevant workflows from the vector database using RAG
• POST /identify_task accepts an IdentifyTaskRequest (text, optional subject, optional metadata, optional thread_id) and returns one of:
  • identified with task: Task, and context_items: List[ContextItem]
  • no_task with task: null, and empty context_items
• POST /enrich_task_with_workflows accepts an EnrichTaskRequest (task, optional thread_id), attaches candidate workflows, and returns the enriched task
• POST /analyze_traces accepts an AnalyzeTracesRequest (thread_id) and analyzes all traces from the thread to extract patterns and update knowledge files
• POST /populate_workflows accepts { workflows: List[Workflow] } and returns inserted IDs/count for the manual workflow collection
• GET /health for health checks

Note: All endpoints that invoke LLM agents accept an optional thread_id parameter. When provided, it is used for DeepEval logging to enable request tracing and observability. If not provided, a new UUID is generated automatically.

Knowledge Files: The analyzer agent maintains three runtime-generated knowledge files in the knowledge/ directory: user_preferences.txt, task_patterns.txt, and workflow_trends.txt. These are populated by analyzing traces from Confident AI and should not be committed to version control.

Environment Variables:

• OPENAI_API_KEY (required): For LLM calls and embeddings
• CONFIDENT_API_KEY (required for trace analysis): For fetching traces from Confident AI API
• CHROMA_PERSIST_DIR (optional): Vector DB persistence directory (default: ./chroma_db)

Setup

Conda environment for clean local dev environments.

conda create -n "agents_ucsd" python==3.11
conda activate agents_ucsd
pip install -r requirements.txt
uvicorn app:app --reload --port 8080

If you want to use docker:

Build locally:

docker build -t workflow-planner .
docker run --rm -p 8080:8080 -e OPENAI_API_KEY=your_key workflow-planner

The container starts with:

uvicorn app:app --host 0.0.0.0 --port $PORT

Testing

Run integration tests:

python tests/test_suite.py

Run unit tests:

python -m unittest tests.task_unit_test -v

Configuration:

• Set WORKFLOW_API_URL environment variable to point at a running server (defaults to http://127.0.0.1:8080)
• Mock tasks in ./prompts/*.txt can include rejected_workflows, proposed_workflow, and feedback fields