Neural Search Talks — Zeta Alpha cover image

ColBERT + ColBERTv2: late interaction at a reasonable inference cost

Neural Search Talks — Zeta Alpha

00:00

Colbert's Architecture for Re-Rating Documents

Colbert: I think we can describe the architecture of Colbert a bit more in detail. So they have it with mask tokens, I believe, as a query expansion idea. They learn some representation for these mask tokens and then they include them in that maximum similarity computation. It's fairly straightforward. Their insight here is probably to keep the actual embeddings rather than trying to have a single vector. Which improves effectiveness, but you also have to store them.

Transcript
Play full episode

Remember Everything You Learn from Podcasts

Save insights instantly, chat with episodes, and build lasting knowledge - all powered by AI.
App store bannerPlay store banner