AMES: Approximate Multi-modal Enterprise Search with Late Interaction Retrieval

Introducing AMES (Approximate Multimodal Enterprise Search), an integrated architecture for retrieving the latest multimodal interactions that is backend agnostic. AMES demonstrates that well-designed multimodal interactive retrieval can be implemented within a production-grade enterprise search engine without architectural redesign. Text tokens, image patches, and video frames are embedded in a shared representation space using multi-vector encoders, allowing cross-mode retrieval without direct retrieval logic. AMES uses a two-stage pipeline: a corresponding token-level ANN search for each Top-M MaxSim parameterized text, followed by improved Exact MaxSim reconfiguration. ViDoRe V3 benchmark tests show that AMES achieves competitive performance within a scalable, production-ready Solr-based system.