VCG: A Multimodal Retrieval Framework for E-Commerce Video Feeds under Extreme Cold-Start Conditions

Researchers present VCG, a multimodal retrieval system that addresses the cold-start problem in e-commerce video feeds by using vision-language models to match users and videos in a shared semantic space rather than relying on behavioral history. The system achieved a 50% uplift in video completion rates during A/B testing and demonstrates that CLIP-based discriminative embeddings outperform generative LLM approaches for retrieval tasks.

The shift from static product catalogs to dynamic video feeds represents a fundamental change in how e-commerce platforms engage users, but it creates a critical technical challenge: new videos have no interaction history for traditional recommendation algorithms. VCG tackles this by leveraging multimodal AI to understand video content semantically rather than behaviorally, enabling immediate recommendations for unwatched content. This approach addresses a real market need as platforms like TikTok Shop and Instagram Reels increasingly drive commerce.

The research reveals important insights about embedding techniques in recommendation systems. While large language models excel at semantic attribute understanding, they create embedding space collapse—where dissimilar items cluster together—making them unsuitable for retrieval tasks. CLIP-based discriminative embeddings avoid this problem by maintaining distinct semantic representations, directly translating to better recommendation diversity and user engagement.

For the e-commerce and AI industries, this work validates that vision-language models can solve previously intractable cold-start problems in high-stakes commercial applications. The 50% improvement in deep video completion suggests real business value, indicating that content-based understanding scales beyond traditional collaborative filtering in video-heavy environments. This finding accelerates adoption of multimodal AI in production systems where behavioral signals are sparse or biased.

The demonstrated bidirectional retrieval capabilities—product-to-video, video-to-product, and zero-shot search—hint at broader applications across inventory management and content discovery. Future attention should focus on how these techniques scale across geographic markets and product categories, and whether they maintain performance when video content distribution shifts.

• →Vision-language models enable zero-shot video recommendations for e-commerce feeds, eliminating reliance on behavioral history for new content.
• →CLIP-based discriminative embeddings outperform generative LLM embeddings for retrieval tasks, avoiding embedding space collapse.
• →VCG achieved 50% improvement in deep video completion metrics during production A/B testing, demonstrating measurable business impact.
• →Position and duration biases in video feeds significantly distort engagement signals, making multimodal content understanding essential for fair recommendations.
• →Bidirectional retrieval between products and videos creates new pathways for inventory discovery and content monetization in dynamic e-commerce environments.