LLM Exo-Graph 🧠🕸️

An advanced knowledge graph engine that externalizes LLM memory into Neo4j, creating a persistent, searchable brain for AI systems.

🌟 Why Exo-Graph?

Traditional LLMs have ephemeral memory. LLM Exo-Graph creates an exocortex - an external brain that:

• 📝 Persists knowledge across sessions
• 🔍 Searches with both semantic and graph algorithms
• 🧩 Connects information through relationships
• ⚡ Scales beyond context window limitations

🎯 The Power of Graph Structure

Subject → Relationship → Object = Triplet(metadata)

Our graph structure captures not just entities, but the rich context of their relationships:

God → CREATED → man = (summary: God created man in his own image) [conf: 0.90]
God → DIVIDED → waters = (summary: God divided the waters) [conf: 0.90]
light → EXISTS → light = (summary: there was light) [conf: 0.90]

Benefits of This Approach

1. Enhanced Graph Search

   • Traverse relationships with Cypher queries
   • Find indirect connections (friend-of-friend)
   • Discover patterns and clusters

2. Superior Vector Search

   • Summaries provide rich semantic context
   • Embeddings capture relationship meaning
   • Hybrid search combines graph + semantic

3. Temporal Intelligence

   • Track relationship changes over time
   • Handle contradictions gracefully
   • Maintain complete history

🏗️ How It Works

Entity Extraction Pipeline

graph LR
    A[Natural Language Input] --> B[LLM Processor]
    B --> C{Entity Extraction}
    C --> D[Subject Recognition]
    C --> E[Relationship Detection]
    C --> F[Object Identification]
    D --> G[Graph Edge Creation]
    E --> G
    F --> G
    G --> H[Neo4j Storage]
    G --> I[Vector Embedding]
    I --> J[Semantic Index]

Loading

Entity Standardization Process

graph TD
    A[Raw Entity/Relationship] --> B[BiEncoder Embedding]
    B --> C[Category Classification]
    C --> D{Similarity Check}
    D -->|High Similarity| E[Use Existing Standard]
    D -->|Low Similarity| F[CrossEncoder Verification]
    F --> G{Cross-Validation Score}
    G -->|Score > Threshold| H[Merge with Standard]
    G -->|Score < Threshold| I[Create New Standard]
    E --> J[Standardized Output]
    H --> J
    I --> J
    
    K[Existing Categories] --> C
    L[Cached Embeddings] --> D
    
    style B fill:#e1f5fe,stroke:#01579b,stroke-width:2px
    style F fill:#f3e5f5,stroke:#4a148c,stroke-width:2px
    style J fill:#e8f5e8,stroke:#2e7d32,stroke-width:3px

Loading

Item Processing Workflow

graph TD
    A[InputItem] --> B[LLM Entity Extraction]
    B --> C[Standardization Process]
    C --> D{Negation Detection}
    
    D -->|Positive Statement| E[Duplicate Check]
    D -->|Negation| F[Conflict Detection]
    
    E -->|New Relationship| G[Create Edge]
    E -->|Duplicate Found| H[Skip/Ignore]
    
    F -->|Conflict Found| I[Temporal Resolution]
    F -->|No Conflict| J[Log Error]
    
    G --> K[Neo4j Storage]
    I --> L[Obsolete Existing]
    L --> M[Update Metadata]
    
    K --> N[Vector Embedding]
    M --> N
    N --> O[Index Update]
    
    P[Temporal Metadata] --> G
    P --> I
    Q[Confidence Scoring] --> G
    Q --> I
    
    style D fill:#fff3e0,stroke:#e65100,stroke-width:2px
    style I fill:#f3e5f5,stroke:#4a148c,stroke-width:2px
    style N fill:#e8f5e8,stroke:#2e7d32,stroke-width:2px
    style J fill:#ffebee,stroke:#c62828,stroke-width:2px

Loading

Key Processing Features:

1. 🔍 Standardization: Entities and relationships are normalized using BiEncoder + CrossEncoder
2. ⚠️ Negation Handling: "Alice no longer works at Google" → obsoletes existing relationship
3. ⏰ Temporal Resolution: Automatic conflict resolution with date-based transitions
4. 🎯 Confidence Scoring: Each relationship has confidence metadata for reliability
5. 🔄 Duplicate Prevention: Exact matches are detected and skipped
6. 📊 Vector Integration: All changes immediately update semantic search indexes

🚀 Quick Start

Prerequisites

# Using Docker (Recommended)
docker-compose up -d

# Or use Neo4j Cloud
# Set NEO4J_URI=neo4j+s://your-instance.neo4j.io

Installation

From PyPI (Recommended):

pip install llm-exo-graph

From Source:

git clone https://github.com/your-org/llm-exo-graph
cd llm-exo-graph
pip install -e .

