Retro Relevance: Lessons Learned Balancing Keyword and Semantic Search

“Semantic search has revolutionized how we approach information retrieval, offering a nuanced understanding of human language that surpasses traditional keyword-based search methodologies for natural language queries. However, this technology is not without its limitations. This presentation critically examines scenarios where semantic search falls short of expectations, failing to consistently deliver the most relevant results. Semantic search’s challenges and limitations include niche queries and domain-specific jargon. While semantic search excels in understanding context and semantics, it sometimes overlooks the importance of exact matches and user expectations, leading to a gap between delivered results and user satisfaction.

To bridge this gap, our presentation talks about leveraging traditional search relevance building blocks with an eye toward performing hybrid queries that combine traditional BM25 keyword search with semantic search. Some of the tools we use include using query rules to contextually pin documents to the top of the result set, and synonyms to capture domain specific language that embedding models may not necessarily know about.

In conclusion, our talk will offer practical strategies for enhancing search relevance that can be complementary to semantic search. By adopting a comprehensive approach that includes synonyms, hybrid queries, and pinned documents, we can improve the performance of semantic search systems, ensuring they meet the evolving needs and expectations of users.”

Kathleen DeRusso

Elastic

Kathleen DeRusso is a Principal Software Engineer at Elastic. She works on the Search Relevance team providing tools, guidance and intuitive user workflows to empower developers building search use cases through the Elastic stack. Prior to joining Elastic, she has several years' experience building applications that use search. Kathleen is also an avid supporter of initiatives to support women in technology.