RecurSearch is an AI-driven recursive web research tool designed to mimic OpenAI's Deep Research. It can search the web recursively to generate comprehensive research reports with or without citations.
- Recursive Web Search: Perform multi-layered web searches, diving deeper into follow-up questions.
- AI-Powered Agents: Specialized agents for generating, validating, and picking research questions and web content.
- Stats support: the reports are backed by stats
- Images: the reports contain related research focused searched images
- Pages: the report can be generated up to 2 pages
- Logging with reports, the local generated copy by all the agents plus log data is kept in local-log and local-report files to explore agents answers
- Citations Support: Option to generate research reports with citations.
- Customizable Depth: Control the search depth of research.
- Auto Generated Report: the reports, logs, and local report are automatically generated locally by agent
- Recursearch also has support for Mastra's AgentNetwork
Ensure you have Node.js v20+ installed, then add the dependencies:
cd recursearch
npm iSet up env variables:
OPENAI_API_KEY=your-openai-api-key
TAVILY_API_KEY=your-tavily-api-key
EXA_API_KEY=your-exa-api-key
Run Mastra Dev server to create reports
npm run dev
Go to search agent and ask it to generate reports.
For AgentNetwork go to networks tab and ask it to generate reports.
RecurSearch leverages multiple AI agents with distinct roles:
-
Search Agent Understands users query to call the searchAgentTool which kicks of MAS, its also responsible understand and format the report to meet user defined requirements.
-
Question Agent: Generates thought-provoking follow-up research questions based on content.
-
Custom Question Agent: Validates whether a newly generated question is thematically unique or overlaps with existing ones.
-
Question Picker Agent: Compares pairs of research questions and selects the most research-focused one.
-
Stats Agent Responsible for generating stats question for given running question to support research content by data
-
Image Agent Responsible for generating image search query for given running answer, to ensure images are research quality
-
Content Picker Agent: Evaluates and picks the most relevant and credible web source from two given options. [When Citations are required]
This Agents collabrate in a team to recursively search the web, and in a decision tree manner pick the best path in the tree to come up with the report.
RecurSearch integrates a robust AgentNetwork comprising the following specialized agents:
- simpleSearchAgent: Conducts straightforward web searches to gather initial data for research topics.
- imageSearchAgent: Searches for high-quality, research-relevant images to enhance the report.
- reportFormattingAgent: Ensures the final report adheres to user-defined formatting and presentation requirements.
- QuestionAgent: Generates insightful follow-up questions to deepen the research process.
- QuestionPickerAgent: Selects the most research-focused question from a set of generated options.
- CustomQuestionAgent: Validates the uniqueness of newly generated questions to avoid redundancy.
- ImageQueryAgent: Creates precise image search queries to ensure the inclusion of relevant visuals.
- StatsAgent: Generates statistical queries and integrates data to support research findings.
- citationAgent: Identifies and incorporates credible sources to provide citations for the research report.
There is also a reflection version that uses the above agents
- PlannerAgent: Creates a structured plan for the research process, outlining the sequence of tasks and agent interactions.
- PlannerEvalAgent: Evaluates the effectiveness and feasibility of the research plan created by the PlannerAgent.
- ReportEvalAgent: Assesses the quality, accuracy, and completeness of the final research report, providing feedback for improvements.
The Debate Version introduces agents designed to simulate a debate-like process to refine research findings and ensure diverse perspectives are considered. These agents include:
- ResearchPresenterAgent: Engages in a debate with other agents, presenting arguments for or against specific research findings or questions.
- ResearchCriticAgent: Generates counterarguments to challenge the points raised by the Presenter, fostering a balanced discussion.
- ReportJudgeAgent: Evaluates the arguments and counterarguments to determine the most credible and well-supported conclusions.
This version is particularly useful for exploring controversial topics or areas with conflicting information, as it encourages a thorough examination of multiple viewpoints. The reports generated by the Debate Version aim to present a balanced and well-rounded perspective on the research topic. Note the reports do not contain stats/facts as this agents don't use tools this was done on purpose to see how they argue without web data.
These agents collaborate in a multi-agent system, working together to refine and enhance the research process, the reports generated by AgentNetwork is not as accurate as Recursearch, yet they can generate a good report template.
can view digrams to understand the flow.
the reports/tree contain recursearch's reports that were solely generated by the agents and I didn't touch/modify them. There are preview screenshots for proofs.
The examples include reports on
- The Influence of Chronic Stress on Immune System Function
- Impact of AI-Generated Content on Journalism and Media Credibility
- Impact of Music Therapy on Individuals with Depression
the reports/network/normal contain the AgentNetwork's version report, these reports are non-deterministic in terms the quality of facts/citations and images but still are good template reports.
The reports/network/planaftereval contains a reflection version
Again I didn't touch/modify them.
Some examples include reports on
- General Report on Recommendation Algorithms and User Behavior on Social Media
- Impact of Music Therapy on Individuals with Depression
- The Influence of Chronic Stress on Immune System Function
the reports/network/debate contain research report that have been debated and judge, this reports don't contain real data from web or stats as this shows how agents argue for their points without research from outside.
Some examples include reports on
- Benefits of debate on research report
- Music theraphy influence on depression
- Phone usage vs mental health
this tool was created to research how agents can generate research reports, the tool must not be used unethically, the authors of recursearch take no responsibility in doing so.
Recursearch was inspired by following paper:
GPL
Please see the license requirements
- @jalpp