Vespa
used in
1
stack
Search and AI serving platform combining vector, text, and structured-data retrieval with machine-learned ranking.
Apache 2.0 licensed platform originally developed at Yahoo, written in Java and C++. Combines vector search, full-text search, and structured queries with native tensor computation and distributed machine-learned ranking evaluated at serving time. Designed for billions of constantly changing documents and thousands of queries per second, and used in production by companies including Spotify and Perplexity. Available self-hosted or as the managed Vespa Cloud service.
Strengths
- Unified hybrid retrieval over vectors, text, and structured data in one engine
- Native tensor formalism for running ranking and inference at serving time
- Proven at very large scale with real-time writes and sub-100ms query latency
- Long production history predating the current vector-database wave
Tradeoffs
- Steep learning curve; application packages, schemas, and ranking expressions are complex to master
- Heavy operational footprint when self-hosting compared to single-binary vector stores
- Overpowered for simple embedding-lookup use cases where a lighter store suffices