Vespa

Vespa is the open-source search engine that grew out of Yahoo's search program and is run today by Vespa.ai: one Apache-licensed engine that stores text, vectors and structured fields, updates them as streams, and runs machine-learned models inside the query path, so ranking logic is code you version. The vendor pitch is industrial numbers — scaling to billions of constantly changing data items, thousands of queries per second with latencies below 100 milliseconds, and a streaming update mode billed as roughly twenty times cheaper than re-indexing. Self-hosting is free software with a real operational weight; Vespa.ai also runs the managed Vespa Cloud, deployed on AWS and other clouds, with a free trial and pricing against continuously running nodes rather than per-query meters.

That billing shape is the judgement: Vespa rewards teams with steady traffic and corpora that never sit still, because you pay for nodes and updates instead of Algolia-style per-search fees, and burst traffic is capacity you pre-bought or auto-scaled, not a meter you feared. Residency is structurally yours — on self-host every byte stays in your account, and the cloud lets you pin regions. The tax is the learning cliff: application packages, schemas, tensor ranking features — below a few million documents it is overkill. Where hybrid lexical-plus-vector relevance at billion-document scale is the actual requirement, few engines have more battle-tested years under load than Vespa.