Harnessing AI for Detection and Correction of Hallucinations in Large Search Systems

This paper outlines our approach to tackling a core challenge in designing a robust e-commerce search system for over 10 million stock-keeping units (SKUs). The system employs AI models (OpenAI, Llama 3) for three major functions: categorization, tagging, and semantic searching using vector embeddings. Categorization refers to classifying incoming information from different third parties into a taxonomy essential for data accuracy and standardization. Tagging aids search by linking relevant tags for easier SKU filtering. Lastly, vector search, using Elastic Search and cosine similarity, enables efficient searches and retrieval of relevant information. To enhance tagging and categorization accuracy, mathematical models using embeddings and cosine similarity were applied to reduce hallucination effects.