I get same score between euclidean distance and cosine similarity for all questions

[stepkurniawan]

Hi all,

Im using LangChain x FAISS. And everytime I compare the distance score between the default euclidean and cosine, it will always give me the same score, so I dont know maybe I did something wrong (or not). Could you check please?

This is how I change the distance metrics:

faiss_distance_strategy = get_faiss_distance_strategy(self.index_distance) # can be default eucledian, or DistanceStrategy.MAX_INNER_PRODUCT, or DistanceStrategy.COSINE

# langchain from_documents
self.db = self.vectorstore_obj.from_documents(split_docs, self.embedding, distance_strategy=faiss_distance_strategy) 

self.db.save_local(self.save_path)

and this is how I check the distance score:

similarity_search_with_score(query, k=top_k)

and here is the problem, everytime I get a result, for example, I get a document A, with distance 0.39182765 using euclidean, I also got the same score with cosine, even down to the 8th digits behind comma. With MAX_INNER_PRODUCT i get something different, probably because its searching for maximum, and not minimum.

is it because everything got normalized?

Thank you!

[algoriddle]

This would be a better question for langchain: https://github.com/langchain-ai/langchain/issues

[xrd]

In the README, it does indicate normalization plays a part for cosine similarity.

https://github.com/facebookresearch/faiss?tab=readme-ov-file#introduction