Agentic Research Paper Knowledge Graph Navigator

An agentic, graph-reasoning web application designed to navigate, traverse, and discover research papers across a scientific citation network. Combining hybrid vector search, community-based topic clustering, topological graph analysis, and a multi-agent orchestration pipeline built with LangGraph & LangChain.

---

🚀 Key Features

• Real Scientific Data: Ingests thousands of papers dynamically from the OpenAlex API with full metadata, author names, venues, citations, and reconstructed abstracts.
• Dual Database Layer & Fallback: Exposes a Docker Compose setup for Neo4j (Graph) and PostgreSQL + pgvector (Relational & Semantic), with a seamless SQLite + NetworkX + Numpy in-memory/file-based fallback if Docker is inactive.
• Topological Graph Insights: Calculates PageRank centrality (to find foundational papers), betweenness centrality (to detect bridge papers connecting fields), and Louvain community detection (for clustering).
• Multi-Agent Reasoning (LangGraph):
  • Router/Orchestrator Agent: Handles intent classification and ambiguity checks (asks clarifying questions first).
  • Research Planner Agent: Formulates execution objectives.
  • Retriever Agent: Executes hybrid vector + keyword search ranked by a multi-factor score.
  • Graph Insight Agent: Extracts network topologies and clusters.
  • Reading Path Agent: Computes chronological/dependency-based reading roadmaps.
  • Idea Agent: Proposes novel research gap hypotheses backed by graph metrics.
  • Verifier Agent (Critic): Fact-checks agent statements against raw paper abstracts.
  • Responder Agent: Formulates final grounding outputs.
• Interactive D3 Force Graph UI: A clean, research-oriented dashboard visualizing paper nodes (sized by citations, colored by topic) and animating links along citation paths.

---

🛠️ Tech Stack

• Frontend: Next.js (React), HTML5, TypeScript, Tailwind CSS, Lucide React, react-force-graph-2d (D3 Canvas).
• Backend: FastAPI, Uvicorn, Python 3.11.
• Orchestration: LangGraph, LangChain, OpenAI / DeepSeek APIs.
• Database: PostgreSQL + pgvector & Neo4j (Docker) / SQLite & NetworkX (Fallback).
• Embeddings: SentenceTransformers (all-MiniLM-L6-v2 - 384 dimensions, local).

---

⚙️ Installation & Setup

1. Configure Environment variables

Create a .env file in the root directory and configure your API key (the system supports DeepSeek by default, compatible with OpenAI SDK). Ensure your OPENAI_API_KEY is defined.

2. Setup Python Backend Environment

pip install -r requirements.txt

3. Run Ingestion Pipeline (Seeding)

Fetch and embed 2,500 agent-related research papers from OpenAlex (computes embeddings locally on CPU):

python backend/scripts/seed_data.py 2500

---

🏃 Running the Application

Option A: Local Dev Runs (SQLite & NetworkX Fallback)

This is the fastest, zero-config method.

1. Start Backend API Server:

   uvicorn backend.main:app --host 0.0.0.0 --port 8000

2. Start Frontend Web Client (Next.js):

   cd frontend
   npm run dev

3. Open browser at: http://localhost:3000.

Option B: Docker Stack (Postgres + Neo4j)

1. Spin up databases:

   docker compose up -d

2. Once databases boot, run the seeding script python backend/scripts/seed_data.py 2500. It will automatically detect Postgres and Neo4j, populate them, and start Uvicorn/Next.js.

💡 Example Queries

• Foundational search: "What are the most important papers on browser agents?"
• Comparison: "Compare Yao et al., 2023 [ReAct] with Toolformer."
• Roadmap: "I want to learn web agents. Give me a reading path."
• Research gap: "Suggest a new research direction based on agent evaluation limitations in the corpus."