Exa's People Search Benchmarks (exa.ai)

Exa has unveiled a cutting-edge people search system that effectively indexes over a billion profiles using a hybrid retrieval model, which combines fine-tuned Exa embeddings with their proprietary search technology. This framework is engineered to handle more than 50 million updates weekly, making it a significant resource for industries such as sales, recruiting, and market research, where precise profile searches are essential. Exa's approach not only focuses on the efficient retrieval of information but also includes an open-source evaluation dataset that establishes a new benchmark for people search on the web. The launch is particularly notable for its methodology, which analyzes historical queries to categorize them into three primary types: role-based searches for specific job positions, skill or role-based discovery, and individual lookups by name and affiliation. By generating targeted queries and utilizing LLMs for complex search evaluations, Exa aims to distinguish its technology based on actual retrieval performance rather than mere memorization. This innovative evaluation strategy, alongside a structured taxonomy developed using Claude Opus 4.5, positions Exa's people search system as a potentially transformative tool in the AI/ML landscape, driving more effective and relevant searches in a data-rich environment.