README.md

metadata

title: Graph RAG
emoji: 🧠
colorFrom: blue
colorTo: purple
sdk: docker
app_port: 7860
pinned: false
license: mit

CORTEX — Agentic Graph RAG Platform

CORTEX is a production-grade, agentic Knowledge Graph platform that transforms unstructured documents and web content into an intelligent, queryable knowledge graph — with a full-featured React UI, streaming AI chat, real-time graph visualization, simulation personas, and deep ontology governance.

🌐 Live Demo: https://webcortex.netlify.app

---

✨ What's Been Built

🖥️ Full-Stack Application

Layer	Stack
Backend API	FastAPI (async) + Python 3.12
Task Queue	Celery + Redis
Graph + Vector DB	Neo4j 5.x (unified)
LLM Layer	OpenAI, Anthropic, Google Gemini, Ollama
Frontend	React 18 + TypeScript + Vite
Unified Start	npm run rag (concurrently launches all 3 processes)

---

🚀 Features

📥 Document Ingestion Pipeline

• Multi-format ingestion: PDF, TXT, MD, DOCX, CSV, XLSX, PPTX, JSON
• Web scraping: Single-page scrape via POST /api/documents/scrape
• Deep web crawling: Multi-depth Playwright-powered crawler (POST /api/documents/crawl) via Crawl4AI
• Async Celery workers: Upload returns instantly with a task_id; background workers build the graph
• Re-ingest: Admin can trigger re-processing of any stored document
• Document preview & download: In-browser preview of text/Markdown; PDF download via API

🔭 Ontology Management

• Auto-generation: LLM analyzes document chunks to propose entity types & relationship types
• LLM-powered refinement: POST /api/ontology/refine — refine schema with optional human feedback
• Versioning: Each schema change bumps the version (v1.0 → v1.1, etc.)
• Document-scoped stats: /api/ontology/stats?document_id=... returns entity/relationship breakdowns for a specific document
• Visual editor: Ontology view in UI with editable entity types and relationship types
• Ontology Drift Detection: Automated drift detection compares live graph against new chunk samples; exposes pending/approved/rejected drift reports with admin approve/reject workflow

🤖 Agentic Retrieval System

• LangGraph orchestration: State-machine ReACT agent with multi-step reasoning and fallback mechanisms
• Tool routing: Dynamically selects from Naive Vector, Hybrid, Metadata Filtering, Global Community Search, HippoRAG, Graph of Thoughts (GoT), and Cypher.
• Retrieval Modes: Switch seamlessly between AUTO, HYBRID, HIPPO-RAG, LOCAL GRAPH, GLOBAL COMMUNITY, GOT, CYPHER, NAIVE, and SIMULATION.
• Streaming responses: Server-Sent Events (SSE) with real-time reasoning steps surfaced in the UI
• Multi-turn conversations: Persistent conversation threads stored in Neo4j, per-user
• Document-scoped queries: Filter retrieval to a specific document via document_id
• Graph of Thoughts (GoT): Advanced multi-hop reasoning mode exploring the graph neighborhood
• HippoRAG: Personalized PageRank (PPR) fallback strategies mapped to entity seeding
• Global Community Search: Vector similarity search over pre-computed hierarchical Leiden reports
• LLM-as-a-Judge (inline): Optional per-response quality scoring with hallucination risk, grounded/ungrounded claims, and confidence reasoning displayed in chat
• Confidence display: Confidence score, hallucination risk, and judge reasoning shown directly in the chat bubble

📊 RAGAS Evaluation & Quality Dashboard

• POST /api/eval/score: Run RAGAS-style evaluation on any Q&A pair (faithfulness, relevancy, context precision, hallucination detection)
• GET /api/eval/dashboard: Aggregate evaluation history — avg scores, hallucination rate, trend timeline
• Results persisted in Neo4j for longitudinal quality tracking

