Real-Time Recommendations: How ZoomInfo Reached 50x Peak Throughput
Blog post from Pinecone
ZoomInfo transformed its contact recommendation system from costly, stale offline batch processing into real-time vector search across more than 390 million contact embeddings using Pinecone, improving relevancy and recall by 2x, increasing user engagement by over 50%, and supporting 50 times more peak requests. Its retrieval engine supports contact lookalikes, personalized recommendations, and API-driven agent workflows by representing both structured and unstructured contact information as embeddings and reranking search results. Scaling occurred through three main steps: partitioning vectors into namespaces to narrow search scope and reduce latency and costs, replacing synchronous LlamaIndex calls with direct asynchronous Pinecone SDK operations to enable concurrency, and moving sustained high-volume workloads from on-demand infrastructure to Dedicated Read Nodes after read-unit rate limits became a bottleneck. Dedicated Read Nodes provide dedicated, warm query infrastructure with fixed hourly pricing and no read rate limits, while replicas can be added to scale throughput and shards expand storage; on-demand indexes remain more suitable for variable or unpredictable traffic.
| Trend | Post Mentions | Total Month Mentions | Posts | Companies | MoM |
|---|---|---|---|---|---|
| Vector Search | 7 | 2,199 | 456 | 148 | -17% |
| Real-time | 6 | 6,521 | 1,586 | 287 | -12% |
| MCP | 1 | 11,103 | 967 | 228 | -17% |
| Serverless | 1 | 1,056 | 257 | 105 | -44% |
Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.