Metadata Integrity: Sequential processing ensures every page is tagged with its source file. Relationship-Aware: Captures connections across different pages and documents. High Performance: Leverages Groq's LPUs for near-instant extraction and querying. Scalable: Built on Neo4j Aura for cloud-native graph storage.
LLM: Llama 3.3 (via Groq) Parser: LlamaParse Database: Neo4j Aura (Graph) Environment: Python / Jupyter Notebook
graph TD
subgraph Ingestion
A[PDF Documents] --> B[LlamaParse]
B --> C[Markdown Text]
C --> D[Groq Entity Extraction]
D --> E[(Neo4j Knowledge Graph)]
end
subgraph Retrieval
F[User Query] --> G[Groq Cypher Generator]
G --> H[Neo4j Traversal]
H --> I[Graph Context]
I --> J[Groq Final Synthesis]
J --> K[Expert Answer]
end
This repository includes a research_traditional_rag under Test_RAG folder (based on foundational tutorials by Krish Naik). I maintained this as a baseline to compare performance against the current GraphRAG implementation.
Key Findings:
- Traditional RAG: Excellent for simple fact retrieval but struggled with "global" queries across multiple document sections.
- GraphRAG (Current): Significantly improved entity relationship mapping and cross-document synthesis. Refer Folder -
/notebook/Graph_RAG/graphRAG.ipynb
Graph_RAG project uses the Apple 2024-2025 10-Q filing as the primary test case.
The document is located in the /data/financial_pdf folder. It was selected for its complex
financial tables and nested entity relationships, which perfectly demonstrate
the power of GraphRAG over traditional RAG.