🗺️ Graph Intelligence

• D3 force-directed visualization: Interactive knowledge graph with zoom, pan, node selection, and a details modal
• Graph Export: Export full or document-scoped graph as JSON, Cypher, or GraphML
• Community Detection: Hierarchical Leiden community clustering with POST /api/graph/communities/assign
• Community listing: GET /api/graph/communities — top communities by entity count
• Temporal Queries: GET /api/entities/{entity_name}/at-time — retrieve entity relationships at a historical point in time
• Semantic Entity Deduplication: Multi-stage entity resolution with configurable similarity thresholds (POST /api/entities/deduplicate)
• Entity Enrichment: LLM-synthesized profile summaries for every entity, stored as e.summary (POST /api/entities/enrich)
• Entity Chat (scoped): POST /api/entities/{entity_name}/chat — multi-turn conversation scoped entirely to a single entity's graph neighborhood
• Graph Memory Updater: Push raw text directly into the live knowledge graph without re-ingesting a document (POST /api/graph/update)

📝 Analytical Report Agent (ReACT)

• POST /api/report: ReACT multi-step report agent using InsightForge / PanoramaSearch / QuickSearch tools
• Decomposes topic into sub-questions → retrieves graph data → synthesizes sections → compiles structured markdown report
• Exposed in the Insights view (copy/download report as Markdown)

🎭 Simulation & Persona Engine

• Persona generation: Celery task that generates personas from graph entities (POST /api/v1/simulation/generate_personas)
• Simulation ticks: Background tick loop (POST /api/v1/simulation/tick)
• Live persona interview: POST /api/v1/simulation/interview — roleplay chat with any graph entity injecting their Neo4j memory as system context
• SimulationRunView: Dedicated UI view for managing and interacting with simulation personas

🛡️ Admin Dashboard

• System statistics: Node count, relationship count, LLM provider, environment
• User management: List users, update scopes/roles (RBAC)
• Document vault: View and delete all ingested documents
• Graph CRUD: Search, inspect, and delete graph nodes from the admin panel
• Ontology governance: Review and approve/reject pending ontology proposals
• Celery task monitor: View active and reserved tasks from the admin panel
• Self-demotion guard: Admins cannot demote their own account
• Re-ingest button: Re-queue any stored document from the document vault
• User activity metrics: Per-user conversation count, message count, last active timestamp

🔐 Authentication & Security

• JWT authentication: Token-based auth with configurable expiry
• RBAC scopes: read, write, admin scopes enforced per endpoint
• User registration: POST /api/auth/register
• Pydantic validation: All API inputs validated at the model layer
• Cypher injection prevention: Schema validation and query whitelisting
• File upload limits: File size and MIME type enforcement

🌐 Frontend (React/TypeScript)

Seven fully implemented views accessible from the CORTEX top navigation bar:

Route	View	Description
/	Home	Animated stats dashboard — documents, entities, relationships, graph health
/process	Process	Upload files or scrape/crawl URLs; view ingestion queue and document list
/ontology	Ontology	View/edit the live ontology schema; run LLM refinement; inspect entity/relationship stats per doc
/interact	Interact	Streaming AI chat with reasoning steps, confidence, hallucination risk; conversation history
/simulate	Simulate	Simulation persona management and live interview interface
/insights	Insights	Topic-driven analytical report generation with copy/download
/admin	Admin (admin-only)	Full admin panel for users, docs, tasks, ontology governance

🔭 Observability

• OpenTelemetry: Distributed tracing (silenced from console; configured for export)
• Health check: GET /api/system/health — Neo4j, Redis, Celery worker status
• System stats: GET /api/system/stats — document, entity, relationship, chunk counts
• User stats: GET /api/system/my-stats — per-user conversation and message activity

