A couple Qs for the model
First off, congrats on the release! Saw the paper on Twitter and was impressed.
I'm currently evaluating the use of ColBERT in production. I've been using BERT retrieval + FlashRank reranking and haven't been super impressed with the results. My documents are primarily legal, so perhaps BERT doesn't generalize well to it.
My Qs are:
Can this be used as both an indexer, retriever, and reranker all-in-one? Or, does the retrieval step also do the reranking? I'm leaning 99% you need retrieval + reranking separately, but I'm not familiar enough with the ColBERT architecture to be certain.
What is the context length of this model? It appears to be 512 based on my read of the config files, but I'd love an external confirmation.
Very excited to see the MTEB benchmarks! I haven't really bothered with fine-tuning BERT models yet, but the RAGatouille repo makes it look so simple I'm inclined to give it a shot. I think shoving the Documents into Gemma's 1M context window and asking questions is a pretty good way to generate synthetic fine-tuning data.