vespa

Vespa is an AI‑search platform that combines text, vector, tensor, and structured data search in a single system. It offers real‑time updates, distributed inference for ranking, and support for generative AI and recommendation use cases.

If you need to build a search or AI application that must handle billions of items and deliver results in under 100 ms, Vespa is designed for that scale. The platform lets you query over vectors, tensors, text, or structured fields and then apply a trained relevance model to rank the results. You can also retrieve documents, call machine‑learned models for additional inference, and stream the answers to your application in one request. The same engine is used for hybrid search, where vector similarity and keyword relevance are combined, and for recommendation systems that mix content retrieval with model scoring.

For developers working on recommendation, personalization or navigation in e‑commerce, Vespa provides a single source for product data, a search API, and a ranking API that takes structured and unstructured attributes. The platform also supplies a “streaming search” mode that reduces indexing costs for personal or private search scenarios, saving up to 20 × compared with full indexing. The documentation lists case studies from Spotify, Yahoo, and Farfetch, showing deployments that meet sub‑100 ms latency and need to scale to thousands of queries per second. Each of these customers integrates keyword analysis, vector search, and ranking models inside Vespa to serve highly relevant content to users.

The platform is available as a fully managed cloud offering on AWS, or can be self‑hosted. Vespa Cloud includes continuous deployment, upgrades, and strong security controls. It also ships a GitHub repository with samples and a blog that covers new features such as the fast‑search map fields. The pricing page states that a free trial is available, but no specific pricing tiers are published on the site.

Key features

• Distributed search engine for text, vectors, tensors, and structured data
• Integrated machine‑learned ranking inference for relevance
• Sub‑100 ms latency at scale of billions of items
• Hybrid search combining vector similarity and keyword matching
• Streaming search mode for personal/private queries, 20× cheaper than indexing
• Fully managed deployment on AWS with continuous upgrades