---

🏗️ Architecture

┌─────────────────────────────────────────────────────────────────────────────┐
│                          React Frontend (CORTEX)                             │
│    Home │ Process │ Ontology │ Interact │ Simulate │ Insights │ Admin        │
└─────────────────────────────┬───────────────────────────────────────────────┘
                              │ HTTP / SSE
┌─────────────────────────────▼───────────────────────────────────────────────┐
│                     FastAPI Gateway (port 8000)                              │
│          JWT Auth · RBAC Scopes · CORS · OpenTelemetry                      │
└──────┬──────────────────────┬──────────────────────┬────────────────────────┘
       │                      │                      │
┌──────▼──────┐   ┌───────────▼──────────┐  ┌───────▼────────────────────┐
│  Ingestion  │   │  ReACT Agent System  │  │  Report Agent (ReACT)      │
│  Pipeline   │   │  - Vector Search     │  │  - InsightForge            │
│  - Parser   │   │  - Graph Traversal   │  │  - PanoramaSearch          │
│  - Ontology │   │  - Cypher Gen (GoT)  │  │  - QuickSearch             │
│  - Extractor│   │  - Community Search  │  │  - Markdown output         │
│  - Web      │   │  - Temporal Queries  │  └────────────────────────────┘
│    Crawler  │   │  - LLM-as-a-Judge    │
└──────┬──────┘   └─────────┬────────────┘
       │                    │
┌──────▼────────────────────▼──────────────────┐
│              Neo4j 5.x Database               │
│  Entities · Chunks · Relationships ·          │
│  Vector Index · Conversations ·               │
│  EvalResults · DriftReports · Users           │
└───────────────────────────────────────────────┘
       │
┌──────▼──────────────────────┐
│  Celery Workers (Redis)     │
│  - Async document ingestion │
│  - Persona generation       │
│  - Simulation ticks         │
└─────────────────────────────┘

---

📦 Project Structure

graph-RAG/
├── src/graph_rag_service/
│   ├── api/
│   │   ├── server.py          # Main FastAPI app + all API routes (1900 lines)
│   │   ├── auth.py            # JWT auth + RBAC helpers
│   │   ├── admin.py           # Admin sub-router
│   │   ├── simulation.py      # Simulation / persona interview router
│   │   └── models.py          # All Pydantic request/response models
│   ├── core/
│   │   ├── abstractions.py    # Abstract base classes (GraphStore, VectorStore, LLMProvider)
│   │   ├── models.py          # Domain data models
│   │   ├── neo4j_store.py     # Full Neo4j implementation (graph + vector)
│   │   ├── llm_factory.py     # Multi-LLM provider factory + UnifiedLLMProvider
│   │   ├── entity_resolver.py # Semantic entity deduplication
│   │   └── storage.py         # File storage abstraction
│   ├── ingestion/
│   │   ├── pipeline.py        # End-to-end ingestion orchestrator
│   │   ├── document_processor.py  # Multi-format document parsing
│   │   ├── ontology_generator.py  # LLM ontology generation + refinement
│   │   ├── extractor.py       # Entity + relationship extraction
│   │   ├── web_crawler.py     # Playwright-based deep web crawler (Crawl4AI)
│   │   └── persona_generator.py   # Simulation persona generation
│   ├── retrieval/
│   │   ├── agent.py           # LangGraph ReACT retrieval agent
│   │   ├── tools.py           # Retrieval tools + RAGEvaluator (RAGAS)
│   │   └── report_agent.py    # ReACT analytical report agent
│   ├── services/
│   │   ├── graph_memory_updater.py   # Push raw text → live graph
│   │   ├── entity_enricher.py        # LLM entity profile summaries
│   │   └── ontology_drift_detector.py # Automated schema drift detection
│   ├── workers/
│   │   └── celery_worker.py   # Celery app + ingest_document_task
│   ├── observability/
│   │   └── tracing.py         # OpenTelemetry setup (console suppressed)
│   ├── config.py              # Pydantic settings (all env vars)
│   └── main.py                # Uvicorn entry point
├── frontend-react/
│   └── src/
│       ├── views/
│       │   ├── Home.tsx            # Animated stats dashboard
│       │   ├── Process.tsx         # Document upload + URL scrape/crawl
│       │   ├── Ontology.tsx        # Schema editor + stats
│       │   ├── InteractionView.tsx # Streaming chat + conversation history
│       │   ├── SimulationRunView.tsx # Persona simulation UI
│       │   ├── InsightsView.tsx    # Report generation + copy/download
│       │   ├── AdminDashboard.tsx  # Full admin panel
│       │   └── Login.tsx           # Login page
│       ├── components/
│       │   └── GraphCanvas.tsx     # D3 force-directed graph + node modal
│       ├── context/
│       │   └── AuthContext.tsx     # JWT auth context + hooks
│       └── App.tsx                 # Router + top-nav (CORTEX branding)
├── tests/                      # Test suite
├── data/uploads/               # Uploaded documents (local storage)
├── .env.example                # All configurable environment variables
├── pyproject.toml              # Python project + uv dependencies
├── package.json                # Unified start scripts (npm run rag)
├── ARCHITECTURE.md             # Detailed architecture design doc
└── QUICKSTART.md               # 5-minute quick start guide

