GraphRAG_Beginning

In this repository I will be telling how to get started with GraphRAGs, I have used Neo4J for creating the Knowledge Graphs Graph RAG Setup Guide Prerequisites Ollama

Local LLM runner for executing open-source language models (e.g. Llama2, Mistral)

Installation Guide

Required for local LLM inference without cloud dependencies

Neo4J

Graph database for storing knowledge graph relationships and vector embeddings

Desktop Version: 5.22.0+ (mandatory for vector index support)

Desktop Setup Guide

• Vector indexes enable efficient similarity search on embeddings

Docker

• Containerization platform for Neo4J deployment
• Alternative to desktop app using provided docker-compose.yaml

Neo4J Configuration

Desktop Setup

1. Create new DBMS in Neo4J Desktop
2. Use version 5.22.0 or higher
3. Start database and enable Bolt connector
4. Note credentials for Python connection:
   • bolt://localhost:7687
   • Username: neo4j
   • Password: [your_password]

Also install APOC pluggin for the respective database which you would be using.

Docker Setup

I have added the docker-compose.yaml file.

Critical Dependencies Breakdown

Library	Purpose	Mathematical/Functional Relevance
langchain	Core RAG framework	Implements retrieval chains: Retrieval → Ranking → Generation
langchain-ollama	Local LLM integration	Enables llama2, mistral model inference via Ollama
neo4j	Graph database driver	Implements Cypher query execution with graph.schema validation
tiktoken	Token counting	BPE tokenization for context window management (CL100k base)
yfiles_jupyter_graphs	Graph visualization	Force-directed layout algorithms for knowledge graph display
langchain-openai	Cloud LLM fallback	Alternative to Ollama using text-embedding-ada-002 embeddings
python-dotenv	Secret management	Secure credential loading for Neo4J/OpenAI connections

Version Strategy

• --upgrade --quiet ensures latest stable versions with clean output
• Pinned versions in requirements.txt recommended for production

Why This Matters for Graph RAG:

• Hybrid Retrieval

1. neo4j handles graph-aware retrieval (relationship traversal)
2. langchain-openai provides vector similarity (cosine distance in embedding space)

• Local Inference Stack

1. langchain-ollama + llama2 enables private document processing
2. 50% reduction in cloud API costs compared to GPT-4

• Debugging Tools

1. yfiles_jupyter_graphs visualizes graph structure using Fruchterman-Reingold layout
2. json-repair handles malformed LLM outputs (common in complex queries)

Cosine Similarity: Core Concept

Mathematical Definition Measures similarity between two vectors by calculating the cosine of the angle between them:

Similarity = cos(θ) = (A · B) / (||A|| ||B||)

Output Range: -1 (opposite) to 1 (identical). In text embeddings, values typically range 0-1 due to

Why It Matters in Graph RAG

1. Semantic Matching

   • Text chunks/documents are encoded as dense vectors (embeddings) using models like sentence-transformers
   • Cosine similarity identifies semantically related content even without exact keyword matches

2. Vector Index Efficiency

• Neo4J uses this metric for:

CREATE VECTOR INDEX chunk_embeddings FOR (n:Chunk) ON n.embedding OPTIONS {indexConfig: {`vector.dimensions`: 384, `vector.similarity`: 'cosine'}}

• Enables fast nearest-neighbor search (O(1) to O(log N) complexity)

3. Magnitude Invariance Ignores vector length differences, focusing purely on directional alignment. Critical for comparing:

• Short queries vs long documents
• Paragraphs vs sentences

Example in Your Pipeline

When processing a query "Explain neural networks":

• Query → Embedding vector Q
• Calculate cosine similarity with all chunk embeddings C_i

Similarity = cos(θ) = (Q · C_i) / (||Q|| ||C_i||)

• Retrieve top-K chunks with highest scores

Integration Note: Your system combines this with graph-aware retrieval (relationship traversal in Neo4J) for hybrid results ranking.