ResearchGPT — Autonomous Research Intelligence

ResearchGPT is a production-grade autonomous research system designed to synthesize complex information across documents and the web using a 6-agent pipeline. It performs deep research, data extraction, parallel reasoning, and synthesis with automated knowledge graph generation.

Built with FastAPI, React, MongoDB Vector Search, and Ollama, powered by grok-3 / grok-3-mini (with Gemini Flash fallback).

---

🏗️ Architecture: The 6-Agent Pipeline

The core of ResearchGPT is a deliberate, sequential agentic pipeline that maximizes reasoning depth while minimizing context window fatigue and hallucinations.

1. Planner Agent (agents/planner.py)

• Takes raw user queries and decomposes them into a structured ResearchPlan.
• Determines research scope, required domains, search strategies, and analytical depth.
• Injects insights from past user sessions (Memory Agent feedback loop).

2. Retrieval Agent (agents/retrieval.py)

• Executes multi-source parallel searches across MongoDB Vector Search, Full-Text Search, arXiv, and Semantic Scholar.
• Ranks, deduplicates, and fuses results using Reciprocal Rank Fusion (RRF).
• Fetches raw PDF content and local Nomic embeddings for precise semantic matching.

3. Analysis Agent (agents/analysis.py)

• Reads raw retrieved documents and extracts structured Findings.
• Identifies core claims, methodologies, limitations, and recurring themes.
• Runs in highly parallelized batches using bounded concurrency with grok-3-mini.

4. Reasoning Agent (agents/reasoning.py)

• Contradiction Detection: Cross-references all claims to detect consensus and conflict.
• Evidence Graph: Constructs a node-link graph mapping documents to themes and claims.
• Gap Analysis: Evaluates findings against the original ResearchPlan to identify what's missing.

5. Synthesis Agent (agents/synthesis.py)

• Synthesizes all data into 4 simultaneous output formats:
  1. Narrative Literature Review (Markdown with inline citations)
  2. Actionable Research Gaps
  3. Comparative Analysis Matrix (JSON/Table)
  4. Document-Theme Citation Matrix (Boolean grid)

6. Memory Agent (agents/memory.py)

• Summarizes the completed research session.
• Extracts new nodes and edges for the global topic knowledge graph.
• Updates the user's persistent "Expertise Vector" stored in MongoDB.

---

🛠️ Tech Stack

Frontend (Client)

• React 18 + Vite
• TypeScript + Tailwind CSS
• Zustand (State management)
• React Router DOM
• Glassmorphism aesthetic

Backend (ML Service)

• Python 3.11 + FastAPI
• Anthropic / xAI / Google LLM Routers (Exponential Backoff + Fallbacks)
• nomic-embed-text via Ollama (Local Embeddings)
• Asyncio (Parallel agent execution)

Infrastructure & Data

• MongoDB Atlas (Vector Search & Document Store)
• WebSockets (Real-time agent streaming to UI)

---

🚀 Getting Started

1. Launch MongoDB & Ollama Ensure MongoDB Atlas URI is set and Ollama is running nomic-embed-text.

2. Start Backend ML Service

   cd ml_service
   uvicorn main:app --reload --port 8000

3. Start Frontend

   cd frontend
   npm run dev

ResearchGPT — Built for deliberate, autonomous intelligence.