Visual document retrieval · Under review

MINER

Keep one vector. Retrieve more with it.

MINER (Mining Multimodal Internal RepreseNtation for Efficient Retrieval) selects, aligns, and fuses internal representations from a frozen retriever into one dense embedding.

+4.5% largest relative average nDCG@5 gain

42.4× smaller index than late interaction

5.3× faster search than late interaction

1 dense vector stored per page

How it works

From internal layers to one embedding

Each click mutates the same scene. Query on the left, page on the right. The backbone stays frozen while MINER probes, masks, and fuses internal layers.

Query tokens and page patches enter one frozen backbone.

Results

Improved dense retrieval without a multi-vector index.

Across the ViDoRe suite, MINER improves eight of nine backbone averages while retaining one embedding per page.

Jina · ViDoRe V2 · measured

Quality against serving cost

Showing retrieval quality against index size. MINER improves Jina by 1.9 nDCG points with no added index storage.

Dense Jina MINER Late interaction

Full-corpus Qdrant measurements; higher nDCG and QPS are better, lower storage is better.

Index-scale efficiency.

Late-interaction systems improve fine-grained matching by storing many token vectors for every page, so their storage and search costs grow across the corpus. Dense retrievers store one vector but ordinarily use only the final-layer embedding. MINER instead combines selected internal representations into a single D-dimensional embedding compatible with the same dense index. Across ViDoRe V1, V2 and V3, eight of nine backbone averages improve.

Backbone averages across the ViDoRe suite

Benchmark	Metric	Backbone	Dense baseline	MINER	Relative change
ViDoRe V1	nDCG@5	Jina	84.3	84.4	+0.1%
ViDoRe V1	nDCG@5	Eager	82.5	84.8	+2.8%
ViDoRe V1	nDCG@5	MoCa	85.7	85.6	−0.1%
ViDoRe V2	nDCG@5	Jina	53.3	55.2	+3.6%
ViDoRe V2	nDCG@5	Eager	56.0	58.5	+4.5%
ViDoRe V2	nDCG@5	MoCa	58.3	59.1	+1.4%
ViDoRe V3	nDCG@10	Jina	49.0	49.7	+1.4%
ViDoRe V3	nDCG@10	Eager	47.6	48.8	+2.5%
ViDoRe V3	nDCG@10	MoCa	49.1	49.5	+0.8%

Jina efficiency on the full ViDoRe V2 benchmark

Method	Vectors per page	QPS	Index storage	Average nDCG@5
Dense Jina	1	15.41	22.24 MB	53.3
MINER-Jina	1	14.36	22.24 MB	55.2
Jina late interaction	Many	2.75	943.38 MB	57.6