diff --git a/public/images/docs/evaluation/explore-playground/columns-evals.gif b/public/images/docs/evaluation/explore-playground/columns-evals.gif new file mode 100644 index 00000000..35212130 Binary files /dev/null and b/public/images/docs/evaluation/explore-playground/columns-evals.gif differ diff --git a/public/images/docs/evaluation/explore-playground/evaluations-list.png b/public/images/docs/evaluation/explore-playground/evaluations-list.png new file mode 100644 index 00000000..ac75cf8c Binary files /dev/null and b/public/images/docs/evaluation/explore-playground/evaluations-list.png differ diff --git a/public/images/docs/evaluation/explore-playground/filter-evals.gif b/public/images/docs/evaluation/explore-playground/filter-evals.gif new file mode 100644 index 00000000..53c04887 Binary files /dev/null and b/public/images/docs/evaluation/explore-playground/filter-evals.gif differ diff --git a/public/images/docs/evaluation/explore-playground/tags-evals.gif b/public/images/docs/evaluation/explore-playground/tags-evals.gif new file mode 100644 index 00000000..dede66ef Binary files /dev/null and b/public/images/docs/evaluation/explore-playground/tags-evals.gif differ diff --git a/src/lib/navigation.ts b/src/lib/navigation.ts index 96db3f7f..72f09e72 100644 --- a/src/lib/navigation.ts +++ b/src/lib/navigation.ts @@ -327,10 +327,9 @@ export const tabNavigation: NavTab[] = [ { title: 'Explore playground', items: [ - { title: 'Overview', href: '/docs/evaluation/guides/explore-playground' }, + { title: 'The Evaluations page', href: '/docs/evaluation/guides/explore-playground' }, { title: 'Test an eval', href: '/docs/evaluation/guides/explore-playground/test-an-eval' }, { title: 'Usage & analytics', href: '/docs/evaluation/guides/explore-playground/usage-analytics' }, - { title: 'Eval logs', href: '/docs/evaluation/guides/explore-playground/eval-logs' }, ] }, { title: 'Create a custom eval', href: '/docs/evaluation/guides/custom-evals' }, diff --git a/src/pages/docs/evaluation/guides/explore-playground/index.mdx b/src/pages/docs/evaluation/guides/explore-playground/index.mdx new file mode 100644 index 00000000..45fdb0f5 --- /dev/null +++ b/src/pages/docs/evaluation/guides/explore-playground/index.mdx @@ -0,0 +1,71 @@ +--- +title: "The Evaluations page" +description: "Browse, filter, and open every eval from the Evaluations page" +--- + +The Evaluations page is your eval library. Every built-in eval and every one your team creates lives here, so this is where you browse what's available, narrow down to the evals you need, open one to test or edit it, and create your own. + +The Evaluations page listing evals with their type, output type, and tags, plus search, tag filters, a Columns control, and a Create evals button +*The Evaluations page: browse every eval, filter by tag, rearrange the columns, or create your own* + +## The evaluations table + +Each row is one eval, and the columns tell you what it is at a glance: + +- **Evaluation Name**, the eval's name, like `groundedness` or `prompt_injection` +- **Type**, whether it runs on its own (single) or aggregates several child evals (composite) +- **Eval Type**, how it reaches its verdict: [Agent, LLM-as-Judge, or Code](/docs/evaluation/concepts/eval-types) +- **Output Type**, what it returns: pass or fail, a percentage score, or a category label, covered in [Output types](/docs/evaluation/reference/output-types) +- **Tags**, the categories the eval belongs to, like Red Teaming, RAG, or Safety +- **30 day chart** and **30 day error rate**, its recent run volume and failure rate +- **Created By**, System for a built-in eval or your workspace for a custom one +- **Last updated**, when the eval last changed + +The screenshot shows the default columns. A few more are hidden, like **Versions**, how many versions the eval has, and you turn them on from the Columns control. + +## Tags + +Every eval carries one or more **tags** that say what it checks and where it fits, and they are how you make sense of a library this size. They fall into a few kinds: + +- **Concern**, the risk the eval guards against: Safety, Red Teaming, Data Leakage, Hallucination, Harmful Objects +- **Technique or check**, how it measures: RAG, Retrieval Systems, NLP Metrics, Output Validation, Output Format, Code +- **Modality**, the kind of content: Image, Audio, Text, PDF, Conversation +- **Domain**, the use case: Medical, Finance, Agents, Chatbot behaviors + +These are the main tag chips; individual evals also carry finer tags like Quality, Bias, or Data Privacy. An eval usually carries more than one, so `groundedness` is tagged both RAG and Retrieval Systems, and `prompt_injection` is tagged Red Teaming. + +## Find an eval + +Three controls narrow the list, and they stack: + +- **Search** by name in the search box +- **Tag chips**, the row under the search: click a tag to show only the evals that carry it, and stack tags to narrow further +- **Filter**, for finer control than the chips: build a condition on name, type, eval type, output type, tags, or who created the eval. Unlike a single tag chip, you can combine conditions, like Output Type is Pass/Fail *and* Created By is your workspace + +Clicking tag chips on the Evaluations page to filter the list down to the evals that carry each tag +*Click a tag to filter the list to the evals that carry it, and stack tags to narrow further* + +Opening the Filter dropdown on the Evaluations page and building a condition on a field like name, eval type, output type, or tags +*Open Filter and build a condition on any field, from eval type to tags to who created it* + +## Arrange the columns + +The **Columns** control lets you show, hide, and reorder the table columns. Keep the ones you care about, like Output Type and Tags, and drop the rest so the table shows what matters to you. + +Opening the Columns control on the Evaluations page and toggling which columns show in the table +*Open Columns to show, hide, and reorder the table columns* + +## Create your own + +**Create evals**, at the top right, opens the flow to build a [custom eval](/docs/evaluation/guides/custom-evals): give it a name, write its rule, and pick its output type. It then joins the library alongside the built-in ones, filterable and taggable the same way. + +## Dive deeper + + + + Every built-in template with its type and required inputs + + + Build your own from a rule and an output type + +