Could you share the research that was done with vector+graph performance comparison?

Hi Zack, Welcome to the community! The split-n-embedding approach by nature: - chunking assumes the information density/how data is spread - embedding model may get literally similar chunks but not related on domain knowledge sense The KG on the other hand is: - fine-grained segmentation of info - interconnection-relevance oriented While the pure KG approach: - Normally is string exact match based on start-node searching - Not good at persisting detailed information like text chunks in Vector Store Thus combining the two is optimal in some use cases. We for now cherry-picked some cases that KG helps https://colab.research.google.com/drive/1tLjOg2ZQuIClfuWrAC2LdiZHCov8oUbs#scrollTo=_Cherry_picked_Examples_that_KG_helps We are creating a proprietary SDK/tool suite(early stage now) to enable advanced RAGs with the help of Graph, Vector, planning-based graph exploration approaches, etc(paper on the way) 😄. Feel free to discuss more about it.

Wonderful thanks so much @wey! I've shared this with my team, as we're currently doing research on KG adoption.

I indeed recently put NebulaLLM(our in-house graph rag/Chain of Exploration/text2cypher SDK) into an agent-based PoC. Where KG-CoE is just one tool to be called about specific kinds of knowledge retrieval and it worked fine. The agent is just a function calling based implementation and CoE was wrapped as a function/tool

你有例子吗?