Question about graph-building time and configuration for large-scale LightRAG indexing

[Jasper2099]

Hi HKUDS team,

Thank you for open-sourcing LightRAG.

I would like to ask about your own practical graph-building performance with LightRAG.

Could you please share:

1. Your rough end-to-end average graph-building throughput in tokens/s

2. For your largest case around 5M tokens, the total time required for full graph construction

3. The configuration used for that case, especially:

   • LLM model
   • embedding model
   • local serving or API
   • GPU/CPU setup
   • GPU type and count
   • tensor parallel settings
   • graph storage backend
   • vector storage backend
   • chunking strategy
   • concurrency settings
   • timeout settings

If possible, it would also be helpful to know whether your reported tokens/s refers to model generation only or the end-to-end graph-building pipeline.

Thank you very much.