This case study examines the development and deployment of Pair Search, a modern search engine system developed by the Singapore government to improve access to Parliamentary records (Hansard). The project represents a significant step forward in making government information more accessible while also preparing for the future of AI-enabled information retrieval. **Project Context and Challenges** The Hansard database contains over 30,000 parliamentary reports dating back to 1955, presenting several key challenges: * Legacy search was purely keyword-based, leading to poor result quality * Documents spans multiple decades with varying formats requiring standardization * Need to serve both human users and AI systems effectively * Requirement for high performance while using sophisticated algorithms **Technical Architecture and Implementation** The system is built on Vespa.ai as the core search engine, chosen for its scalability and advanced text search capabilities. The search process is implemented in three distinct phases: *Document Processing Phase* The team tackled the complex task of standardizing decades of parliamentary records into a uniform format suitable for modern search operations. This involved careful consideration of changing data formats and structures over time while maintaining the integrity of the historical records. *Retrieval Phase* The system implements a hybrid approach combining: * Keyword-based search using Vespa's weakAnd operator * BM25 and nativeRank algorithms for text matching * Semantic search using e5 embeddings, chosen for their balance of performance and cost-effectiveness compared to alternatives like OpenAI's ada embeddings *Re-ranking Phase* A sophisticated three-phase approach was implemented to maintain speed while using complex ranking algorithms: * Phase 1: Initial filtering using cost-effective algorithms * Phase 2: ColbertV2 model-based reranking for improved relevance * Phase 3: Global phase combining semantic, keyword-based, and ColbertV2 scores into a hybrid scoring system **Production Deployment and Monitoring** The system has been soft-launched with specific government departments including: * Attorney General's Chambers * Ministry of Law legal policy officers * Communications Operations officers at MCI and PMO * COS coordinators Performance metrics are being actively monitored, including: * Daily user count (~150) * Daily search volume (~200) * Result click-through patterns * Number of pages viewed before finding desired results **Future Development Plans** The team has outlined several strategic directions for system enhancement: *Data Expansion* * Planning to incorporate High Court and Court of Appeal case judgments * Exploring integration with other government data sources *Search Enhancement* * Implementing LLM-based index enrichment through automated tagging * Developing question generation capabilities * Exploring query expansion using LLMs to improve retrieval accuracy *RAG Integration* The system is designed to serve as a retrieval backend for RAG applications, with specific focus on: * Providing API access for both basic search and RAG-specific retrieval * Supporting the Assistants feature in Pair Chat * Enabling integration with other government LLM applications **Technical Lessons and Best Practices** Several key insights emerge from this implementation: *Architecture Design* * The three-phase search approach helps balance speed and accuracy * Hybrid scoring systems outperform single-metric approaches * Careful attention to document processing and standardization is crucial for historical data *Model Selection* * E5 embeddings provide a good balance of cost and performance * ColbertV2 reranking adds significant value to result quality * Combining multiple ranking approaches yields better results than relying on a single method *Production Considerations* * The system maintains high performance despite complex algorithms through careful phase design * API-first design enables broader application integration * Continuous monitoring of user interaction metrics guides optimization **Impact and Results** The initial deployment has shown promising results: * Significant positive feedback from government users * Improved search result quality compared to the previous system * Successfully handling a growing user base with consistent performance * Recognition at the parliamentary level, with mention by the Prime Minister This case study demonstrates the successful implementation of modern search and RAG capabilities in a government context, showing how careful attention to architecture, model selection, and user needs can result in a system that effectively serves both human users and AI applications. The project also highlights the importance of planning for future AI integration while maintaining current performance and usability standards.