How ZoomInfo Uses Pinecone to Deliver Real-Time Contact Recommendations at Scale

ZoomInfo provides sales and marketing teams worldwide with access to hundreds of millions of professional contact records, enriched with firmographic data and AI-powered search capabilities. For its customers, the ability to quickly identify the right person inside a target account—and act on that information—is directly tied to pipeline and revenue. Even small improvements in how contacts are surfaced translate into significant time savings and better outcomes for go-to-market teams.

ZoomInfo’s existing contact discovery experience required users to manually search, filter, and navigate large volumes of contact data. Identifying the most relevant buyers inside a target account could take hours of manual work. The platform lacked a way to surface ranked, personalized recommendations the moment a user viewed a company profile—a gap the Applied AI team set out to close.

ZoomInfo’s Applied AI team built a real-time contact recommendation system using a ~400M-parameter text embedding model and Pinecone as the vector database. Within three weeks, the team had loaded millions of embeddings and validated sub-second latency without manual index configuration. Pinecone’s serverless slab architecture stores vectors in large contiguous units, ensuring writes and reads proceed in parallel without blocking. For production-scale traffic, the team deployed Pinecone Dedicated Read Nodes—isolated read replicas with warm memory and local SSD—delivering predictable low-latency performance under sustained high-QPS workloads. The system scaled to over 390 million vectors across more than 100,000 namespaces without re-architecting the underlying infrastructure.

The production system achieved a P50 latency of ~60ms at ~40 QPS, keeping end-to-end recommendation service latency under one second for standard loads and under five seconds for peak traffic above 100 RPS. Relevancy improved 2x, user engagement rose by more than 50%, and the system served 50x more peak customer requests than the prior implementation. What had required hours of manual research was reduced to a single click: the moment a user views a company profile, the platform surfaces the most relevant contacts instantly. As Carlos Nunez, VP of Engineering and Applied AI, described it: the team went from managing infrastructure to improving the recommendation model itself.

ZoomInfo plans to expand Pinecone-powered recommendations across additional products, customer segments, and internal applications. The deployment demonstrates a broader shift in B2B software: platforms that can surface the right information instantly, rather than after manual search, are redefining what go-to-market intelligence means. With Pinecone’s combination of serverless scale, slab-based storage, and Dedicated Read Nodes, ZoomInfo has a clear path to extend real-time semantic search to new surfaces without operational overhead.