How Vanguard Uses Pinecone to Boost Customer Support with 12% More Accurate Responses

Vanguard, one of the world's largest investment management firms, has long prioritized delivering exceptional client experiences — including responsive, knowledgeable customer support. With millions of clients relying on Vanguard for retirement planning, investments, and financial advice, the quality and speed of support interactions carry real financial consequences. The company's Center for Analytics and Insights (CAI) team, operating within the Chief Data Analytics office, was tasked with modernizing how customer service representatives access information during live calls.

The core challenge was a retrieval problem. Vanguard's support teams were using keyword-based search to locate relevant financial documents, but this approach only surfaced links to lengthy source files — leaving agents to manually sift through dense content to find specific answers. This inefficiency drove up call times and eroded customer satisfaction. During peak periods like tax season, Vanguard's traditional workaround was to hire additional representatives to absorb the volume, adding significant operational cost without addressing the root cause.

To move beyond keyword search, the CAI team first experimented with JSON storage and cosine similarity-based retrieval. These early solutions proved too slow, struggled to scale with growing datasets, and frequently returned results that lacked contextual relevance. The team then evaluated a range of vector database options — including pgvector, Faiss, and Redis — before selecting Pinecone. Key decision factors included Pinecone's support for hybrid search (combining BM25 sparse embeddings with dense embeddings), real-time indexing capabilities, advanced metadata filtering for compliance, and enterprise-grade security features such as AWS PrivateLink. Pinecone also worked directly with Vanguard to provision a dedicated AWS account and cluster tailored to their security and performance requirements.

The resulting system, called Agent Assist, is an internal RAG-powered chat assistant built on top of Pinecone Serverless. Financial documents stored as HTML pages are scraped, preprocessed with a custom chunking strategy, and encoded into dual dense and sparse embeddings — with sparse embeddings trained in-house using BM25. Hybrid retrieval is configured with an Alpha value of 0.5 to balance precision across domain-specific financial terminology. To ensure agents always access current information, documents are tagged daily as "live" or "stale" using metadata filtering, with outdated documents archived to DynamoDB for regulatory compliance.

Since deploying Agent Assist, Vanguard has seen measurable gains across accuracy, efficiency, and compliance. Hybrid retrieval improved search result accuracy by over 12% compared to dense-only retrieval. Call times dropped as agents could surface precise answers in real time, and the team no longer needs to scale headcount during peak seasons to manage volume. Metadata tagging also introduced stronger audit traceability, supporting Vanguard's compliance obligations. Looking ahead, Vanguard plans to expand its use of RAG and Contextual-Aware Generation (CAG) systems, with Pinecone serving as a foundational layer in its broader AI knowledge ecosystem.