With Optional Dependencies:

# For document processing
pip install "llm-exo-graph[documents]"

# For development
pip install "llm-exo-graph[dev]"

# All features
pip install "llm-exo-graph[all]"

Basic Usage

from llm_exo_graph import ExoGraphEngine, InputItem

# Initialize with auto-configuration
engine = ExoGraphEngine()

# Or with custom encoder models
config = {
    "encoder_model": "all-mpnet-base-v2",
    "cross_encoder_model": "cross-encoder/ms-marco-MiniLM-L-12-v2"
}
engine = ExoGraphEngine(config=config)

# Feed knowledge
engine.process_input([
   InputItem("Marie Curie discovered radium in 1898"),
   InputItem("Radium glows green in the dark"),
   InputItem("Marie Curie won the Nobel Prize twice")
])

# Query naturally
response = engine.search("What did Marie Curie discover?")
print(response.answer)
# → "Marie Curie discovered radium in 1898."

🤖 MCP Integration (Model Context Protocol)

What is MCP?

MCP enables AI assistants like Claude to directly interact with your knowledge graph via Server-Sent Events (SSE), creating a persistent memory layer that survives across conversations.

Quick Setup with Docker

1. Start the MCP Server

   # Use the notebook docker-compose for MCP development
   docker-compose -f docker-compose.notebook.yml up -d
   
   # This starts:
   # - Neo4j on port 7687/7474
   # - MCP SSE server on port 3000

2. Configure Claude Desktop

   // ~/Library/Application Support/Claude/claude_desktop_config.json
   {
     "mcpServers": {
       "exo-graph": {
         "command": "npx",
         "args": [   
             "-y",
             "mcp-remote",
             "http://localhost:3000/sse",
             "--allow-http"
         ]
       }  
     }
   }

3. Restart Claude Desktop - The MCP server will connect automatically

Graph Data Examples

After setup, Claude can work with rich graph relationships like these from our Biblical knowledge graph:

God → CREATED → man (God created man in his own image) [conf: 0.90]
God → DIVIDED → waters (God divided the waters) [conf: 0.90]  
light → EXISTS → light (there was light) [conf: 0.90]
God → SAID → "Let there be light" (God spoke creation into existence) [conf: 0.95]
man → MADE_IN_IMAGE_OF → God (humanity reflects divine nature) [conf: 0.85]
waters → SEPARATED_BY → firmament (division of waters above and below) [conf: 0.88]

Using MCP in Claude

Once configured, Claude gains persistent memory and can:

💾 Store Knowledge Permanently

Claude: "I'll remember that John works at OpenAI as a researcher"
→ Creates: John → WORKS_AT → OpenAI (researcher role) [conf: 0.95]

🔍 Query Across Sessions

User: "What did we discuss about John yesterday?"
Claude: "You told me John works at OpenAI as a researcher. I have that stored in the knowledge graph."

🔗 Discover Connections

User: "How is John connected to AI research?"
Claude: "Through the knowledge graph, I can see John → WORKS_AT → OpenAI → FOCUSES_ON → AI Research"

📊 Analyze Patterns

User: "Show me all employment relationships you know about"
Claude: "I found 15 employment relationships in the graph, including John at OpenAI, Alice at Google..."

⏰ Track Changes Over Time

User: "John left OpenAI and joined Google"
Claude: "I've updated the graph - obsoleted John's OpenAI relationship and created a new Google relationship with today's date."

🌐 REST API

Quick API Usage

# Start API server
cd kg_api_server
python app/main.py

# Add knowledge
curl -X POST http://localhost:8080/api/v1/process \
  -H "Content-Type: application/json" \
  -d '{"items": [{"description": "Einstein developed E=mc²"}]}'

# Search
curl http://localhost:8080/api/v1/search?query=Einstein

API Endpoints

• POST /api/v1/process - Add knowledge
• GET /api/v1/search - Natural language search
• GET /api/v1/entities/{name} - Get entity details
• DELETE /api/v1/edges/{id} - Remove relationships

📊 Visualization

Generate beautiful graph visualizations:

python visualize_graph.py

Creates three outputs in /output:

• 📄 knowledge_graph_relationships.txt - Human-readable relationships
• 🖼️ knowledge_graph_static.png - Publication-ready visualization
• 🌐 knowledge_graph_interactive.html - Interactive exploration

