Taleva Achieves 90.3 on PeopleSearchBench Recruiting

Taleva achieved an Overall score of 90.30 on the 30-query Recruiting subset of PeopleSearchBench, an open benchmark for AI-powered people-search systems. That is 22.07 points above the strongest official published Recruiting baseline when the benchmark's three component scores are combined using its equal-weight Overall formula.

The score comes from a completed Taleva benchmark run. The supporting evaluator output, candidate files, per-query metrics, and run configuration are retained internally and available for investor due diligence. This public release presents the headline result, benchmark methodology, system boundary, and comparison with the official published baselines.

Download the full technical report (PDF)

Key takeaways

• Taleva achieved 90.30 Overall on PeopleSearchBench's 30 recruiting queries.
• The strongest official Recruiting baseline is Lessie at 68.23 Overall, calculated from the three component values published by the benchmark.
• The resulting difference is 22.07 points on the benchmark's 0–100 scale.
• The result measures the complete Taleva search system, not one language model or prompt in isolation.
• The supporting run archive is retained internally and available for investor due diligence.

The result at a glance

PeopleSearchBench defines Overall as the equal-weight mean of Relevance Precision, Effective Coverage, and Information Utility. The official repository publishes the Recruiting component scores for each baseline but not a separate Recruiting Overall column, so the baseline Overall values above apply that published formula and round to two decimal places.

The Taleva run and the official baseline runs were conducted separately. The comparison is useful context, but it is not a controlled head-to-head under a single frozen environment.

What PeopleSearchBench measures

PeopleSearchBench contains 119 natural-language queries across four scenarios: Recruiting, B2B prospecting, deterministic or expert search, and influencer or key-opinion-leader discovery. Taleva's announced result covers only the 30-query Recruiting subset because it is the category aligned with Taleva's product and data.

The benchmark evaluates three dimensions:

• Relevance Precision asks whether the returned people fit the request and whether the strongest matches appear first.
• Effective Coverage asks whether the system completes the task and returns enough qualified people.
• Information Utility asks whether each result contains enough relevant, actionable information to be useful.

PeopleSearchBench decomposes a request into checkable criteria and verifies candidate claims against profile data and live web evidence. This is more structured than assigning one holistic score to a result list, although parts of the evaluation still depend on model-assisted judgement and live search infrastructure.

What the result measures inside Taleva

The score is a system-level result. A Taleva search begins with a recruiter's natural-language brief and passes through several connected stages:

1. The search agent interprets the role, location, experience, skills, and contextual preferences in the brief.
2. Constraints that can be represented reliably become structured retrieval filters.
3. Requirements needing contextual judgement become profile-level evaluation criteria.
4. The retrieval system builds a candidate pool from Taleva's professional-profile index.
5. Candidate evaluation ranks and explains the strongest matches before they are delivered.

PeopleSearchBench evaluates the final candidate output. It therefore does not isolate the contribution of the agent prompt, retrieval index, ranking logic, profile coverage, evaluation models, or result presentation. The 90.30 result should be read as a measurement of the combined Taleva search system.

Scope of the result

The result establishes that Taleva achieved 90.30 Overall on the Recruiting subset, 22.07 points above the strongest official published Recruiting baseline. Its scope is deliberately precise.

The result does not extend to:

• a per-metric win, because Taleva's three component scores are not included in this edition;
• a confidence interval, because repeated scoring runs and query-level samples are not included;
• state-of-the-art performance across every people-search task, because this result concerns Recruiting only; or
• a controlled comparison with systems evaluated on different dates or configurations.

Why evaluation conditions matter

People-search evaluation is sensitive to time. Professional profiles change, search indexes refresh, web evidence moves, and model providers revise systems. A candidate list can remain fixed while its later verification result changes. Result depth, timeouts, retries, and judge configuration can also affect the final score.

PeopleSearchBench publishes its queries, scoring definitions, baseline results, and evaluation code in an open repository. That transparency makes the benchmark methodology inspectable and provides a consistent framework for comparing people-search systems.

What the full technical report includes

The downloadable report explains:

• the benchmark's scenarios and scoring dimensions;
• how the official Recruiting baseline Overall values were calculated;
• the boundaries of the Taleva system being evaluated;
• threats to validity, including web-index and judge drift;
• the exact scope of the result; and
• a privacy-aware evidence checklist for diligence and external replication.

Read the Taleva PeopleSearchBench–Recruiting technical report

Supporting evidence and next steps

The supporting run archive is retained internally and available for investor due diligence. A privacy-reviewed external package can include the benchmark commit, Taleva deployment identifier, run date, query count, result depth, evaluator configuration, failure accounting, aggregate component scores, and per-query metrics. Where candidate-level data cannot be redistributed, hashes and a controlled reproduction path can provide provenance without exposing personal data.

Future robustness work can repeat evaluation over frozen candidate lists, retain every failed or zero-result query, and report score variation across runs. This would extend the completed result with an externally reproducible statistical package.

Frequently asked questions

Frequently Asked Questions

Did Taleva score 90.3 on the entire PeopleSearchBench benchmark?+

Where are the supporting run artifacts?+

How was the 68.23 Lessie baseline calculated?+

Does the result measure Taleva's AI agent by itself?+