Offline benchmark of the retrieval agent across all labelled benchmark CVs: mean Precision@5 against the ground-truth relevant roles. A retrieved job counts as relevant when its track matches the CV's labelled track and its seniority is compatible with the candidate's level. The label is never exposed to the embedding or retrieval step, so there is no label leakage.
Each run parses every labelled CV, retrieves roles semantically and re-ranks them exactly as the live agent does.
No benchmark run yet. Click the button above to run the agent evaluation.