Overview

Nexarag is an open-source platform for building knowledge graphs from research papers and querying them with AI, enabling transparent and reproducible literature analysis without the hallucinations of traditional RAG systems. Deploy locally with full privacy control or integrate with any LLM via the standardized Model Context Protocol (MCP).

See also: * Usage and Development Guides * Contribution Guidelines

Feedback

We’d love your feedback on Nexarag, including bug reports, feature requests, documentation fixes, and tutorial or training-material ideas.

• Bugs and feature requests: open an issue on GitHub: https://github.com/KevinMoonLab/Nexarag/issues
  Please include:

  • what you expected vs. what happened
  • steps to reproduce (or a minimal example)
  • OS + Docker version (and GPU details if relevant)
  • relevant logs or screenshots

• Private or security-sensitive reports: email us at nexarag.ai@gmail.com
  If possible, include the same details as above and note why it should not be posted publicly.

Test Coverage

• Manual Test Coverage Report
• Backend Automated Testing
• Frontend Automated Testing

Pre-Requisites

• Docker
• (Windows Only) WSL2
• (MacOS Only) Ollama Desktop
• (Optional) Claude Desktop (for MCP)

Quick Start

Follow the instructions below or visit our YouTube channel for detailed tutorials:

• MacOS Deployment and MCP Configuration
• WSL + GPU Deployment

1. Download Docker compose file

Choose a Dockerfile compatible with your OS and hardware:

• MacOS
• Linux/WSL (CPU)
• Linux/WSL (GPU)

[Optional]: Move the Docker compose file to a location on your drive, e.g. ~/Nexarag.

2. Run Nexarag

From the same directory as the downloaded docker-compose.yml, run: bash docker compose up -d Visit Nexarag in your browser at http://localhost:5000 (or 5100 on MacOS).

3. Pull Ollama Models

To support all internal features, Nexarag requires:

• An embedding model, such as nomic-embed-text:v1.5
• A language model, such as gemma3:1b
• An MCP-capable model, such as qwen3:8b

Browse the full library of Ollama models here and choose any model from these families that your hardware supports. Defaults that run on most hardware are provided below.

Linux/WSL

Models can be pulled through the command line in the ollama Docker container.

bash docker exec -it nexarag.ollama /bin/bash ollama pull nomic-embed-text:v1.5 ollama pull gemma3:1b ollama pull qwen3:8b

MacOS

Pull models directly from your command line.

bash ollama pull nomic-embed-text:v1.5 ollama pull gemma3:1b ollama pull qwen3:8b

4. (Optional) MCP Integration

Claude Desktop

Add the following to your Claude Desktop Config:

json { "mcpServers": { "nexarag": { "command": "npx", "args": [ "-y", "mcp-remote", "http://localhost:9000/mcp" ], "env": { "MCP_TRANSPORT_STRATEGY": "http-only" } } } }

Ollama

First install pipx, then run:

bash pipx install ollmcp

To start the MCP client: bash ollmcp -u http://localhost:9000/mcp -m qwen3:8b

Semantic Scholar

Please note that we are rate-limited by the Semantic Scholar API, so enriching BibTex uploads with data and updating the graph after adding papers from a Semantic Scholar search may take several minutes to complete.