---

⚡ Quick Start

Prerequisites

• Python 3.12+
• Node.js 18+
• Neo4j 5.x (running) with APOC and Graph Data Science (GDS) plugins installed
• Redis (running)
• Ollama (optional, for local LLMs)

1. Clone & Install

git clone <repository-url>
cd graph-RAG

# Installs Python deps (uv), frontend (npm), and Playwright Chromium
npm install

2. Configure Environment

cp .env.example .env
# Fill in NEO4J_URI, NEO4J_PASSWORD, and your LLM API keys

3. Start Neo4j (Requires GDS Plugin)

docker run -d --name neo4j \
  -p 7474:7474 -p 7687:7687 \
  -e NEO4J_AUTH=neo4j/password \
  -e NEO4J_PLUGINS='["graph-data-science", "apoc"]' \
  neo4j:5.18.0

4. Start Redis

docker run -d --name redis -p 6379:6379 redis:alpine

Hugging Face Spaces / All-in-One Docker

You can easily deploy CORTEX to Hugging Face Spaces or as a standalone container. Our Dockerfile automatically installs Python 3.12, Redis, Neo4j, and the GDS plugin to run the entire platform on a single port (7860).

5. Launch Everything

npm run rag

This starts three color-coded processes concurrently:

Process	URL
API Server	http://localhost:8000
API Docs	http://localhost:8000/docs
React Frontend	http://localhost:5173

Default credentials: admin / admin

---

🔑 Environment Variables

Copy .env.example to .env and configure:

# Neo4j
NEO4J_URI=bolt://localhost:7687
NEO4J_USER=neo4j
NEO4J_PASSWORD=password

# Redis
REDIS_HOST=localhost
REDIS_PORT=6379

# LLM Provider (openai | anthropic | gemini | ollama)
DEFAULT_LLM_PROVIDER=gemini
GOOGLE_API_KEY=your-key-here

# Optional: OpenAI / Anthropic
OPENAI_API_KEY=sk-...
ANTHROPIC_API_KEY=sk-ant-...

# Optional: Ollama (local)
OLLAMA_BASE_URL=http://localhost:11434
OLLAMA_MODEL=deepseek-r1:7b
OLLAMA_EMBEDDING_MODEL=nomic-embed-text

# Feature flags
ENABLE_LLM_JUDGE=true

# Security
SECRET_KEY=change-this-in-production
ACCESS_TOKEN_EXPIRE_MINUTES=1440

---

🌐 API Reference

Authentication

Method	Endpoint	Description
POST	/api/auth/register	Register new user
POST	/api/auth/login	Login → JWT token
GET	/api/auth/me	Get current user info

Documents

Method	Endpoint	Description
POST	/api/documents/upload	Upload file (PDF, DOCX, TXT, MD, CSV, XLSX, PPTX, JSON)
POST	/api/documents/scrape	Scrape single URL → ingest
POST	/api/documents/crawl	Deep multi-page Playwright crawl → ingest (API Only)
GET	/api/documents	List all ingested documents
DELETE	/api/documents/{id}	Delete document + graph chunks
GET	/api/documents/{id}/download	Download source file
GET	/api/documents/{id}/preview	Preview text content
GET	/api/documents/status/{task_id}	Ingestion task status

