Vespa

Vespa is an open-source AI search platform and vector database from Norway's Vespa.ai, for RAG, search and recommendation, self-hosted or on Vespa Cloud.

Vespa lets an application search, rank and run machine-learned models over vectors, tensors, text and structured data in a single query. It is aimed at systems where retrieval quality decides the outcome, such as generative AI answers, AI agents, product search and recommendations. The engine began at the web search engine AllTheWeb in 2001 and later grew up at Yahoo. Since 2023 it has been developed by Vespa.ai, an independent company spun out of Yahoo and based in Trondheim, Norway. Perplexity, Spotify and Yahoo run on it. As a vector database, Vespa builds HNSW indexes on tensor fields, applies metadata filters inside the nearest neighbor search and stores several embeddings per document. Hybrid queries mix dense embeddings with keyword signals and filters in one ranked result. Phased ranking adds second-phase and global-phase expressions, including cross-encoder rerankers. Embedders such as the Hugging Face embedder run inside the platform when documents are fed and when queries arrive. Binarized vectors and paged attributes on disk let indexes grow beyond memory, while content clusters spread data across nodes and keep replicas through a node failure. Streaming search gives each user a personal index over their own data. Vespa is built for engineering teams that run search or recommendation at large scale and want retrieval, ranking and inference in one engine rather than several services. Teams that want full control self-host the open-source engine on their own hardware or on Kubernetes, and can buy a support contract for it. Teams that would rather not operate clusters run the same engine on Vespa Cloud, a managed service on AWS that rolls out new platform versions automatically. An enclave option keeps the data plane inside the customer's own cloud account. Typical deployments answer questions over company documents with retrieval-augmented generation and rank products on e-commerce sites that mix text, images and structured navigation. Others select content or ads for personalization by combining retrieval with machine-learned model evaluation. Data scientists bring ranking models from XGBoost, LightGBM, TensorFlow or ONNX, and Vespa evaluates them while the query runs. It suits teams whose search results, recommendations or agent answers are the product their users see.