DeepSearch Agents

A minimal multi-step research agent framework for DeepSearch AI. Leverages OpenAI chat models, SerpAPI web search, and customizable tools to automatically decompose complex queries, gather evidence, and produce verified answers with references.

Table of Contents

1. Features
2. Installation
3. Configuration
4. Usage
5. Project Structure
6. Development
7. License

Features

• Task decomposition & planning with recursive sub-tasks
• Integrated web search (SerpAPI) & page retrieval (Jina)
• Content summarization & quote extraction
• Automated answer evaluation & refinement
• Configurable LLM models, token limits, and execution policies
• Token usage tracking and usage-based branching

Installation

git clone https://github.com/your-org/deepsearch-agents.git
cd deepsearch-agents
pip install .       # or: pip install -e .

Configuration

Configure API keys and model/execution settings before running.

1. Create a .env file or export environment variables:

   OPENAI_API_KEY=your_openai_api_key
   TRACING_OPENAI_API_KEY=your_tracing_key
   SERPAPI_API_KEY=your_serpapi_api_key
   JINA_API_KEY=your_jina_api_key
   OPENAI_BASE_URL=optional_custom_base_url

2. Edit settings.yaml to customize:

   • models: LLM names, temperatures, max tokens, tool options
   • execution: max task depth, max turns, token usage limits

Usage

Run the CLI entrypoint to issue a query:

python main.py --query "How has the SPX performed in the last 30 days?"

The agent will:

1. Reflect on the question and generate sub-questions
2. Perform web searches and visits to gather knowledge
3. Summarize content and build answers
4. Evaluate and refine the final answer with references

Project Structure

. 
├── main.py               # CLI entrypoint
├── settings.yaml         # Default model & execution configs
├── pyproject.toml        # Project metadata & dependencies
└── src/
    └── deepsearch_agents/
        ├── conf.py       # Configuration loader (env + YAML)
        ├── context.py    # Task & context management
        ├── planner.py    # Task planning & orchestration agent
        ├── tools/        # Built-in LLM tools (search, visit, ...)
        └── llm/          # LLM API integration & response parsing

Development

• Python >= 3.10
• Install dev dependencies (if provided):

  pip install -e .[dev]

• Linting & formatting: pre-commit run --all-files
• Type checking: mypy src
• Contributions welcome via issues and pull requests.

License

MIT License. See LICENSE for details.