Exa AI Research Blog | Semantic Search & Neural Network Search Engine

Exa has launched a state-of-the-art people search benchmark to enhance its AI search engine capabilities, particularly benefiting sales, recruiting, and market research teams. The new system involves a hybrid retrieval method combining Exa embeddings with Exa search, capable of indexing over a billion people and handling 50 million updates weekly. The initiative includes the release of an open-source people search evaluation, featuring a dataset with 1,400 queries and an evaluation harness for reproducibility. Analyzing 10,000 historical people queries revealed three primary search categories: role-based search, skill or role-based discovery, and specific individual lookup. Data generation for these categories involved creating synthetic queries for executives and founders across various company sizes, ensuring profiles have public verification but are not widely known to prevent memorization. The evaluation methodology employs recall and NDCG metrics for targeted lookups and uses an LLM judge for role-based discovery due to the subjective nature of results, validating matches through online profile content.