Splore AI: Elasticsearch to Vespa for Semantic and Hybrid Retrieval

Client Overview

Splore AI specializes in streamlining complex business processes using AI technologies, focusing on enhancing decision-making and operational efficiency. They serve industries such as finance, legal, and tech by integrating generative AI and multi-agent systems. However, their existing search infrastructure, based on Elasticsearch, was unable to meet the demands of their growing, data-intensive operations.

Challenge

Splore AI needed to improve their search infrastructure to keep up with the increasing volume and complexity of their data. Their previous system was struggling with scalability and the advanced search features required for their evolving needs. Specific challenges included:

• Scalability & Performance: The existing system couldn't handle petabytes of data across multiple indices with fast retrieval.
• Relevance Ranking: Search results were limited to timestamp-based ranking, without machine learning optimization.
• Hybrid Search Needs: While Splore AI had a vision for advanced AI powered search, their Elasticsearch setup couldn't adequately support the future needs at the time.

Solution

After evaluating the limitations of Elasticsearch, Searchplex recommended Vespa for its advanced capabilities in hybrid search, ranking optimization, and handling large-scale data. The solution was tailored to address Splore AI's unique needs.

• Migration to Vespa: We mapped Elasticsearch's schema to Vespa's format, optimized data structures for fast retrieval, and built efficient data ingestion pipelines. Additionally, we incorporated user roles and access permissions into the schema to meet business requirements.
• Hybrid Search Integration: Vespa's hybrid search capabilities combined BM25, TF-IDF, and vector search in a single, unified ranking pipeline.
• ML-based Ranking: Developed multiple ranking strategies tailored for different communities and user roles to improve retrieval quality, ensuring results were more relevant and personalized.
• API & Performance Optimization: Developed a custom Domain-Specific Language (DSL) for search queries and optimized indexing and retrieval for sub-second response times.

Result

The transition to Vespa provided significant improvements in both performance and functionality, enabling Splore AI to meet their search demands at scale:

• 50x Improvement in Search Latency: Query response times were reduced from over 1 second to ~20ms.
• Enhanced Search Relevance: Vespa's hybrid search model improved the blending of keyword and semantic search, delivering more accurate results.
• Future-Ready Architecture: The solution provides a foundation for future machine learning-driven ranking and personalization, enabling ongoing improvements. Splore AI now enjoys a scalable, high-performance search infrastructure that supports both their current and future needs, empowering their AI-powered business process automation solutions with superior search capabilities. Teams evaluating migration to Vespa or broader Vespa architecture work can use this as a concrete reference point for what search modernization looks like in practice.