A client building an AI retrieval system hit a wall when their vector search pipeline moved beyond prototype. As embedding volume and user traffic increased, queries slowed down, memory pressure spiked, and filtering became a bottleneck. The stack worked in the lab, but production revealed latency issues and throughput ceilings that blocked scale.

In this talk, we’ll walk through how we benchmarked and tuned OpenSearch 3.x for vector workloads, comparing index configurations (HNSW vs IVF), adjusting parameters for recall vs latency, and optimizing filtering, hybrid scoring, and indexing strategies. We’ll share what worked, what didn’t, and the OpenSearch features that made scaling possible.

You’ll learn how to:
1. Choose and tune vector index types for real AI workloads
2. Handle filtering, recall, and latency trade-offs at scale
3. Stabilize performance with 3.x features like segment replication and async indexing

If you’re running semantic search, RAG pipelines, embeddings, or vector databases, this talk gives you field-tested lessons instead of theory or marketing slides.