Most conversations about visual document retrieval fixate on NDCG. The number that actually wrecks a production budget is index size — and that’s where DistilVDR is more interesting than its accuracy headline.
The system distills an 8B vision-language teacher down to a 524M single-vector retriever, end-to-end, both the query and document sides. That “both sides” matters: prior compression work either trained a small multi-vector encoder from scratch or distilled only the query side, and neither gives you a compact single-vector retriever you can actually index cheaply. DistilVDR-HiRes keeps 86.9% of the teacher’s NDCG@5 across ViDoRe v1+v2+v3 while beating every reproduced sub-1B baseline on the resolution-sensitive v3 set.
The design choice I’d steal is the asymmetry. Document encoding is offline and can afford visual capacity; the query encoder runs per request, so they keep it at 70M parameters. Supervision comes entirely from the frozen teacher’s embedding space — no relevance labels, no negative sampling, no contrastive term — which quietly deletes a whole data-pipeline that most retrieval training drags along. The operational payoff: one million documents fit in an index 15.6x smaller than the strongest sub-1B multi-vector baseline, indexed an order of magnitude faster.
This is the single-vector-versus-late-interaction tradeoff made concrete. Multi-vector recall is lovely right up until you price the index footprint and the tail latency at corpus scale. A single-vector student that holds ~87% of an 8B teacher’s quality is exactly what I’d A/B behind a hybrid retriever before assuming I need the big model in the serving path. The HF paper page has the per-benchmark ViDoRe breakdowns worth reading before you trust that average on your own corpus.
The uncomfortable question: if a 524M student recovers 87% of the recall at a fraction of the index cost, how much of your retrieval budget is currently buying the last 13%?