Already using OpenSearch and Elasticsearch within its content stack, Thomson Reuters Labs evaluated both alongside Vespa for a new legal search application covering more than two million active U.S. statutes documents. The team selected Vespa for its ability to bring sophisticated search logic into a single application.

Vespa consolidated chunking, embedding, enrichment, and ranking within the search platform, reducing separate services and making experiments easier to implement and evaluate. Integrated learning-to-rank gave scientists and engineers control over custom features, feature logging for training, and model inference within the same framework.

The application moved from proof of concept into production serving customers, reducing the need to rebuild successful research in a separate production stack.

Read the Thomson Reuters Labs engineering blog

Technical resources for OpenSearch users

  • Lucene-Based Benchmark

    Compare Vespa and Elasticsearch across vector, lexical, and hybrid search, including concurrent updates. Relevant to Lucene-based architectures; results were measured against Elasticsearch, not OpenSearch.

    Read the benchmark

  • Etsy: Engineering the move to Vespa

    Hear how Etsy Engineering consolidated five retrieval engines onto Vespa, with practical lessons on partial updates, ranking, evaluation and migration.

    Watch the presentation

  • Vespa Terminology for OpenSearch Users

    Map familiar OpenSearch concepts to Vespa’s application packages, schemas, content nodes, and query processing. See where the architectures overlap and differ.

    Explore the guide

Why consider Vespa?

  • Higher performance and lower infrastructure costs at scale
  • Unified support for retrieval, ranking, and machine learning inference
  • Built for search, recommendations, and personalization
  • Designed for RAG and agentic AI applications
  • Optimized for real-time user experiences

Vespa vs. OpenSearch: compare capabilities

Compare how Vespa and OpenSearch support retrieval, ranking, inference, and operations. Consider which capabilities are built into each platform and where your application may require additional plugins, services, or engineering.

Capability
Vespa
OpenSearch
Retrieval methods
Keyword, vector, hybrid retrieval, and structured filtering
Keyword, vector, hybrid retrieval, and structured filtering
Ranking and reranking
Programmable multi-phase ranking with compiled ranking expressions, tensor operations, and model inference on data or coordinator nodes
Query scoring and rescore; model-based reranking through search pipelines.
Data updates
Real-time searchable updates without soft commits or transaction-log lookups; partial in-memory attribute updates
Near-real-time indexing; expensive partial updates
Best suited for
Real-time, large-scale AI applications requiring integrated retrieval and ranking
Enterprise search and analytics workloads

Independent evaluation

Despite the title, GigaOm’s Radar for Vector Databases assesses more than vector storage and similarity search. It covers ranking, hybrid retrieval, multimodal support, and scalability across 17 platforms, including OpenSearch and Vespa.

The report positions Vespa as a Leader and Outperformer, highlighting capabilities beyond vector search: programmable multi-phase ranking, integrated model inference, native tensor support, and multimodal retrieval.

Read the report

Recommended Resources