关于图拓展的问题

[wzrf]

想问一下这个图是不是只能一次生成，没法生成之后持续构建？如果要新增文档的话应该怎么做，是不是communities/keywords这些要重新计算？

[siyuan-youtu]

Hello. Currently, our graph construction phase mainly consists of two parts: LLM-based triple extraction and community detection.

The first part can be optimized incrementally by saving previously extracted triples for corresponding schemas and documents, which can significantly save time and reduce token usage.

However, for the second part, community detection, we do not recommend an incremental approach, as this stage involves reclustering the entire graph.