What changed
Hugging Face has announced two new retrieval pipelines for the ViDoRe V3 pipeline leaderboard. Both pipelines work exclusively with page images and run on open-weight models using a single GPU. The first pipeline, VisAgentRet, uses GLM-5.3-Flash as a search tool with EVIE-4.5B, achieving an average nDCG of 70.01, the highest score on the leaderboard. The second pipeline, EVIE-4.5B + JEV-27B-VL, reranks the top 16 page images with one forward pass, reaching an average nDCG of 68.25.
What this means in practice
The pipelines are designed to improve the accuracy of information retrieval from page images. VisAgentRet functions as an agent that searches and reads pages, using a combination of query rewriting and reciprocal rank fusion to refine results. The EVIE-4.5B + JEV-27B-VL pipeline reranks the top 16 page images in a single forward pass, achieving a faster query response time of 1.47 seconds per query. Both pipelines are available on GitHub under the Apache-2.0 license, with code repositories named visagentret and Jev-reranker.
Limitations and availability
The pipelines are available for use on GitHub under the Apache-2.0 license. They require a single GPU and are designed to work with page images, though an OCR text version is also available. The performance of the pipelines is reported under specific evaluation conditions, and the results are based on the ViDoRe V3 leaderboard's 8 public English datasets. The supplied announcement does not specify the exact hardware requirements beyond a single GPU, nor does it mention the availability of the models outside of the leaderboard environment.
