Filtered multimodal re-ranking for e-commerce

This study explores the integration of Large Language Model (LLM) and Information Retrieval (IR) components to enable filtered search over multimodal structured data. We identify core integration challenges and introduce a conceptual framework based on two paradigms: filtered retrieval and filtered re-ranking. With the focus on the latter, we employ RT-3 rank transformation to dynamically adjust the ranking scores according to user-defined filters across multiple attribute modalities numerical , categorical, binary, and spatial. This approach avoids redundant computations, provides fine-grained control over modality pri-oritization, and mitigates over-constrained filtering via soft ranking. We implement our method in a publicly available web prototype using two real-world datasets, demonstrating its effectiveness in balancing semantic and contextual relevance while empirically validating improved system efficiency in multimodal search scenarios.