🔧 Configuration

Engine Configuration

from llm_exo_graph import ExoGraphEngine, Neo4jConfig, OllamaConfig

# Custom encoder configuration
config = {
    "encoder_model": "all-mpnet-base-v2",           # BiEncoder model
    "cross_encoder_model": "cross-encoder/ms-marco-MiniLM-L-12-v2"  # CrossEncoder model
}

# Initialize with all configurations
engine = ExoGraphEngine(
    llm_config=OllamaConfig(model="llama3.2"),
    neo4j_config=Neo4jConfig(),
    config=config
)

Available Encoder Models

BiEncoder Models (for semantic embeddings):

• all-MiniLM-L6-v2 (default) - Fast, good quality
• all-mpnet-base-v2 - Higher quality, slower
• sentence-transformers/all-MiniLM-L12-v2 - Balanced

CrossEncoder Models (for relationship validation):

• cross-encoder/ms-marco-MiniLM-L-6-v2 (default) - Fast
• cross-encoder/ms-marco-MiniLM-L-12-v2 - More accurate
• cross-encoder/ms-marco-electra-base - Highest accuracy

Environment Variables

# LLM Configuration (auto-detected)
OPENAI_API_KEY=sk-...           # For OpenAI
OLLAMA_BASE_URL=http://localhost:11434  # For Ollama
OLLAMA_MODEL=llama3

# Neo4j Configuration
NEO4J_URI=bolt://localhost:7687
NEO4J_USERNAME=neo4j
NEO4J_PASSWORD=password
NEO4J_DATABASE=neo4j

# Optional
LOG_LEVEL=INFO

📚 Advanced Features

Custom Model Configuration

Choose encoder models based on your needs:

# High Performance Setup (Fast processing)
fast_config = {
    "encoder_model": "all-MiniLM-L6-v2",
    "cross_encoder_model": "cross-encoder/ms-marco-MiniLM-L-6-v2"
}

# High Accuracy Setup (Better quality)
accurate_config = {
    "encoder_model": "all-mpnet-base-v2", 
    "cross_encoder_model": "cross-encoder/ms-marco-MiniLM-L-12-v2"
}

# Domain-Specific Setup (for scientific/technical content)
domain_config = {
    "encoder_model": "sentence-transformers/allenai-specter",
    "cross_encoder_model": "cross-encoder/ms-marco-electra-base"
}

engine = ExoGraphEngine(config=accurate_config)

Document Processing

from llm_exo_graph import DocumentProcessor

processor = DocumentProcessor()
results = processor.process_directory("./research_papers/")

Temporal Relationships & Negation Handling

# Example: Career transitions with temporal intelligence
engine.process_input([
    InputItem("Alice works as a software engineer at Google"),
    InputItem("Alice no longer works at Google"),  # Negation - obsoletes previous
    InputItem("Alice started working at OpenAI in January 2024")  # New relationship
])

# The system automatically:
# 1. Detects "no longer" as negation
# 2. Finds conflicting relationships
# 3. Obsoletes old relationship with end date
# 4. Creates new relationship with start date

Standardization in Action

# These variations are automatically standardized:
engine.process_input([
    InputItem("John works at Microsoft"),
    InputItem("John is employed by Microsoft"),  # Standardized to "WORKS_AT"
    InputItem("John's employer is Microsoft"),   # Also standardized to "WORKS_AT"
])

# Result: All create the same standardized relationship
# John → WORKS_AT → Microsoft (with different summaries)

Conflict Resolution

# Handles contradictions intelligently
history = engine.get_entity_relationships("Alice")
# Shows both relationships with temporal metadata:
# - Alice → WORKS_AT → Google [obsolete: 2024-01-15]
# - Alice → WORKS_AT → OpenAI [active: 2024-01-16]

🧪 Examples

• 📖 Bible Knowledge Graph
• 🧬 Bio Research Graph
• 📄 Document Processing
• 🔗 API Integration

🛠️ Development

Running Tests

pytest tests/
cd kg_api_server && pytest tests/

Contributing

See CONTRIBUTING.md

📈 Performance

• ⚡ 50-74% faster queries with optimizations
• 🔄 Batch processing for large datasets
• 💾 Intelligent caching layers
• 🎯 Optimized Neo4j indexes

📦 Package Information

• PyPI: https://pypi.org/project/llm-exo-graph/
• Install: pip install llm-exo-graph
• Version: Check latest on PyPI
• Extras: [documents], [dev], [all]

🤝 Community

• 📖 Documentation
• 🐛 Issues
• 💬 Discussions
• 📦 PyPI Package

📝 License

MIT License - see LICENSE

---

LLM Exo-Graph - Giving AI a persistent, searchable memory 🧠✨