Impact index slower than BM25 scoring

[LuisPB7]

Hi everyone, thank you for this amazing library.

I am doing experiments on MSMARCOv1. I indexed the collection and searched with BM25, and got the 0.184 MRR@10 as expected. On my computer, search took around 2 minutes.

Then I switched to an impact index. I only added the "-impact" flag on my indexing and search script. Everything else is the same. Search now takes slightly more than 3 minutes.

In other impact experiments, when I increase the TF of certain terms in a document according to the weights output by some neural model (e.g., SPLADE), the difference between impact scoring and BM25 scoring becomes very large. Notice I'm not doing query/document expansion with new terms - just reweighing the existing terms, i.e., changing their query/document term frequency.

What is the intuition behind search taking longer with an impact index? Am I missing something that's being done under the hood? The posting lists do not get larger, only the stored TFs are different. Thanks!

[lintool]

I think this will answer your question: https://dl.acm.org/doi/10.1145/3576922

Please read first and then ask if you have follow-up? @JMMackenzie @andrewtrotman