Results
Exa’s People Search benchmark contains 1,400 natural-language queries covering roles, locations, and other profile criteria. We evaluated all 1,400 queries, achieving 94.35% precision, 94.86% R@1, and 96.21% R@10. We compare aggregate performance with published results from other systems.
| System | Precision | R@1 | R@10 | Queries |
|---|---|---|---|---|
| Thielos | 94.35% | 94.86% | 96.21% | 1,400 |
| Wrangle Deep † | 91.10% | 94.79% | 98.86% | 1,400 |
| Sixtyfour † | 82.80% | 90.40% | 96.60% | — |
| Juicebox † | 79.00% | — | — | — |
| Exa | 63.30% | 72.00% | 94.50% | 1,399 |
| Brave | 30.20% | 44.40% | 77.90% | 1,373 |
| Parallel | 26.90% | 20.80% | 74.70% | 1,387 |
Sources: Thielos evaluation, Exa benchmark results, and vendor-reported figures (†).
Precision
Average share of returned profiles judged relevant.
R@1
Share of queries with a relevant person ranked first.
R@10
Share of queries with at least one relevant person in the first ten.
Results ranked by score. Navy: Thielos. Solid gray: results published by Exa. Hatched gray: vendor-reported results.
Metric definitions
Precision
How much of the returned list is relevant?
8 relevant profiles out of 10 → 80% for that query
R@1
Is the first result relevant?
Relevant person ranked first → query succeeds
R@10
Is there a relevant person in the first ten?
Relevant person anywhere in the top ten → query succeeds
Illustrative examples. Filled squares represent relevant profiles; outlined squares represent irrelevant profiles. Faded squares are not assessed in the R@1 example.
Evaluation protocol
- Dataset
- Exa People Search · 1,400 public queries
- Evaluation
- All 1,400 queries · October 21, 2026
- Returned depth
- Up to ten ranked profiles per query
- External evaluator
- GPT-6 Luna · Exa People Search rubric
- Evaluation records
- Query-level outputs and grading records retained