Original title: MIDR: Enrichment-Augmented Indexing for Multimodal Document Retrieval
Debanjan Mahata, Atharva Tendle, Daniel Preotiuc-Pietro, Yong Zhuang +1 · September 2026
Searching visually rich documents is hard because standard text scanners often scramble or skip charts, tables, and layouts. Other tools fix this by searching raw images directly, but that requires heavy computing power during every search. Researchers built MIDR to solve this. Instead of waiting for a user to search, MIDR uses an artificial intelligence that understands images to read page pictures ahead of time. It turns visual details into structured text fields before saving them to a searchable text index. You can also mix it with standard text retrieval. In tests across five English domains, MIDR improved search accuracy over basic text search by 23.0% and stayed competitive with top visual systems, while needing about 9x less index memory and cutting query delays by about 2x.
Imagine sorting through a pile of mixed photos and written notes to find a specific recipe. Some notes are on the back of photos, and charts are hard to read without context, so you might miss important steps. In this paper, the system processes each page thoroughly when it's first added, turning all visual and text elements into searchable descriptions. This way, when you search later, everything is already organized and easy to find.
This helps people search through documents with charts, tables, and images more effectively. It reduces the need for complex systems and can handle documents in different languages, making information more accessible.
The abstract does not specify limitations, but the framework is evaluated only on the ViDoRe V3 benchmark and two French domains. It may not generalize to other languages or document types not tested.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
MIDR: Enrichment-Augmented Indexing for Multimodal Document Retrieval — Debanjan Mahata, Atharva Tendle, Daniel Preotiuc-Pietro et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.