node-DeepResearch - An open-source reproducible AI agent from Deep Research, supporting multi-step inference and complex queries.

What is node-DeepResearch?

node-DeepResearch is an open-source AI agent project that continuously searches and reads web pages, using the Gemini language model and Jina Reader tool to progressively reason and answer complex questions until an answer is found or the token budget is exceeded. The project supports multi-step reasoning and complex queries, handling tasks ranging from simple questions to complex multi-step inference. node-DeepResearch provides a Web Server API, allowing users to submit queries via HTTP requests and obtain real-time progress updates.

Main functions of node-DeepResearch

• Continuous searching and readingUse search engines (such as Brave or DuckDuckGo) to find relevant information, read web page content, and continue until you find the answer to your question or exceed your set token budget.
• Multi-step reasoningTo handle complex, multi-step problems, break the problem down step by step and solve it step by step.
• Real-time progress feedbackIt provides real-time progress updates based on the Web Server API, allowing users to track the progress of their queries at any time.
• Flexible search methodsIt supports everything from simple factual questions to complex open-ended questions, such as predicting future trends or developing strategies.

The technical principles of node-DeepResearch

• Language ModelGemini is used as the core language model, responsible for generating responses and reasoning logic. Gemini is a powerful general-purpose language model capable of handling various natural language tasks.
• Web page content processingThis tool uses Jina Reader to convert web page content into a plain text format suitable for language model processing. Jina Reader is an open-source tool specifically designed for processing HTML web page content.
• Search and Information ExtractionThe agent obtains relevant information based on search engines (such as Brave or DuckDuckGo). The search engine provides web page links, and the agent extracts useful information by reading the web page content.
• Multi-step reasoning process:
  • initialization: Set the initial context and variables.
  • Budget reviewCheck the token budget at each step to see if it is exceeded.
  • Generate prompts: Generate input suggestions for the language model based on the current question.
  • Model generation: Call Gemini to generate an answer or the next action.
  • motion processing: Execute the corresponding operation based on the generated action (such as searching, accessing web pages, generating answers, etc.).
  • Results Evaluation: Evaluate whether the generated answer meets the requirements of the question; if not, continue reasoning.
• Loop and TerminationIf the answer is found within the budget, the query ends; if the answer is exceeded or cannot be found, "Beast Mode" is entered to generate the final answer.

The project address for node-DeepResearch

• GitHub repository:https://github.com/jina-ai/node-DeepResearch

Application scenarios of node-DeepResearch

• literature reviewIt helps researchers quickly find and organize literature in relevant fields and generate literature review reports.
• Market researchQuickly collect and analyze market data to understand competitor activities, industry trends, etc.
• News reportIt can quickly collect and organize background information and related reports on news events, helping reporters to quickly write news articles.
• Learning aidsIt helps students solve complex academic problems and provides learning materials and reference answers.
• Policy ResearchIt helps policymakers collect and analyze background information and implementation effects of relevant policies.