July 2020 Summaries

In June 2020, Vespa introduced several enhancements, including support for approximate nearest neighbor vector search using the HNSW algorithm, which allows efficient combination with filters and text search, offering millisecond response times while processing hundreds of millions of documents per node. Vespa also improved its unique Streaming Search feature, increasing read throughput up to fivefold through better parallelism, benefiting applications like personal and high-write searches. Additionally, new rank features were introduced, such as optimized fieldMatch and the globalSequence feature for document ordering, despite some limitations in systems with changing node indexes. Vespa, largely developed by Yahoo engineers, continues to evolve with community feedback and contributions, serving as a robust open-source big data processing and serving engine used by various Yahoo and Verizon Media products.

NLP has undergone a significant transformation with the advent of Transformer models like BERT, which excel in various NLP tasks and have garnered widespread adoption due to the ease of use provided by resources like the Hugging Face Transformers library. Vespa, a distributed application platform, offers a unique approach to deploying these models by moving computation to data, thus reducing network bottlenecks and complexity. It enables efficient evaluation of machine-learned models, like Transformers, by automatically distributing them across content nodes, which handle data-dependent computations. This approach is particularly advantageous for search and recommendation applications that require evaluating numerous data points per query. Vespa's flexibility allows for combining different models and computations, such as using BM25 for initial ranking followed by a Transformer model for re-ranking, and it supports models from various platforms like TensorFlow and PyTorch. The platform's capabilities are demonstrated with a sample application using the MS MARCO dataset, showcasing the integration of a small Transformer model for text ranking while highlighting future improvements aimed at enhancing performance and model support.