Deep-Research-Bundles

Multi-Agent RAG Research System

Deep-Research-Bundles is a comprehensive multi-agent research system that orchestrates six specialized agent types to conduct deep literature review, fact verification, synthesis, and report generation. The system goes beyond simple RAG by implementing a full research workflow with planning, search, analysis, critique, and writing agents that collaborate through structured protocols.

The six agent types are: Planner (decomposes research questions into sub-queries), Searcher (executes multi-source retrieval), Analyst (extracts and synthesizes findings), Critic (evaluates evidence quality and identifies gaps), Writer (generates structured reports), and Coordinator (manages workflow and resolves conflicts). Four execution modes support different use cases: sequential for simple queries, parallel for broad surveys, iterative for deep dives, and adversarial for controversial topics.

The hybrid retrieval system combines Qdrant vector search, Elasticsearch keyword search, and LLM-based reranking for maximum recall and precision. The MCP (Model Context Protocol) integration enables seamless tool use, allowing agents to access web search, academic databases, code repositories, and structured data sources through a unified interface.

Built on FastAPI for the orchestration server, CrewAI for agent management, and LangGraph for workflow definition, the system produces research reports with full provenance tracking and citation management.

• 6 specialized agent types: Planner, Searcher, Analyst, Critic, Writer, Coordinator
• 4 execution modes: sequential, parallel, iterative, and adversarial
• Hybrid retrieval with Qdrant vector + Elasticsearch keyword + LLM reranking
• MCP protocol integration for unified tool access across diverse data sources
• Full provenance tracking and citation management in generated reports