Omar Khattab

Excited to see this making its way to Vespa! (First author of the ColBERT paper.) Happy to help with ideas/etc. about doing this. FYI our implementation is here: https://github.com/stanford-futuredata/ColBERT/tree/v0.2

Very interesting---thanks @jobergum! I see that you're using d=32. Another way to go about this is to use a larger dimension (say 128) but quantize the representations. This may perform...

That's really awesome! The results with medium BERT and a small dimension are better than I expected. Also the latency is pretty great. Is this on CPU? I'm wondering how...

Thank you @jobergum and thanks to the team! This is actually much faster than I thought. I'm wondering how fast you guys could make the end-to-end retrieval version of ColBERT...

Your understanding is correct. The only remark is that q, p, and n in your notation are matrices, not vectors.

We're planning to make improvements to how the package works with pip. In the meanwhile, please follow the import approach shown at the top of the intro notebook: https://github.com/stanford-futuredata/ColBERT/blob/new_api/docs/intro.ipynb (Also,...

@luffycodes Please follow the instructions on the README for installation. Don't use the setup script directly.

@luffycodes Which branch do you want to install? From your other issues, I assume you're interested in ColBERTv2. In that case, please check the README for the new_api branch. From...

Thanks, @luffycodes ! Let me know if the commands for installation on the new_api branch and the corresponding ColBERTv2 notebook don't work for you.

Thanks James! Is there a typo in `dest='n'` in the parsing file?