VESPA-AI

Vespa.ai Yahoo OSS hybrid search + ML inference.

Définition

Vespa.ai features : (1) Tensors : core abstraction multi-dimensional tensors (vectors, matrices, embeddings), tensor computation language YQL extensions. (2) Index types : inverted index (BM25 full-text), ANN HNSW (vector search), columnstores (numeric, attribute fields). (3) Hybrid search : YQL Yahoo Query Language combines text + vector + filters + ranking expressions custom tensor math. (4) Ranking phases : multi-phase ranking (first-phase fast cheap, second-phase expensive ML model inference top-N), TensorFlow + ONNX + XGBoost + LightGBM model inference at query time. (5) Distributed architecture : Cluster Controller + Distributor + Content Nodes + Container Nodes, automatic data redistribution + replication. (6) Stateful + stateless : combine document storage + indexing + inference. (7) Schemas : .sd schema files declarative model definition. Vespa Cloud SaaS offering 2021+ managed. Customers : Yahoo (~1B+ documents indexed), Spotify (recommendations playlists), Spotify Wrapped 2023 powered Vespa.ai, Wayfair, AliExpress.

Origine

Vespa origine Yahoo search engine 2003+ internal ; open-sourced 2017 sous Apache-2.0 ; Vespa.ai Inc. spinoff Yahoo 2023 (focus growth managed offering Vespa Cloud) ; Jon Bratseth founder + ex-Yahoo Search lead.

Exemple en contexte

Spotify Wrapped 2023 personalised yearly summary uses Vespa.ai : ingest user listening history + audio features 600M users, ANN HNSW retrieve similar listening clusters, tensor compute personality genre breakdown, ML model inference 'your top genre', delivered ~500M Wrapped pages within hours.

Termes liés

• Elasticsearch vector — alternative search engine.