An assistant that retrieves whole pages can spend much of its context on repeated navigation, background and duplicated claims. Reducing that material can save time, but the selection process becomes part of the evidence chain. If it removes an exception or a qualification, the answer can become more confident while becoming less faithful to the source.
Exa’s October 19 release of Exa 3.0 combines a larger embedding model, an expanded index and generally available Dynamic Highlights. The company says its highlight model selects material across results, assigning more space to stronger pages and reducing repetition. It reports substantially fewer tokens than full-page text and introduces a lower-cost Lite search option alongside Auto. These are provider-reported comparisons and product terms at release. Primary source.
Selection changes what the agent sees
The release’s broader coverage includes product listings, local businesses and time-sensitive information. That can help an application find useful records, but coverage and correctness remain separate questions. A retrieved listing may be relevant without being current, and a short extract may omit the date needed to interpret it.
An evaluation should therefore preserve both the highlight and the underlying page. Reviewers can check whether an incorrect answer began with a missing source, a poor selection within the source or a reasoning error after retrieval. Those failures call for different fixes. Merely increasing the answer model’s size may not repair missing evidence.
Measure the complete request
The immediate comparison for a developer is a fixed set of real questions run through full-page and highlight-based retrieval. Track answer support, qualification retention, latency and total cost. Include questions whose correct answer depends on a small exception, a date or disagreement between sources; these are more informative than easy fact lookups alone.
The Lite and Auto options also invite a task-specific choice. A cheaper request may cost more overall if it causes extra searches or manual review. Conversely, a higher-cost search may provide no useful gain for a straightforward lookup.
More selective retrieval could make grounded assistants economical across a wider range of workloads. The next evidence is stable answer quality after index and source changes, with selection failures visible rather than hidden inside an aggregate benchmark. The product release supplies new controls; the application still needs its own evidence standard.
References
Questions, news tips or corrections? Contact [email protected].
