AI App Context Sandbox

Draft and publish states, testing against the analysis agent, field-level validation, Reset to published version, and the no-history and single-app limits.

The AI App Context Sandbox is the AI App Context tab in App Version Details, where a builder edits the AI context yaml for one version of an app, saves it as a draft, tests it against the analysis agent, and publishes it when satisfied. It replaces editing App Context at the listing level, where every save went live immediately for all versions.

You may also see the AI App Context Sandbox referred to as the sandbox, the context editor, semantic authoring, semantic model, or the AI Context tab; builders often ask about it as "where do I edit what Intelligence knows about my app?"

Applies to: Insight Cloud, Create area, App Version Details. Internal to Seek and Seek-supported publisher organizations today.You'll need: The Builder role (the Build App permission) or Seek Admin. Org Admins and those with View Creator Area permissions can view but not edit. Other roles do not see the Create area.

Last reviewed: 2026-09-18

Where to find the AI App Context Sandbox

In the left sidebar, click Create, open an app from App Management, choose the version in Select app version, then click AI App Context in the left panel.

What you can do with the AI App Context Sandbox

  • Edit the AI context yaml document for one specific app version, independently of other versions.

  • Save context as a draft without affecting subscribers.

  • Ask the analysis agent test questions that use the draft context, in isolation from the live experience.

  • See validation errors that name the exact field and explain the fix, then correct them yourself.

  • Publish the draft so the version's live AI context updates for all subscribers.

  • Reset to published version to discard a draft and start again from the live context.

How the AI App Context Sandbox works

Each app version has its own AI context. When a new version is created it inherits the previous version's context, and you can then change the new version's context without touching older versions.

Every version has at most one draft and one published context. Saving writes the draft. Publishing replaces the published context with the draft. Subscribers are only impacted by published context.

When you save a draft, the sandbox validates it against the app's Snowflake schema and structure. If validation fails, each error shows the field, the error type, and a plain-language explanation.

Test questions run against the draft context for that app only. Saving a new draft clears the test conversation and starts a fresh session against the new draft. Within one session you can carry on a conversation. Test runs are traced in Langfuse with the context that was tested.

Example

A builder publishes version 3 of a CPG app, which adds two tables. Version 3 inherits version 2's context. The builder upgrades to Version 3 and runs the default variant for the publisher org. The builder opens version 3 in AI App Context, adds the new tables to the context, saves the draft, and asks "What were total units by region last month?" The agent answers using the draft. The builder compares the SQL to the verified query in the context, adjusts a metric description, saves again, re-tests, and publishes. Version 2's context is unchanged, so subscribers still on version 2 are unaffected.

Limits and what the AI App Context Sandbox can't do

  • The AI App Context Sandbox does not keep a history of drafts or published contexts. There is one draft and one published context per version; you cannot revert to an earlier draft or an earlier published version. Reset to published version is the only rollback.

  • The AI App Context Sandbox does not save test conversations. Saving a new draft clears the chat, and conversations do not persist across page refreshes.

  • The AI App Context Sandbox tests one app at a time. It does not run test questions across multiple apps.

  • Test questions require the app's variant to have run. If it has not, the sandbox shows a "variant not found" error and asks you to run it first.

  • The sandbox agent gives the same facts, numbers, and conclusions as Intelligence for the same question and the same published context. Wording can differ between the two, because the agents are not deterministic.

  • The AI App Context Sandbox does not change how contexts are stored or integrate with SBT.

  • The AI App Context Sandbox is built for apps published with SBT 1.0 (app specs), which keep a context per version. Legacy apps do not keep per-version history in the platform.

AI App Context Sandbox vs. App Context on the listing

Until the Sandbox ships, App Context is edited at the bottom of an app's listing (Org Settings → Publishing → Catalog → Edit App Listing → AI Context), and saving there applies to every version at once. The Sandbox moves editing to App Version Details, per version, with draft and publish states. See App Context for what the context document contains.

If the AI App Context Sandbox isn't working

If the AI App Context tab says This section is coming soon, the Sandbox has not been enabled for your organization yet. If a test question returns "variant not found," run the app's variant and try again. If validation errors do not make sense after fixing the named field, contact Seek support with the app name, version, and the error text.

Related articles

  • App Context: what the AI context document contains and how Intelligence uses it

  • App Version Details: the page that hosts the AI App Context tab

  • App Management: where you open an app from