Hybrid RAG (GraphRAG and Azure GPT-RAG)

[xtoshi]

I'm working on implementing a hybrid RAG (Retrieval-Augmented Generation) system using GraphRAG and Azure GPT-RAG libraries. My goal is to process unstructured documents like employee handbooks and create a system that leverages both graph-based and vector-based retrieval for enhanced context and improved responses. I'd like to validate my approach and seek advice on best practices.
Here's my proposed workflow:

Process documents and generate vector embeddings for text chunks using Azure GPT-RAG.
Store embeddings in Azure Cognitive Search.
Use GraphRAG to create a knowledge graph from the documents.
Implement a hybrid retrieval system that combines:
a) Vector similarity search from Azure Cognitive Search
b) Graph traversal using GraphRAG
Merge results from both methods to create an enhanced context.
Use this rich context to generate responses with a language model.

My main questions are:

Is this approach aligned with GraphRAG's capabilities and best practices?
How can I most effectively combine the results from vector search and graph traversal to create the best possible context?
Are there any specific considerations or potential pitfalls when integrating vector embeddings with GraphRAG?
Can you suggest any optimizations or improvements to this workflow?
Are there any examples or case studies of similar implementations using GraphRAG that I could refer to?

I'd greatly appreciate any insights, suggestions, or resources that could help me implement this hybrid approach effectively. Thank you in advance for your expertise!

[xtoshi]

Any recommendations or suggestions here?

[Victor-Chow]

Maybe you need R2R

[SpikeYangRc]

Any recommendations or suggestions here?

Hi @xtoshi Do you have any update for Hybrid RAG research?

[joshua-atolagbe]

@xtoshi . Here is a paper implementation https://arxiv.org/pdf/2408.04948