Gemini Website Search

The Search workspace publishes a conventional documentation search experience from the same documents as Gemini RAG. It is a separate component from the Assistant: queries run entirely against the App's local RDBMS index and do not call a Gemini model.

The local index is maintained for every File Store whether or not a Search widget has been created. This keeps ingestion simple and lets you add Search later without changing import definitions. Existing File Stores are detected and queued for indexing when the App starts; Rebuild index can also explicitly queue every current document.

Local indexing and RDBMS support​

The C# extension selects its search implementation from the configured OrmLite dialect provider's DbKind:

Native full-text support is initialized automatically. If the database feature is unavailable, cannot be created with the current permissions, or a native query fails, Search transparently uses a bounded LIKE query over document titles, headings, and content. SQL Server's Full-Text Search is an optional server component and must be installed in the SQL Server instance to use the native provider.

PDF, Word, PowerPoint, and Excel documents can still be uploaded to Gemini, but are not locally searchable unless their text is first converted into a supported text or Markdown document. Their local Search status explains why no sections were created.

Create and tune a Search widget​

Choose New Search and configure its visitor-facing title, input placeholder, optional launcher tooltip, and no-results message. Document scope uses the same category, doc type, status, locale, product, version, and tag dropdowns as Assistants. The server enforces this scope, so a host page cannot expand it.

When a File Store combines several websites or repositories, visitors can search all of that knowledge from one box without first knowing where an answer lives. Per-deployment scope can expose the full corpus or only an approved slice, and ranking weights let you promote authoritative titles, heading matches, preferred document types and fresher content consistently across every source.

Search first asks the active database provider for matching candidates, then applies one consistent ranking model across every RDBMS. You can tune each part while the Test the local index results refresh automatically:

The same saved ranking configuration is used by the test panel, live preview, and published widget. Changing a weight issues a fresh query, providing immediate feedback without merely reordering an incomplete client-side result set. Each control's default value and accepted range is listed under Ranking defaults below.

Search interaction and result behavior​

The Search dialog queries as the visitor types, debounces requests, cancels superseded work, and keeps the previous result list visible until the next response is ready. Matching text is bold and underlined in light themes and bold white in dark themes by default; its highlight color can be overridden.

Results are grouped by document and use the Markdown frontmatter title when available. Visitors can:

• open Search by clicking the launcher, pressing Ctrl/⌘+K, or pressing /;
• move through results with Up and Down and open the selection with Enter;
• load additional results automatically by scrolling;
• press Escape or the esc button to close the active layer;
• return from an open document with Escape or the left-arrow button; and
• revisit recently opened documents, retained in that browser's localStorage.

When a result has a Source URL, the published widget navigates to that canonical page and anchor. The administrative live preview opens it in a new browser window. When a Markdown document has no Source URL, the widget opens a second dialog containing a sanitized rendered copy; Escape closes that document first and returns focus to the Search results.

If Search and Assistant widgets are embedded on the same page, Search keeps Ctrl/⌘+K and the Assistant automatically moves to Ctrl/⌘+Shift+K. The / shortcut can be enabled or disabled independently. Shortcuts do not fire from editable inputs.

Appearance and embedding​

Search uses Shadow DOM isolation and serves its saved configuration together with the widget as one self-contained classic script:

<script
  src="https://app.example.com/chat/ext/gemini/public/searches/widget.js?g=abc123"
  async>
</script>

Customize Auto, Light, Dark, Nord, Matrix, or Soft Pink, the font family, highlight color, launcher style, corner, and top/right/bottom/left offsets. Auto first follows the host page's color-scheme value in localStorage when it is light or dark, then falls back to the visitor's operating-system preference.

As with Assistants, the launcher can float in any corner or render inside a host element using its saved Mount element selector or a data-mount override:

<span id="search-slot"></span>
<script
  src="https://app.example.com/chat/ext/gemini/public/searches/widget.js?g=abc123"
  data-mount="#search-slot"
  async>
</script>

Set exact or wildcard Allowed origins and a per-client request limit before publishing. Search uses the same familiar deployment lifecycle as Assistants: Save draft, Publish, Unpublish, Regenerate ID, Archive, Restore, and typed confirmation before permanent deletion. Regenerating the ID invalidates every old embed immediately.

Ranking defaults​

Results combine database relevance with configurable weights. Change any weight below the Live Preview to re-query immediately against the same term, so relevance is tuned against real results rather than a reordered client-side list. The same saved configuration is used by the test panel, the live preview, and the published widget.

Weights accept half-steps. Freshness uses sourceUpdatedAt, falling back to the upload and then the creation time. Document type preference lists every docType present in the store; 0 is neutral, positive values promote and negative values demote, and up to 50 types can be weighted. Reset defaults restores the whole table to the values above.

A result opens its canonical Source URL; Markdown without one opens a rendered document preview.

Customization reference​

Every setting is stored per Search deployment and served with the widget. Values are validated and clamped on the server, so an out-of-range value is corrected rather than rejected.

data-mount on the script tag overrides the saved Mount element, and data-mount="none" forces the floating launcher. An invalid or unmatched selector logs a console warning and falls back to floating.

Use Run diagnostics before release to validate publication, File Store access, index contents, allowed origins, and the widget endpoint.

Use Run diagnostics before release to validate publication, File Store access, index contents, allowed origins, and the widget endpoint - see Operations & Troubleshooting.

See Search Analytics & Privacy for search-quality metrics and optional page-view analytics, or AI Assistants to publish a grounded chat widget alongside Search.