Controlling the semantic search similarity metric

[Ronnie (ronniemi)]

I noticed in the code that the default semantic search similarity metric seems to be L2. Where can I configure it to other metric? such as cosine

[Juyel (juyelhushen)]

If you're referring to the vector store used by GraphRAG's semantic search, the similarity metric is determined by the underlying vector-store implementation rather than by the LLM configuration.

For example, if you're using a vector store that supports both L2 and cosine distance, cosine generally needs to be selected when the vector index is created/configured. It isn't simply a ranking option applied after retrieval.

I'd check which vector-store implementation your GraphRAG version is using first. If you're using the default store in that version, the relevant configuration/code path is the vector-store/index configuration rather than settings.yaml's LLM settings.

Also, be careful when comparing L2 and cosine: if the embeddings are L2-normalized, squared L2 distance and cosine similarity induce the same ranking (they are monotonically related), although the returned score values are different.

If you share the GraphRAG version and the vector-store configuration you're using, I can point to the exact setting/code path for changing it to cosine.