<p>Interactive video retrieval (IVR) remains a challenging problem due to its multimodal nature, ambiguous user intent, and strict real-time constraints. While advances in vision-language models (VLMs) have substantially improved retrieval accuracy, existing systems still struggle to balance scalability, latency, and robustness, especially under loosely defined queries. This paper introduces a unified, multimodal, and context-aware framework for large-scale IVR that systematically addresses these challenges. First, a fast and scalable design leverages GPU-accelerated indexing, parallel inference, and lightweight fusion to achieve sub-second responsiveness at million-scale. Second, a multimodal architecture integrates vision-language embeddings, scene text, speech transcripts, and object-level cues through adaptive score fusion, mitigating modality bias and ensuring robust retrieval across diverse query types. Third, a multi-stage sequential retrieval paradigm handles ambiguous or partially ordered queries via coarse-to-fine expansion and context-aware alignment, enabling accurate retrieval under uncertain temporal boundaries. The framework is validated on V3C, domain-specific datasets, and further evaluated under time-critical, user-interactive conditions in the Video Browser Showdown (VBS) and Ho Chi Minh AI Challenge (AIC), benchmarks widely recognized for reflecting real-world scalability and interaction performance. Consistent top-tier results across multiple years demonstrate the robustness and generalizability of the proposed design beyond competition settings. By unifying scalability, multimodality, and temporal reasoning within a single architecture, this work contributes an empirically validated framework that advances the state of the art in interactive video retrieval.</p>