BROWSE
Detail
Conference Paper
ZeRA: Zero-Reindex Multimodal RAG via Heterogeneous Embedding Alignment for Lightweight Query Encoding
Cited 0 time in 
Share 
Share
- Authors
- Sung Jin Kim, Dasom Ahn, Hye Rim Kim, Sangwon Kim, Kwang-Ju Kim, Byoung Chul Ko
- Issue Date
- 2026-08
- Citation
- International Conference on Pattern Recognition (ICPR) 2026, pp.666-678
- Publisher
- Springer
- Language
- English
- Type
- Conference Paper
- Abstract
- Multimodal retrieval-augmented generation (RAG) systems often rely on large-scale vision encoders, making query-time inference expensive and hindering practical deployment. Re-indexing existing knowledge bases with lighter encoders is typically infeasible. We propose zero-reindex alignment (ZeRA), a lightweight-inference RAG framework that enables direct retrieval against pre-indexed high-capacity vision embeddings without reindexing. ZeRA aligns lightweight query embeddings to a frozen teacher embedding space using a small multilayer perceptron mapper, preserving retrieval compatibility while avoiding large-scale encoders at inference time. Experiments on Encyclopedic-VQA and InfoSeek demonstrate that ZeRA maintains competitive retrieval and end-to-end performance with reduced inference cost, achieving end-to-end accuracy within 5.3% of the teacher model on Encyclopedic-VQA (All).
- KSP Keywords
- Embedding space, End to End(E2E), End-to-End Performance, Heterogeneous embedding, Knowledge Bases, Multimodal Retrieval, Query embeddings, Teacher Model, high-capacity, large-scale, multilayer perceptron