SOLVRO MCP

Knowledge Graph RAG System for ToPWR
Intelligent assistant for Wrocław University of Science and Technology

Quick Start • Architecture • Features • API

┌─────────────┐     ┌─────────────┐     ┌─────────────┐     ┌─────────────┐
│  Frontend   │────▶│  ToPWR API  │────▶│  MCP Server │────▶│    Neo4j    │
│    :80      │     │    :8000    │     │    :8005    │     │    :7687    │
└─────────────┘     └─────────────┘     └─────────────┘     └─────────────┘
React + Nginx         FastAPI              FastMCP          Knowledge Graph

• PWrChat UI - React chatbot (session sidebar, dark/light mode toggle, persistent theme)
• Intelligent Query Routing - Guardrails system determines query relevance
• Natural Language to Cypher - Converts questions to graph queries
• Knowledge Graph RAG - Retrieval-Augmented Generation with Neo4j
• MCP Protocol - Standard Model Context Protocol interface
• Observability - Optional Langfuse tracing integration
• Docker Ready - One command deployment

Quick Start

# Setup
just setup
cp .env.example .env  # Edit with your API keys

# Run with Docker
just up      # Start Neo4j + MCP Server + API
just logs    # View logs
just down    # Stop services

Architecture

System Overview

┌─────────────┐     ┌─────────────┐     ┌─────────────┐     ┌─────────────┐
│  Frontend   │────▶│  ToPWR API  │────▶│  MCP Server │────▶│    Neo4j    │
│    :80      │     │    :8000    │     │    :8005    │     │    :7687    │
└─────────────┘     └─────────────┘     └─────────────┘     └─────────────┘
React + Nginx         FastAPI              FastMCP          Knowledge Graph

RAG Pipeline

The heart of the system is a LangGraph-based RAG pipeline that intelligently processes user queries:

Pipeline Flow:

1. Guardrails - Fast LLM determines if query is relevant to knowledge base
2. Cypher Generation - Accurate LLM converts natural language to Cypher query
3. Retrieval - Execute query against Neo4j knowledge graph
4. Response - Return structured context data

Data Pipeline

Separate ETL pipeline for ingesting documents into the knowledge graph:

Pipeline Steps:

1. Document Loading - PDF and text document ingestion
2. Text Extraction - OCR and content extraction
3. LLM Processing - Generate Cypher queries from content
4. Graph Population - Execute queries to build knowledge graph

Configuration

Copy .env.example to .env and configure:

########################################
# LLM / AI Provider Keys
########################################

# OpenAI API key (optional)
OPENAI_API_KEY=

# DeepSeek API key (optional)
DEEPSEEK_API_KEY=

# Google Generative AI / PaLM API key (optional)
GOOGLE_API_KEY=

# CLARIN LLM API key (optional, used by API & client)
CLARIN_API_KEY=


########################################
# Langfuse Observability
########################################

LANGFUSE_SECRET_KEY=
LANGFUSE_PUBLIC_KEY=
LANGFUSE_HOST=https://cloud.langfuse.com


########################################
# Neo4j Database
########################################

# URI used by data pipeline, MCP server and graph config
NEO4J_URI=bolt://localhost:7687
NEO4J_USER=neo4j
NEO4J_PASSWORD=


########################################
# Data Pipeline Runtime Controls
########################################

# Max parallel pages processed per batch
DATA_PIPELINE_MAX_CONCURRENCY=4

# Minutes after which a stuck in-progress hash can be reclaimed
DATA_PIPELINE_CLAIM_STALE_MINUTES=30


########################################
# MCP Server Networking
########################################

# Bind host for the MCP server process
MCP_BIND_HOST=0.0.0.0

# Host/port used by API and MCP client to reach the MCP server
MCP_HOST=127.0.0.1
MCP_PORT=8005

Commands

# Docker Stack
just up          # Start all services (including frontend at :80)
just down        # Stop services
just logs        # View logs
just ps          # Service status
just nuke        # Remove everything

# Local Development
just mcp-server  # Run MCP server
just api         # Run FastAPI
just kg "query"  # Query knowledge graph

# Frontend
just frontend-install  # Install npm dependencies
just frontend-dev      # Start dev server at :3000 (requires running API)
just frontend-build    # Build for production

# Quality
just lint        # Format & lint
just test        # Run tests
just ci          # Full CI pipeline
uv run --with pytest python -m pytest tests/data_pipeline/test_pipeline_concurrency.py -q
                # Run pipeline concurrency/idempotency tests only

# Data Pipeline
just prefect-up  # Start Prefect
just pipeline    # Run ETL

Project Structure

src/
├── mcp_server/      # MCP server + RAG pipeline
├── mcp_client/      # CLI client
├── topwr_api/       # FastAPI backend
├── config/          # Configuration
└── data_pipeline/   # Prefect ETL flows

frontend/
├── src/
│   ├── api/         # API client
│   ├── hooks/       # useUserId, useSessions, useChat, useTheme
│   ├── components/  # Sidebar, Chat, shared UI
│   └── types/       # TypeScript mirrors of backend models
└── package.json     # React + Vite + TailwindCSS

docker/
├── compose.stack.yml    # Main stack (Neo4j + MCP + API + Frontend)
├── compose.prefect.yml  # Data pipeline
├── Dockerfile.mcp       # MCP server image
├── Dockerfile.api       # FastAPI image
├── Dockerfile.frontend  # React + Nginx image
└── nginx.conf           # SPA fallback + API proxy

API Usage

Chat Endpoint

curl -X POST http://localhost:8000/api/chat \
  -H "Content-Type: application/json" \
  -d '{"user_id": "user1", "message": "Czym jest nagroda dziekana?"}'

Response:

{
  "session_id": "abc123",
  "message": "Nagroda dziekana to wyróżnienie przyznawane...",
  "metadata": {
    "source": "mcp_knowledge_graph",
    "trace_id": "xyz789"
  }
}

Session Management

# Get session history
curl http://localhost:8000/api/sessions/{session_id}/history

# List user sessions  
curl http://localhost:8000/api/users/{user_id}/sessions

Tech Stack

License

MIT © Solvro