代码中是不是没用到社区图的检索结果

[StrongBirds]

你好，我不知道是不是我看错了，有个疑问，在enhanced_kt_retriever.py的第1141行，为什么不用社区图的检索结果
return {
"scored_triples": scored_triples
}

[siyuan-youtu]

What specific operation does the "community graph" retrieval you mentioned refer to?

Does line 188 of the file youtu-graphrag/models/retriever/faiss_filter.py match what you described?

[StrongBirds]

What specific operation does the "community graph" retrieval you mentioned refer to?您提到的“社区图”检索具体指的是哪种操作？

Does line 188 of the file youtu-graphrag/models/retriever/faiss_filter.py match what you described?文件 youtu-graphrag/models/retriever/faiss_filter.py 中的第 188 行与您描述的是否一致？

是的，你这里有社区检索的结果，但是在enhanced_kt_retriever.py的第1159行,只取了返回结果中的scored_triples的值？并没有使用返回结果中的comm_nodes的值。所以很疑惑，不知道是不是自己理解的不对。

[JunnanHansonDong]

@StrongBirds Thank you for your careful attention to this detail.
You're absolutely right that the current version provides triple_only_retrieval() as the default mode, that's also why in this version we name this function triple_only, which was indeed an engineering optimization to simplify integration with the frontend for a lightweight demo as an open-sourced project for the public.

However, we’ve retained the ​​full community-based top-down filtering path​​ in the codebase (retrieve_via_communities()) as you see, ensuring compatibility with the full methodology described in the paper.

• ​​How to Leverage Communities​​

  • Even in the current triple_only mode, the returned nodes/triples have already included communities and members (e.g., if a community node appears as a head/tail entity) and also after neighbor expansion. This could somehow obtain the high-level summarization indeed.

  • Or you can easily release the comm nodes out from the retrieved results in dual path retrieval as described in the paper.

Im happy to help if you would like to dive deeper to include full comm nodes in the results.

[StrongBirds]

@StrongBirds Thank you for your careful attention to this detail.感谢您对此细节的仔细关注。 You're absolutely right that the current version provides triple_only_retrieval() as the default mode, that's also why in this version we name this function triple_only, which was indeed an engineering optimization to simplify integration with the frontend for a lightweight demo as an open-sourced project for the public.您说得对，当前版本默认提供 triple_only_retrieval() 模式，这也是为什么在这个版本中我们将其命名为 triple_only ，这确实是一个为了简化与前端集成并作为一个轻量级演示的公开源代码项目而进行的工程优化。

However, we’ve retained the ​​full community-based top-down filtering path​​ in the codebase (retrieve_via_communities()) as you see, ensuring compatibility with the full methodology described in the paper.然而，正如您所看到的，我们在代码库中保留了完整的基于社区的从上到下过滤路径（ retrieve_via_communities() ），以确保与论文中描述的完整方法兼容。

• ​​How to Leverage Communities​​如何利用社区

  • Even in the current triple_only mode, the returned nodes/triples have already included communities and members (e.g., if a community node appears as a head/tail entity) and also after neighbor expansion. This could somehow obtain the high-level summarization indeed.即使在当前的 triple_only 模式下，返回的节点/三元组已经包括了社区和成员（例如，如果社区节点作为头/尾实体出现）以及邻域扩展之后。这确实可以以某种方式获得高级总结。
  • Or you can easily release the comm nodes out from the retrieved results in dual path retrieval as described in the paper.或者，你可以像论文中描述的那样，在双路径检索中轻松地将社区节点从检索结果中释放出来。

Im happy to help if you would like to dive deeper to include full comm nodes in the results.如果您想深入研究，包括完整的社区节点在结果中，我很乐意提供帮助。

感谢作者认真细致的回复，我会先尝试自行修改代码使用社区检索结果，对比在使用与不使用社区时对最终回答效果的差异。如有疑问，还望作者不吝赐教。

[StrongBirds]

你好，还有一个疑问，在tree_comm.py中_refine_cluster函数，社区合并之后产生的新的中心节点会不会不存在于center_sim_matrix矩阵中的风险？

[StrongBirds]

作者你好，在tree_comm.py的_build_batch_prompt函数中，comm_info = {
"id": comm_id,
"center": center_name,
"members": member_names[:10],
"size": len(members)
}标黑的为什么这样取值，如果在求中心节点的时候已经知道哪些节点比较重要，为什么不取比较重要的节点？

[StrongBirds]

作者你好，在enhanced_kt_retriever.py中的_build_node_text_index函数，我看是对字符串做了split()，如果在中文场景下这种做法是否需要改进，如果改进，推荐使用什么方法？

[JunnanHansonDong]

@StrongBirds Hi, responses hereunder one by one:

• _refine_cluster and center_sim_matrix. You can see that the max_iter has been pre-defined as 1, so there will be no updates to the matrix. This design is generally made to boost efficiency.
• members in comm_info. The size of each community is not that big, so top 10 is enough to represent this comm. Anyway, you are welcome to raise a PR or modification locally to include key entities in this column for better representation.
• split(). In Chinese scenarios, jieba segmentation could be a potential alternative.

Thanks for your great interest and careful attention to our project again.

[StrongBirds]

• you are welcome to raise a PR

@StrongBirds Hi, responses hereunder one by one:您好，以下逐条回复：

• _refine_cluster and center_sim_matrix. You can see that the max_iter has been pre-defined as 1, so there will be no updates to the matrix. This design is generally made to boost efficiency._refine_cluster 和 center_sim_matrix。您可以看到 max_iter 已经被预定义为 1，因此矩阵将不会有更新。这种设计通常是为了提高效率。
• members in comm_info. The size of each community is not that big, so top 10 is enough to represent this comm. Anyway, you are welcome to raise a PR or modification locally to include key entities in this column for better representation.comm_info 中的成员。每个社区的大小并不大，所以前 10 个足以代表这个社区。无论如何，您都可以提交一个 PR 或本地修改，以包括此列中的关键实体，以获得更好的表示。
• split(). In Chinese scenarios, jieba segmentation could be a potential alternative.split(). 在中文场景下，jieba 分词可能是一个潜在的替代方案。

Thanks for your great interest and careful attention to our project again.感谢您再次对我们项目表示出极大的兴趣和关注。

谢谢回复，阅读完后续代码之后，我会逐步尝试修改。