Launch Week: Pinecone for agents, search, recommendations, and more

Launch Week is here at Pinecone! Every day this week, we're rolling out new features to better support you in building accurate and performant AI applications at scale in production. This blog will serve as your go-to hub for all announcements, updated daily with the latest reveals. Stay tuned, and follow us on socials to keep up with the action!

Day 1: Optimizing Pinecone for agents (and more)

Agentic workloads differ from traditional search and recommendation systems, involving millions of small namespaces with bursty, unpredictable query patterns.

Over the past year, we’ve evolved our serverless architecture to tackle these challenges with innovations across both our query and write paths, including:

• Adaptive indexing using log-structured merge trees
• Support for millions of namespaces per index—without sacrificing performance or cost

We’ve also boosted search and recommendation performance with:

• Enhanced metadata filtering
• High-performance sparse indexing for keyword search
• Automatic replication at high QPS (w/ provisioned capacity options coming soon), and more 

Pinecone serverless can serve 4x as many queries with roughly 1/8th the latency, on half the compute footprint. In this benchmark, Pinecone produces the same results with topK 10 or 100.

Pinecone sparse retrieval latency vs. OpenSearch. Learn more in our sparse retrieval deep dive (https://www.pinecone.io/learn/sparse-retrieval/).

Improved performance of metadata filtering in Pinecone. For high cardinality metadata, the filters are applied much more efficiently boosting recall.

Learn more about these performance optimizations in our deep dive by CEO Edo Liberty and Principal Product Manager Nathan Cordeiro.