// the find
vespa-engine/vespa
The AI search platform
Vespa is Yahoo's production search and vector database platform, open-sourced and battle-tested at scale serving hundreds of thousands of queries per second. It handles the full pipeline: ingest, index, ANN vector search, tensor computation, ranking with ML models, and serving — all in one system. The target is teams who need more than what a standalone vector DB gives them but don't want to bolt together five separate services.
First-class tensor computation at serving time means you can run re-ranking models (ONNX, XGBoost, LightGBM) directly inside the query path without a separate inference server. The combination of exact lexical search, ANN vector search, and structured filters in a single query is genuinely difficult to replicate by composing Elasticsearch with a vector DB. Daily releases Monday–Thursday from main, with a factory build pipeline visible at factory.vespa.ai, is an unusual cadence that signals active maintenance. The schema definition language and application package model (services.xml + .sd files) give you a declarative, version-controlled configuration surface that scales to multi-node clusters without rewriting application code.
The operational footprint is heavy — C++ and Java components, AlmaLinux 8 as the supported build host, OSGi bundles, and a bespoke config server mean you are running a serious distributed system, not a library you drop in. The learning curve is steep: the query language (YQL), the ranking expression framework, and the schema language are all Vespa-specific and require real study before you're productive. Self-hosting requires genuine infrastructure discipline; if you can't commit to that, you're pushed toward Vespa Cloud, which adds cost and vendor dependency. Documentation is thorough but dense and assumes you already understand distributed search concepts — someone coming from 'I just need vector similarity search' will find it overwhelming.