Efficient Global Search with GraphRAG: Vectordb vs GraphDB for Production

[TWALSH77]

I'm implementing GraphRAG for a production application and facing challenges with the database choice for efficient global search while maintaining graph structure. My current observations:

The main GraphRAG repository suggests using vector databases like Lance or Azure AI Search for deployment.
Traditional vector databases excel at low-latency, high-throughput Euclidean operations but don't inherently maintain graph structures.

Graph databases like Neo4J or Neptune DB seem more suitable for maintaining graph structures but might struggle with large-scale vector operations.

I'm unsure how to implement global search efficiently while preserving the graph structure. What's the best approach for a production-ready GraphRAG implementation that can handle a large document corpus?
Specific concerns:

How to maintain graph structure with traditional vector databases?
Are graph databases like Neo4J suitable for large-scale vector operations?
Is there a way to leverage the strengths of both vector and graph databases?
What are the best practices for scaling and deploying such a system?

Potential ideas:

• Storing node and edge information as metadata alongside vectors.
• Using composite indices or additional fields to represent connections.
• Implementing graph traversal logic in the application layer.
• Hybrid Graph / Vector Db

Any insights on the correct database/infrastructure choice for inference in a production application would be greatly appreciated.