chunk-level retrieval和graph-level retrieval的区别

[liulangjita]

感谢蚂蚁的优秀项目~
chunk-level retrieval和graph-level retrieval的区别是什么，graph-level retrieval应该是通过图谱查询的数据结果，chunk-level retrieval是通过向量数据库召回的数据。他们是如何融合到最终结果里的，从向量数据库召唤的结果会影响最终回答的可信度吗

[leywar]

Let me try to answer your question. The goal of KAG's Graph-Level retrieval is to achieve more accurate reasoning. First, KAG-Solver will perform precise reasoning on the predefined schema knowledge part. Then, in the graph structure obtained by open information extraction, the relevant subgraph structure is retrieved through semantic reasoning and solved using a large language model. Chunk-Level retrieval is used as a supplement to alleviate the problem of sparse structured knowledge. Supporting documents are retrieved through Graph-Chunks inverted index, DPR and other means, and then the large language model is called to solve sub-problems one by one. The overall execution process is to perform Graph-Level reasoning and retrieval first, and then perform Chunk-Level retrieval.

[leywar]

The goal of vector retrieval is to find the most relevant chunks. The effect on multi-hop fact reasoning is acceptable. However, reasoning with logical calculations relies more on the structured knowledge of the graph.

[mars216]

提供搜索的目的是找到最相关的内容，对多跳事实推理的效果尚可，但逻辑计算推理更依赖于图谱的构成知识。

对于图搜索是使用的规则匹配，还是向量搜索呀，为啥在给的V0.5的用户规则手册里面，上面文字写规则匹配，例子给出向量搜索呀？同理文本块在论文里边说有一个实体了，那是如何进行向量检索的？