Research Agent

A command-line research assistant that combines local LLM conversation with a small multi-agent research pipeline. The app chats normally, and when the user asks to research a topic it searches the web, summarizes findings, checks the summary, and formats citations into a report.

Features

• Conversational CLI interface
• Local LLM calls through Ollama
• Research trigger routing via ConversationManager
• DuckDuckGo web search helper
• Summarization, fact-checking, and citation agents
• Formatted research report output

Project Structure

.
├── main.py                     # CLI entry point
├── conversation_manager.py     # Chat loop, tool-call parsing, research routing
├── orchestrator.py             # Coordinates the research pipeline
├── Agents/
│   ├── base_agent.py           # Shared Ollama chat helper and tool-call parsing
│   ├── search_agent.py         # Builds/searches research queries
│   ├── summariser_agent.py     # Produces concise summaries
│   ├── fact_checker_agent.py   # Reviews summaries for unsupported claims
│   └── citation_agent.py       # Formats source citations
└── Tools/
    ├── web_search.py           # DuckDuckGo search helper
    ├── summariser.py           # Lightweight extractive summarizer
    └── citation.py             # Basic APA-style citation formatter

Requirements

• Python 3.10+
• Ollama installed and running locally
• The llama3 model pulled in Ollama
• Python package:
  • requests

Setup

Install Python dependencies:

pip install requests

Start Ollama:

ollama serve

Pull the default model if needed:

ollama pull llama3

The model name is configured in conversation_manager.py and Agents/base_agent.py:

MODEL = 'llama3'

Change that value if you want to use a smaller or different local model.

Usage

Run the assistant:

python main.py

You can chat normally:

You: hello

To trigger the research pipeline, ask for research explicitly:

You: research renewable energy storage trends

Exit with:

quit

How It Works

1. main.py starts a terminal chat loop.
2. ConversationManager sends user input to the local Ollama model.
3. If the model returns a run_research tool call, the manager passes the topic to ResearchOrchestrator.
4. ResearchOrchestrator runs:
   • SearchAgent
   • SummariserAgent
   • FactCheckerAgent
   • CitationAgent
5. The compiled report is sent back through the conversation manager for a final assistant response.

Known Issues

• Some agent methods currently treat _chat() as if it returns a full Ollama response object, but BaseAgent._chat() returns only the message content string. Those methods may need to parse the returned string directly.
• Package directories are named Agents and Tools, while imports use lowercase agents and tools. This may work on case-insensitive filesystems like Windows, but can fail on case-sensitive systems.
• Agents/_init_.py and Tools/_init_.py are named with one underscore on each side. Python package initializers are normally named __init__.py.

Notes

This is a local-first prototype. It depends on Ollama for LLM responses and DuckDuckGo's public instant-answer endpoint for search results, so research quality and coverage can vary by topic.