MUMBAI, India, Nov. 2 -- Intellectual Property India has published a patent application (202641125239 A) filed by Sns College Of Technology on October 21, 2026, for Zero-Shot Multimodal Visual Product Identification System And Method For Conversational Commerce Using Contrastive Image-Text Embeddings, Vector Similarity Search And Ocr-Based Disambiguation.
Inventors include Mr. S. Haribabu; Vishal S; Swathi S; Ramani Krishnan M; and Dhilip Kumar S.
The application for the patent was published on October 30, 2026, under issue no. 44/2026.
Abstract: The present invention relates to a business-to-business software-as-a-service platform configured to enable businesses, such as automobile spare parts dealers and lubricant oil vendors, to identify products from customer-submitted photographs received through a chat interface, such as WhatsApp, and to respond with matching product information within the same conversation. The proposed system provides a unified platform combining a secured administrator portal, catalog and inventory management, multimodal vector ingestion, a customer chat intake workflow, zero-shot visual similarity search, optical character recognition (OCR) based disambiguation, and configurable confidence thresholding.The system receives clean catalog images uploaded by an authenticated administrator, organizes them by folder, stock keeping unit (SKU), and tag, converts each image into a multimodal embedding vector using a contrastive image-text model (Jina CLIP), and stores each vector alongside its product metadata in a unified vector-enabled database cluster (MongoDB Atlas). Each product is further classified as either a visual product or a text-based product. Upon receipt of a noisy, real-world photograph from a customer through the chat interface, the system encodes the photograph directly as a base64 payload and submits it to the embedding model without intermediate storage, executes an approximate nearest-neighbour vector search ($vectorSearch) against the catalog embeddings, and retrieves the highest-scoring candidate products. Where the candidates are text-based products of near-identical appearance, such as branded oil bottles, the system applies OCR to the customer photograph and uses the extracted text as a tiebreaker among the top five visual matches. The resulting best match is compared against a configurable confidence threshold, and the system returns either the identified product, comprising name, price, description, and similarity score, or a no-match response, in under two seconds. The invention eliminates dependence on large labelled training datasets and on customer-entered text searches, and provides fast, accurate, and trusted product identification within conversational commerce.
Disclaimer: Curated by HT Syndication.
