Exa Upgrades Highlights: 500-character summaries match the accuracy of 8,000-character full texts.

AIMPACT Update, April 24 (UTC+8): According to monitoring by Beating, AI search API company Exa has launched a major upgrade to its Highlights feature. Highlights is Exa’s capability to extract dense, query-relevant summary snippets from web pages, generated in real time per request (not cached) and completed within 100ms. On the SimpleQA benchmark, 500-character Exa Highlights match the accuracy of 8,000-character original full texts, reducing token consumption by 16x; 4,000-character Highlights outperform 32,000-character full texts. Improvements are especially pronounced in long-document scenarios such as technical documentation, API references, SDK guides, and research papers: with a 500-character budget, Highlights achieve 60% accuracy versus just 6% for full texts. Exa’s proprietary /answer, Deep, and Websets agent endpoints all use Highlights as their underlying retrieval mechanism, without accessing original page content. Highlights is not a new feature but an existing content extraction option within the Exa API. This upgrade occurs at the model level—所有 requests enabling the enableHighlights parameter automatically benefit without requiring code changes. (Source: BlockBeats)