FuseRank (Demo): Filtered Vector Search in Multimodal Structured Data

We describe and demonstrate our work on multimodal filtered vector search in tabular data. It offers a practical way for businesses to vectorize their product assortments or any other business-critical data and then simultaneously retrieve and filter this information using state-of-the-art similarity search. Our methodology is based on the extended vector space model, with multiple modalities represented as sub-vectors that get concatenated and compared to the query vector via a dot product operation. It is a flexible framework that allows manipulating the influence of each modality on the overall item ranking via modality weights. We share the source code, the demonstration video, and the screenshot of the application. We also provide a brief description of its main building blocks, the supported data types, and modality filters. The application is bundled with two public datasets and pre-computed text embeddings so that it can be easily run without prior preparation.