rag.insert() is so slow

[SLKAlgs]

The rag.insert() process is too slow. Can I divide the document into multiple parts, specify different GPUs for insert in the same project, and finally build a knowledge graph about the complete document?

[LarFii]

This approach could lead to conflicts. However, one possible solution is to divide the document into multiple parts, cache all the extraction results from the LLMs for each part, and then merge the caches before performing the insertion. I haven’t tried this yet, but it should be theoretically feasible.

[nihirv]

To mitigate #315, I decided to modify my code to call rag.insert() on each page of a document (i.e. [rag.insert(page) for page in doc] vs rag.insert(doc)), and am also finding this method very slow