NEW Knowledge Graph based RAG: SimGRAG (no training)
Вставка
- Опубліковано 8 лют 2025
- Excellent new Knowledge Graph based RAG system called SimGraphRAG, or simply SimGRAG.
Overview of our four classical KG-based RAG systems, and the new SimGRAG, which outperform them. Short technical deep dive into the new methods and algorithms, plus code via GitHub repo.
All rights w/ authors:
SimGRAG: Leveraging Similar Subgraphs for Knowledge Graphs
Driven Retrieval-Augmented Generation
Yuzheng Cai, Zhenyue Guo, Yiwen Pei, Wanrui Bian, Weiguo Zheng
from Fudan University
#airesearch
#knowledgegraph
#science
#aiagents
#graph
With the automatic audio dubbing from UA-cam /Google you hear a synthetic voice in your regional language.
To hear my original voice in English, switch to "Default" or "English" in the settings. Thank you.
I liked your video but your voice sound is low compared to other YT videos.
SimGRAG reshapes the landscape of KG-based RAG. The GitHub repo holds the keys to its success. Well done, Fudan team!
This is awesome! I’m actually working on implementing this solution in the aerospace domain, so this is incredibly useful! Thanks for sharing, I’ll probably go dig into the white-paper so that I can see what they’re doing differently and what I can learn 😊
Hi man, I am doing the same for another engineering domain, would be nice to discuss and explore!
I used lightRAG with neo4J as gRAG repo and very impressed with its gRAG capabilities.
Game changing development in KG-based RAG systems! SimGRAG shows significant improvement. Looking forward to exploring the GitHub repo. #AIResearch #Science
SimGRAG's optimization of KG-based RAG using similar subgraphs is revolutionary. Huge thanks to the Fudan University team. #AIagents
What about Light RAG?
SimGRAG is a game-changer in KG-based RAG systems. Can't wait to explore its advanced methods and code. #airesearch #knowledgegraph
Hi my friend! Can you explain how it’s different from neo4j graph rag and lightrag?
I love your voice :D
Agree.
Would you say this is similar to HyDE but its more like HyGE where the LLM generates a hypothetical graph to use for retrieval instead of a hypothetical document?
Thank you very much for this. Do you have any particular video or reference to recommend to build “small but smart vector spaces” for a particular domain?
Depends on your technical terms in your domain and your training of the tokenizer, plus optimal size of vocabulary (I work currently with 75K for physics).
How does this compare to more well-established KG tools, such as Infranodus?
14:00 the mapped node for a placeholder in the pattern graph (triples) will Not be evaluated concerning distance ... .really? . so we are sort of putting emphasis on the source node and explicitly known relationships ? could not at least the node type be accounted for? I mean we know in the examples that it is a unknown director or movie ? Or is that so encoded in the known parts via embedding, as to Not matter? I find it quite astonishing..
the retrieval mechanism is too naive, it can’t be really useful in practice.
For example, the system would fail to answer the query “which country is Sacramento located in” because the KG has the structure: Sacramento California United States
Here the query has 2 nodes while the KG subgraph has 3. They wouldn’t match!
You comment includes "would " and "wouldn't". Therefore I assume, you have not implemented a real test, but you are just guessing?