Generative Prompting for Complex Product Retrieval

Complex product retrieval, which involves interpreting complicated language contexts that may not align directly with product metadata, presents significant challenges for accurate query-product matching compared to traditional keyword-based search methods. Earlier approaches that rely on pretrained language models for generating text embeddings in a shared space have shown promise but fall short in leveraging multi-modal context, particularly visual cues. Motivated by Generation-Augmented Retrieval (GAR), we propose a novel method for complex product search that enhances language-heavy queries with generative visual prompts. Our approach incorporates a pretrained LLM to map these enriched queries and product metadata into a unified embedding space, thereby improving relevance matching without necessitating product images. Experiments on the Amazon-C4 dataset demonstrate the superior performance of our method, consistently surpassing state-of-the-art techniques across various product categories. Our contributions include a generation-augmented retrieval model, an innovative use of visual prompts for query enhancement, and a comprehensive evaluation showing the effectiveness of our approach in complex retrieval tasks.