Weave Code
Evals Ui Laravel Package
vizra/evals-ui
Self-hosted Livewire dashboard for vizra/evals. Track score trends per suite, browse run history, drill into each sample's assertions and judge reasoning, compare baselines, and trigger runs from the browser. Install with composer and visit /evals.
Livewire dashboard for the Vizra Evals agent-evaluation framework.
Frequently asked questions about Evals Ui
- How do I install vizra/evals-ui in a Laravel app?
- Run composer require vizra/evals-ui and then visit /evals in your browser. There is no build step and no asset publishing required, because the stylesheet is served directly by the package. You do need vizra/evals installed and its tables populated, since the dashboard reads the data that package already writes.
- What are the requirements for vizra/evals-ui?
- The package requires PHP 8.4 or newer, Laravel 12 or newer, Livewire 3.7+ (or 4.x), and vizra/evals as the underlying evaluation package. If your app is still on an older Laravel major or PHP 8.2/8.3, this dashboard is not an option until you upgrade. Check your composer.json for Livewire before installing, since the dashboard is built entirely on Livewire components.
- Does vizra/evals-ui replace vizra/evals?
- No. vizra/evals-ui is the presentation layer: it renders the score trends, run history, and per-sample drill-downs on top of the tables vizra/evals writes. The evaluation engine, suites, assertions, and judge logic live in vizra/evals, which you install and configure separately. Think of the UI as the Horizon-style admin surface for your eval data, not the harness itself.
- How do I secure the /evals dashboard in production?
- The dashboard follows the Horizon pattern: it is open automatically in the local environment and denied everywhere else until you grant access. Define a viewEvalsDashboard gate, for example in AppServiceProvider::boot(), and return true only for users who should see eval results, such as admins. Until that gate is defined, non-local requests to /evals will be rejected, so you do not accidentally expose judge reasoning or sample data.
- Do I need to publish views, config, or run npm to use the eval dashboard styles?
- No. The stylesheet ships inside the package and is served from there, so there is no publish step and no Node build in your application. That keeps the install to a single composer require plus a visit to /evals. If you want to change markup or copy later, check the package documentation for what is publishable versus what is fixed in the shipped views.
- What does the vizra/evals-ui dashboard actually show?
- It shows score trends per suite over time, the full run history, and a drill-down into every sample so you can read each assertion result and the judge's reasoning. It also supports baseline comparison across runs, so you can see whether a prompt or model change moved scores meaningfully. On an empty install it renders an empty state that prompts you to write your first suite.
- Can I trigger evaluation runs from the browser instead of the CLI?
- Yes. Browser-triggered runs are one of the dashboard's features, so you can kick off an evaluation from the UI rather than dropping into a terminal. That is convenient for quick iteration after a prompt change, but for CI or scheduled regression checks you will still want the CLI or hosted CI checks. Be mindful of who has access to the trigger button when you scope the viewEvalsDashboard gate.
- How do I compare evaluation runs against a baseline?
- The dashboard includes baseline comparison across runs, so you can pick a reference run and see how later suites scored relative to it. This is useful when you change a prompt, model, or tool definition and want to know whether quality regressed or improved. Combined with score trends per suite, it gives you a quick read before drilling into individual failing samples and judge reasoning.
- Is vizra/evals-ui stable enough for production use?
- The package is pre-1.0 (v0.3.x), which under semver means minor version bumps can include breaking changes. Pin an exact version in composer.json rather than floating on ^0.3, and budget review time for each upgrade, especially if you rely on published views or config. The v0.3.2 release itself was documentation and empty-state copy only, with no schema or API changes, but that is not a guarantee for future releases.
- Does vizra/evals-ui load remote assets or phone home?
- The dashboard stylesheet is served by the package itself, so rendering the UI does not depend on a CDN or an external build pipeline. The repository README and vendor links do point at vizra.ai, including a promotional link to Vizra Cloud, so review what renders in your admin panel if you operate in an air-gapped or compliance-screenshot environment. If you need to strip vendor branding, confirm whether the relevant views are publishable before adopting.
Weaver
How can I help you explore Laravel packages today?