Query & Chat

Method	Endpoint	Description
POST	/api/query	Agentic query (streaming or JSON); supports document_id, use_got
GET	/api/conversations	List conversation threads
GET	/api/conversations/{id}	Get conversation + messages
DELETE	/api/conversations/{id}	Delete conversation

Ontology

Method	Endpoint	Description
GET	/api/ontology	Get current ontology
PUT	/api/ontology	Update ontology (admin)
POST	/api/ontology/refine	LLM-powered ontology refinement
GET	/api/ontology/stats	Entity/relationship counts (optional doc filter)
POST	/api/ontology/drift/detect	Trigger drift detection
GET	/api/ontology/drift	List drift reports
POST	/api/ontology/drift/{id}/approve	Approve drift → merge into ontology
POST	/api/ontology/drift/{id}/reject	Reject drift report

Graph

Method	Endpoint	Description
GET	/api/graph/visualization	Graph nodes + edges for D3 rendering
GET	/api/graph/export	Export graph (json | cypher | graphml)
POST	/api/graph/update	Push raw text → merge into live graph
POST	/api/graph/communities/assign	Run Hierarchical Leiden clustering & generate reports
GET	/api/graph/communities	List top communities

Entities

Method	Endpoint	Description
POST	/api/entities/deduplicate	Semantic entity resolution + merge
POST	/api/entities/enrich	Generate LLM summaries for all entities
GET	/api/entities/{name}/summary	Get enriched entity profile
POST	/api/entities/{name}/chat	Multi-turn entity-scoped chat
GET	/api/entities/{name}/at-time	Temporal query (ISO 8601 date)

Reports & Evaluation

Method	Endpoint	Description
POST	/api/report	Generate ReACT analytical report (markdown)
POST	/api/eval/score	RAGAS evaluation of a Q&A pair
GET	/api/eval/dashboard	Evaluation history dashboard

Simulation

Method	Endpoint	Description
POST	/api/v1/simulation/interview	Live persona interview (in-character LLM)
GET	/api/v1/simulation/report	Sandbox analytical report (API Only)
POST	/api/v1/simulation/generate_personas	Queue persona generation task (API Only)
POST	/api/v1/simulation/tick	Advance simulation tick (API Only)

System & Admin

Method	Endpoint	Description
GET	/api/system/health	Neo4j + Redis + Celery health
GET	/api/system/stats	Document, entity, relationship counts
GET	/api/system/my-stats	Current user's activity stats
GET	/api/system/formats	Supported ingestion file formats
GET	/api/admin/stats	Admin-only system stats
GET	/api/admin/users	List all users
PUT	/api/admin/users/{username}/role	Update user scopes
GET	/api/admin/tasks	View Celery tasks
GET	/api/admin/documents	Admin document vault
POST	/api/admin/documents/{id}/reingest	Re-queue document for ingestion
GET	/api/admin/graph/nodes	Search graph nodes
DELETE	/api/admin/graph/nodes/{id}	Delete a graph node

---

🧪 Testing & Benchmarking

# Run tests
uv run pytest

# With coverage
uv run pytest --cov=src/graph_rag_service

# Run SOTA Benchmarks against Hugging Face datasets (e.g. HotpotQA, MuSiQue)
uv run python benchmarks/run_benchmark.py

---

🚀 Production Deployment

Process	Command
API Server	uv run python main.py
Celery Worker	uv run celery -A src.graph_rag_service.workers.celery_worker worker --loglevel=info --concurrency=4 --pool=threads
React Build	cd frontend-react && npm run build

The built React assets can be served directly by FastAPI (static file mount), or deployed to a CDN separately. Neo4j and Redis can be run via Docker, managed cloud services (AuraDB, Redis Cloud), or self-hosted.

---

📄 Additional Documentation

• ARCHITECTURE.md — Deep dive into the system design, data flow, and component interactions
• QUICKSTART.md — 5-minute environment setup guide
• /docs — Interactive Swagger UI (auto-generated from FastAPI)

---

Project Status: Production-grade MVP · Actively developed
License: Proprietary — all rights reserved