Minimal Deep Research

A minimal Gemini/OpenAI-inspired implementation of Deep Research using structured outputs with LangChain, an LLM provider, and a web search API.

Want to try it out? You can use any OpenAI-compatible LLM - whether that's a local model through Ollama or cloud options like Gemini 2.0 (shown in the demo below).

minimal-deep-research.demo.mp4

What's This?

Deep Research combines the power of LLMs with web search (via Tavily or Perplexity) to do what researchers do: ask good questions, gather information, and synthesize findings. But instead of taking hours, it takes minutes.

The core functionality resides in the DeepResearcher class, which:

• Crafts smart search queries based on your research topic
• Scours the web for relevant information
• Summarizes what it finds
• Identifies gaps and digs deeper
• Pulls everything together into a clear final report

Getting Started

1. Grab the code:

   git clone https://github.com/charlesnchr/minimal-deep-research
   cd minimal-deep-research

2. Set up your environment:

   python -m venv venv
   source venv/bin/activate
   pip install -r requirements.txt

3. Add your API keys: Create a .env file in the root directory and add your keys (e.g., PERPLEXITY_API_KEY)

How to Use It

Basic usage is straightforward:

python deep_research.py "Your research question here"

To use some of the optional command-line arguments:

python deep_research.py "What are recent advances in quantum computing?" \
    --max-loops 10 \
    --model-id "o3-mini" \
    --search-api tavily

Configuration Options

• topic: The research topic to investigate
• --max-loops: How many research iterations to run (default: 3)
• --model-id: Which LLM to use (default: llama-3.3-70b)
• --search-api: Pick your search provider - "tavily" or "perplexity" (default: tavily)

Acknowledgements

This was inspired by ollama-deep-researcher, but with a focus on simplicity. The core logic lives in a clean deep_research.py file - no LangGraph topology, nor the LangGraph server required to run it, instead relying on a simple main loop and the deep_research.py as an entrypoint.