Multi-Agent Research

A Streamlit app for running a configurable multi-agent research workflow.

Features

• Configure and launch multi-agent LLM workflows for topic-specific research
• Pick a research profile from the sidebar to switch topic defaults
• Real-time progress and agent status display
• Downloadable markdown reports
• Trusted sources configuration
• Supports OpenAI and Ollama LLMs

Architecture

• multi_agent_workflow.py owns the workflow engine, typed workflow state, agent roles, and normalized runtime events.
• research_profiles.py loads topic-specific research profiles and composes the prompt source policy.
• workflow_adapters.py isolates Tavily search initialization and report persistence from the engine.
• app.py owns the Streamlit UI, run control, and event rendering.
• See docs/agentic-architecture.md for the explicit workflow contract.

Requirements

• Python 3.10+
• See requirements.txt for dependencies

Setup

1. Clone the repository:

   git clone <repo-url>
   cd multi-agent-research

2. Install dependencies:

   pip install -r requirements.txt

3. Create a .env file in the project root and add your API keys as needed:
   • For Tavily (required for web search):

     TAVILY_API_KEY=your_tavily_api_key_here

   • For OpenAI (required if you select OpenAI as LLM Type):

     OPENAI_API_KEY=your_openai_api_key_here

   • To switch the active research profile used by the workflow and UI defaults:

     RESEARCH_PROFILE_PATH=research_profiles/health_science.yaml

     Profiles are YAML files stored under research_profiles/. Copy the default profile and adjust the topic, audience, tone, target word count, report filename, and source tiers to fit a different research domain.
   • To override the default research source tiers used by the workflow and sidebar:

     PRIMARY_TRUSTED_SOURCES=PubMed (pubmed.ncbi.nlm.nih.gov),The Lancet (thelancet.com),Nature Medicine (nature.com/nm)
     SECONDARY_TRUSTED_SOURCES=Google Scholar,ScienceDaily (sciencedaily.com),Healthline (healthline.com)

     Values can be comma-, semicolon-, or newline-separated.
4. Run the app:

   streamlit run app.py

Testing

• Run the workflow tests with:

  .\.venv\Scripts\python.exe -m unittest discover -s tests -p "test_*.py"

• The mocked orchestration test covers the research -> draft -> approve -> publish path without calling external services.

Usage

• Choose a research profile in the sidebar to control the topic, prompt defaults, and source tiers
• Configure workflow parameters in the sidebar
• Start/stop the workflow and monitor progress
• Download generated reports from the UI

File Structure

• app.py — Main Streamlit app
• multi_agent_workflow.py — Workflow logic
• workflow_adapters.py — External service adapters for search and persistence
• llm_factory.py — LLM integration
• tests/ — Config and orchestration tests
• docs/ — Generated reports

License

MIT License