为啥会比LightRag节约token？

[hongdongjian]

1. 构图都是一个个chunk提取，然后该项目比LightRag多了社区的概念，token消耗不是更多么？
2. 多了社区的概念，每次新增文档都需要重新构图，成本不是更高么？

[JunnanHansonDong]

Thanks for your questions and interest.

We respect LightRAG as a meaningful pioneering work. Responses hereunder correspondingly:

• LightRAG:

  • The extraction is unlimited with no targets.
  • The schema template (different from our schema) additionally requires detailed descriptions for all entities and relations in the output.

• Youtu-GraphRAG:

  • We employ a schema to constrain the extraction to targeted entity/relation/attribute types only.
  • We generate no descriptions for all the extracted content.
  • The community is identified by a local algorithm in a limited number, and only a very short description will be generated for each community using LLMs.
  • Incremental updates need no reconstruction and community detection. The new nodes and edges can be instantly identified among existing communities, or be included in a new one if current ones are not applicable. However, we have a more advanced design in phase 2, it will be released asap once done.

You are welcome to have more discussions and have a try.