Project settings

    Rename a project, set its default judge Model, read retention and content capture, and delete a project knowing exactly what goes with it.

    After this page you can fix a project’s name without breaking a link, pick the judge that New run and Online rules preselect, know what is kept and whether prompts are captured, and delete a project with its consequences named first.

    General

    Project settings has tabs General · Environments · Models · Review. General shows the name (editable by an organization Owner or Admin — URLs use ids, so a rename breaks nothing), the default judge Model (a Model of this project; what New run and Online rules pre-select), created by and when, a retention card and a content capture card, and the danger zone.

    Retention and content capture

    Retention comes from the organization’s plan (retentionDays): traces older than that are deleted per Environment. A platform admin can put a retention hold on an organization so its traces stop expiring; the card says when one is on. Content capture is an SDK-side setting (capture_content), so the card is inferred per Environment from recent spans — captured, not captured, or unknown when there are no traces — and links to the snippet to change it. Nothing here is saved; see Configuration and privacy.

    Delete a project

    Owner or Admin only. The confirm dialog counts what goes with it — datasets, runs, Models, Environments, online rules, alerts, monitors, review queues, comments, prompts, agents, simulation runs, RL datasets, topics — and asks you to type the name. Deletion removes the project and then cascades asynchronously module by module: datasets, runs, Models, playground history, online rules, alerts, review queues, comments, prompts, agents and simulation runs, RL datasets and topics are deleted; Environments are deactivated, not deleted, so their tenant id and stored traces stay addressable until retention removes them. Organization-scoped objects — evaluators, collections, Providers — are untouched. Export datasets first if any hold golden data; a failed step in the cascade is logged and does not undo the delete.

    The other tabs: Environments and keys, Models and providers and, for review score types, Score types and review queues.

    EvalKit is built by Syntropylabs. Published on PyPI and npm.