Yahoo's search engine grew up and went open source

What it does

Vespa is a full-stack search and AI serving platform that handles vector search, tensor operations, text search, and structured data in one system. It selects data from large corpora, runs machine-learned models on that data, organizes and aggregates results, and returns them — typically in under 100 milliseconds — while the underlying data keeps changing. The project is the complete open-source codebase; new releases ship from master every weekday morning.

The interesting bit

Most vector databases bolt search onto storage. Vespa inverts that history: it started as Yahoo’s large-scale search engine and absorbed vectors, tensors, and ML inference into the same distributed serving layer. The result is a system already proven at “hundreds of thousands of queries per second” on production internet services, now repurposed for modern RAG and AI workloads without being retrofitted.

Key highlights

• Combines vector search, tensor math, text search, and structured queries in a single request path
• Runs ML model inference at serving time across distributed nodes
• Targets sub-100ms latency for search, recommendation, and personalization use cases
• Ships new releases Monday–Thursday from master; Apache 2.0 licensed
• Java and C++ core; Java modules build on any platform with JDK 17 and Maven 3.8+

Caveats

• Full builds require AlmaLinux 8; macOS developers are limited to Java modules or Docker
• The README notes the getting-started application is “fully functional and production-ready” but immediately suggests adding nodes for redundancy — a hint that single-instance setups are a starting point, not an endpoint

Verdict

Worth a look if you’re building high-traffic search, recommendation, or RAG pipelines and would rather not duct-tape a vector database onto a separate serving layer. Probably overkill if your dataset fits in memory on one machine and your latency requirements are measured in seconds, not milliseconds.