Large Scale Entity Resolution with Milvus at 6sense | Milvus Meetup SF

At the Milvus Meetup in San Francisco, Premal Shah (Co-founder and Chief Architect at 6sense) presented “Large Scale Entity Resolution with Milvus at 6sense”. 6sense manages a massive identity graph of over 600 million B2B people records to power account-based marketing, buyer intelligence, and intent data. In this deep dive, Premal explains why their legacy keyword-based matching on Elasticsearch hit a hard scaling wall as identity resolution workloads grew to tens of thousands of queries per second. He covers the transition from brittle rule-based pipelines to dense vector representations, and why keeping 600M high-dimensional embeddings in memory required aggressive architectural optimization. Premal details how the team slashed memory and hosting overhead by quantizing embeddings from FP32 down to INT8 without sacrificing recall, and why CPU-based HNSW indexing struggled to hit their ambitious 30,000 QPS target due to proxy bottlenecks. Finally, he shares how 6sense unlocked ultra-high throughput by adopting GPU-accelerated indexing with NVIDIA CAGRA on Milvus, reaching 30K+ QPS at sub-millisecond query execution speeds, comparing the total cost of ownership (TCO) between CPU and GPU clusters, and outlining their roadmap for native hybrid search. What you'll learn: The limitations of keyword heuristics and Elasticsearch for high-throughput B2B entity resolution Transitioning 600M+ people records into semantic vector embeddings with custom fine-tuned models FP32 vs. INT8 quantization: how to cut memory footprint and infrastructure TCO while preserving matching accuracy Benchmarking open-source Milvus on CPU HNSW and understanding proxy throughput bottlenecks at scale Deploying GPU-accelerated vector indexing with NVIDIA CAGRA to achieve 30,000+ QPS retrieval throughput A practical TCO and infrastructure breakdown: comparing CPU scale-out vs. GPU scale-up architectures The future roadmap: combining vector search with full-text keyword retrieval in hybrid search pipelines Chapters: 0:00 Introduction & 6sense Overview 1:43 The Challenge: Entity Resolution Across 600M+ B2B People Records 3:15 Limitations of Elasticsearch & Keyword Search 4:10 Moving to Dense Vectors & Custom Embedding Models 6:12 FP32 to INT8 Quantization: Saving Memory & Infrastructure Costs 7:20 Testing Vector Databases & Open-Source Milvus Deployment 9:05 Benchmarking CPU HNSW: Bottlenecks at High QPS 12:05 Unlocking Ultra-High Throughput with GPU CAGRA 14:40 Results: Achieving 30,000+ QPS on Massive Datasets 16:48 CPU vs. GPU Performance & Cost (TCO) Breakdown 18:22 Future Roadmap: Hybrid Search in Milvus 19:30 Q&A: Entity Resolution, Lakehouse & Scaling Bottlenecks This talk was recorded at the Milvus Meetup in San Francisco (525 Market St). The event brought together search, ML infrastructure, and data platform engineers to explore the algorithms, infrastructure, and production patterns behind modern AI-powered retrieval systems. Watch the full event recording: Milvus Meetup San Francisco October 2026 (Full Event): https://youtu.be/itISrq5J-Ik Watch other talks from this meetup: Milvus 3.0: A More Powerful Retrieval Engine: https://youtu.be/j9qjA73Jj-I From Retrieval to Action: Building Production Agents on AWS: https://youtu.be/wmbsMV87AQQ Find us on: Zilliz Website: https://zilliz.com Milvus Website: https://milvus.io Zilliz LinkedIn: https://www.linkedin.com/company/zilliz/ Milvus LinkedIn: https://linkedin.com/company/67143129 Milvus X: https://x.com/milvusio Milvus GitHub: https://github.com/milvus-io/milvus Try Zilliz Cloud: https://cloud.zilliz.com/signup