AnswerDotAI / RAGatouille

Easily use and train state of the art late-interaction retrieval methods (ColBERT) in any RAG pipeline. Designed for modularity and ease-of-use, backed by research.
Apache License 2.0
3.03k stars 206 forks source link

`colbert-ir/colbertv2.0` is non-deterministic? #196

Closed nicokuzak closed 6 months ago

nicokuzak commented 6 months ago

Hello! I have a notebook where I am creating an index and then searching it a bunch of different times for different queries to evaluate colbertv2 on my data.

I duplicated the notebook, seeded it, yet I get different scores and different results in the two notebooks. Is this a known limitation? Is there anything I need to do for determinism in the models?

bclavie commented 6 months ago

Interesting! Among other potential reasons, this might be because of the new (experimental) implementation of k-means without faiss. If you index with the use_faiss=True argument passed, do you get deterministic scores?

nicokuzak commented 6 months ago

@bclavie the two notebooks are different even when using use_faiss=True

nicokuzak commented 6 months ago

Ok I figured it out. There was a set( in my preprocessing of the text data 🤦 . I'll check out if use_faiss actually makes a difference here but it was on me - sorry!