Developers Debunk Late Interaction Misconceptions: Naming and Index Size
Developers in the retrieval community have opened a discussion to clear up two widespread misconceptions about the "Late Interaction" paradigm: misused terminology and overblown concerns about index size.
Confirmed
- Both @lateinteraction and @tomaarsen noted that the community broadly mislabeling the paradigm as "Multi-vector" retrieval is a major problem — the correct name is "Late Interaction."
- Both emphasized that the paradigm's core lies not in how many vectors are used, but in how the scoring function is designed; @tomaarsen added that the name MultiVectorEncoder (MVE) is even more misleading than terms like Late Interaction.
- @lateinteraction pointed out that index size has not been an issue since 2021, except for visual search cases like ColPali; ColBERTv2 and PLAID solved storage costs well before text-davinci-002.
- He further explained that for many applications, a standard PLAID index is actually smaller than today's commonly used oversized single-vector representations — the fear that "index size grows with scale" is unfounded.
Why it matters
- Misused terminology can lead to misunderstanding of the technology's essence, shifting developers' focus away from the scoring function, which is the true core.
- The real bottleneck today is infrastructure and adoption: many still insist on raw matrices and MaxSim computation while ignoring more efficient implementations like KV cache, showing the problem lies in diffusion rather than the technical challenge itself.
2026-08-25 ~ 2026-08-25 · 6 related posts
Primary sources
- Misnaming 'Late Interaction' as 'Multi-vector' obscures the real issue — lateinteraction ·
- Doing the math: int8 dense embeddings cost ~2x the storage of binary ColBERT — antoine_chaffin ·
- Tech discussion: Why it should be 'Late Interaction' not 'Multi-vector' — tomaarsen ·
- [source] Misnaming 'Late Interaction' as 'Multi-vector' obscures the real issue — lateinteraction · 2026-08-25
- [source] Tech discussion: Why it should be 'Late Interaction' not 'Multi-vector' — tomaarsen · 2026-08-25
- Dev on retrieval: Index size hasn't been an issue since 2021 — lateinteraction · 2026-08-25
- Vector Retrieval Myth: Index Size is Not the Bottleneck — lateinteraction · 2026-08-25
- Index size not the bottleneck; underutilization of infra is — lateinteraction · 2026-08-25
- [source] Doing the math: int8 dense embeddings cost ~2x the storage of binary ColBERT — antoine_chaffin · 2026-08-25