RecurSearch is an AI-driven recursive web research tool that performs comprehensive, multi-layered research similar to OpenAI's Deep Research. It uses specialized AI agents to generate detailed research reports with citations, statistics, and relevant images.
- 🔄 Recursive Web Search - Multi-layered searches that dive deeper with intelligent follow-up questions
- 🤖 AI-Powered Agent System - Specialized agents for research generation, validation, and content curation
- 📊 Statistical Support - Reports backed by relevant data and statistics
- 🖼️ Intelligent Image Selection - Research-focused image searches and integration
- 📄 Multi-Page Reports - Generate comprehensive reports up to 2 pages
- 📝 Citation Support - Optional citation generation with credible sources
- 🔧 Customizable Research Depth - Control how deep the research goes
- 📋 Comprehensive Logging - Local logs and reports for transparency and debugging
- 🔍 Quality Evaluation - Automated evaluation for toxicity, bias, and answer relevancy
- 🌐 AgentNetwork Integration - Support for Mastra's AgentNetwork framework
- Node.js v20+ installed on your system
-
Clone and navigate to the project:
cd recursearch npm install -
Set up environment variables in
.env.development:OPENAI_API_KEY=your-openai-api-key TAVILY_API_KEY=your-tavily-api-key EXA_API_KEY=your-exa-api-key
-
Start the development server:
npm run dev
-
Generate reports:
- Go to the Search Agent tab in the Mastra playground running at
localhost:4411and request report generation - The search agent has memory enabled for follow-up questions
- For AgentNetwork features, use the Networks tab in the playground
- Go to the Search Agent tab in the Mastra playground running at
RecurSearch supports multiple AI models. Configure your preferred AI provider and model in config/config.ts by changing the settingRunTimeContext variable.
out of box, Recursearch has support for
- OpenAI
- Claude
- Gemini
more providers and models can easily be added
| Agent | Purpose |
|---|---|
| Search Agent | Orchestrates the research process and formats final reports |
| Question Agent | Generates insightful follow-up research questions |
| Custom Question Agent | Validates question uniqueness to avoid redundancy |
| Question Picker Agent | Selects the most research-focused questions |
| Stats Agent | Generates statistical queries to support findings with data |
| Image Agent | Creates targeted image search queries for visual content |
| Content Picker Agent | Evaluates and selects the most credible web sources |
Comprehensive research with web tools and data integration:
- simpleSearchAgent - Initial data gathering
- imageSearchAgent - Research-relevant image discovery
- reportFormattingAgent - User-defined formatting compliance
- QuestionAgent, QuestionPickerAgent, CustomQuestionAgent - Question management
- ImageQueryAgent - Precise image search queries
- StatsAgent - Statistical integration
- citationAgent - Credible source identification
Enhanced with planning and evaluation agents:
- PlannerAgent - Structures the research process
- PlannerEvalAgent - Evaluates research plan effectiveness
- ReportEvalAgent - Assesses report quality and completeness
Simulates debate-like processes for controversial topics:
- ResearchPresenterAgent - Presents arguments and findings
- ResearchCriticAgent - Generates counterarguments
- ReportJudgeAgent - Evaluates arguments for balanced conclusions
Note: Debate version intentionally excludes web tools to focus on reasoning-based arguments.
All example reports are completely AI-generated without manual editing.
Highest quality reports with full recursive research:
- 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
- How LLMs Work
- Mental Health Research
Each report includes preview screenshots as authenticity proof.
Template-style reports with good structure but variable fact quality:
- Recommendation Algorithms and User Behavior on Social Media
- Music Therapy and Depression Impact
- Chronic Stress and Immune Function
Plan-evaluate-improve methodology reports:
- How LLMs Work (Reflection Version)
Argument-based reports using pure reasoning (no web data):
- Benefits of Debate on Research Reports
- Music Therapy's Influence on Depression
- Phone Usage vs. Mental Health
RecurSearch uses a decision tree approach where agents collaborate to:
- Generate multiple research questions
- Validate question uniqueness and relevance
- Select the most promising research paths
- Search recursively through multiple layers
- Evaluate content credibility and relevance
- Synthesize findings into comprehensive reports
- Assess final output quality and bias
This tool was created for research into multi-agent report generation.
Important:
- Must not be used unethically
- Authors take no responsibility for misuse
- Always verify AI-generated content
- Use responsibly for legitimate research purposes
RecurSearch is inspired by:
GPL - Please see license requirements for full terms.
For detailed architecture diagrams and advanced configuration, see the /docs directory.