Vanguard Pinecone Semantic Search

Vanguard, a leading investment management company, offers a range of financial services including retirement services, advice, and investments. Vanguard's customer support teams previously relied on keyword-based search solutions to find documents where answers to customer questions may live. This required representatives to manually search for answers within dense, lengthy documents, ultimately increasing call times and reducing customer satisfaction. To mitigate lengthy calls during peak seasons, Vanguard would hire additional representatives, adding operational cost and overhead. The Center for Analytics and Insights team at Vanguard was tasked with finding an alternative solution for real-time retrieval over a highly dynamic dataset. They needed to move beyond a keyword search-based system to a semantic or vector search-based system. During early evaluations with JSON storage and cosine similarity-based search solutions, they encountered significant limitations: slow search and generation, inefficient management of growing datasets, and often irrelevant search results. The team ultimately chose Pinecone as their vector database to power hybrid retrieval for Agent Assist, an AI assistant for customer support representatives. Vanguard built a hybrid RAG system on Pinecone with dual dense and sparse embeddings, sparse embeddings trained in-house using BM25. The embeddings and their metadata are ingested and stored in Pinecone serverless. Once indexed, data is queried via hybrid retrieval with Alpha set at 0.5 for optimal precision for financial documents with domain-specific terms and abbreviations. As policies change and new documents are created daily, metadata filtering marks documents as live or stale daily, ensuring only live documents are accessed upon retrieval. Since deploying Pinecone, Vanguard has seen 12% improvement in result accuracy and a 70-95% reduction in token consumption per agent compared to dense retrieval alone, significantly reduced customer wait times through faster and more precise retrieval, and enhanced compliance through metadata tagging for audit purposes.