This project provides a "Deep Research" agent designed to iteratively search, read webpages, and reason to find concise answers to complex queries, distinguishing itself from AI models focused on generating long-form articles. It's targeted at users needing precise, well-researched answers through an automated, multi-step investigation process.

How It Works

The system operates in a loop, driven by a user query. It leverages a Large Language Model (LLM) like Gemini or OpenAI for reasoning, a "Jina Reader" for web scraping and content extraction, and an iterative process of searching, reading, and reasoning. The core advantage lies in its focused approach to deep investigation, breaking down complex questions into sub-queries and refining the search and reading process until a satisfactory answer is found or the token budget is exhausted.

Quick Start & Requirements

• Install via git clone https://github.com/jina-ai/node-DeepResearch.git and npm install.
• Requires Node.js.
• Environment variables for API keys (Gemini, OpenAI, or Jina) are necessary: GEMINI_API_KEY, OPENAI_API_KEY, JINA_API_KEY.
• A free Jina API key with 1M tokens is available at https://jina.ai/reader.
• Official deployment and API available at https://search.jina.ai and https://deepsearch.jina.ai/v1/chat/completions.

Highlighted Details

• Iterative search, read, and reason loop for deep investigation.
• Supports Gemini (latest gemini-2.0-flash), OpenAI, and local LLMs (with structured output support).
• API is OpenAI-compatible, using jina-deepsearch-v1 as the model name.
• Handles citations in GitHub-flavored markdown footnote format.
• Provides a local server option (npm run serve) compatible with OpenAI clients.

Maintenance & Community

• Actively developed by Jina AI.
• Official API and UI are available for testing and use.
• Links to blog posts and implementation guides are provided for deeper understanding.

Licensing & Compatibility

• The repository does not explicitly state a license in the README.
• The API is OpenAI-compatible, facilitating integration with existing tools.

Limitations & Caveats

• The project is under active development, with npm installation not recommended currently.
• Performance and correctness can vary based on the LLM used and query complexity, as demonstrated by examples where answers were partially correct or required many steps.
• Not optimized for generating long-form articles.