do you have option in code not to use node2vec , but use just similarity between nodes texts and connected to nodes

[Sandy4321]

do you have option in code not to use node2vec , but use just similarity between nodes texts and connected to nodes
node2vec is bad in performance and very slow.
seems to be Microsoft graphrag does not use node2vec?

[Sandy4321]

any answer pls

[rasinmuhammed]

The text-similarity-plus-neighbour-overlap intuition lines up with what I
landed on after trying a few approaches. Shipped a library that combines
those signals along with embeddings and configurable weights:
pypi.org/project/nodecanon (disclosure: I'm the author).

from nodecanon.adapters.nanographrag import NanoGraphRAGAdapter
from nodecanon import Resolver

graph = NanoGraphRAGAdapter.from_working_dir('./nano_output/')
result = Resolver().resolve(graph)
NanoGraphRAGAdapter.save(result.graph, './nano_output/')

Or directly from a live GraphRAG instance:

graph = NanoGraphRAGAdapter.from_instance(rag)

The 1-hop neighbour Jaccard is computed on neighbour names rather than ids,
so two duplicate nodes with different ids but the same neighbour set register
as fully overlapping. That handles the case node2vec misses where the
clustering geometry is fine but the surface forms diverge.

Fast mode skips embeddings and runs in milliseconds on graphs of a few
thousand nodes. Full mode uses all-MiniLM-L6-v2 on CPU for the semantic
component.