A local museum has photographs, tools and oral-history records in a single searchable catalog. Visitors want to narrow a search by collection, decade and media type without learning database query syntax. Azure AI Search provides facet buckets and counts for fields marked facetable, while query filters can narrow the results. Neither a colorful sidebar nor a count is proof that a record may be disclosed. This guide designs a public catalog search with illustrative records and numbers, not a benchmark of a live Azure index.
Identify the visitor’s actual filtering task
Begin with a list of questions visitors ask: show photographs from the textile collection; limit results to the 1930s; select audio records that have a public transcript. Those are fields a user can understand and that staff can reliably maintain. A raw internal storage path, donor ID or permission code is not a useful public facet. Choose a small set of low-cardinality labels, keep display labels consistent, and define how an object with multiple collection tags should appear.
Separate the search page from the records’ access policy. A public visitor should never see a restricted interview simply because a facet name happens to match. Access filtering must be applied server-side before or as part of retrieval, using trusted permission metadata and user identity. Do not rely on a user interface that hides a checkbox after downloading unauthorized results. Build a test with one deliberately restricted synthetic record.
Mark eligible index fields as facetable
Microsoft explains that facets can operate on supported string, date, numeric and other nonvector fields when configured in the index. Vector fields themselves are not facetable. A museum may mark collection, medium and decade-group as facetable, and keep long free-text notes as searchable content rather than sidebar categories. Keep a controlled vocabulary for media type so photo, photograph and image do not become three buckets that mean the same thing.
Facetable is an index-schema property; setting it does not clean bad metadata or supply access control. Changing existing index configuration may require an index rebuild or migration depending on the field change. Plan schema updates carefully, preserve object IDs, and compare a sample of records before switching users to a new index. The most useful facet is one whose meaning is stable across curatorial revisions.
Request facet buckets together with search
A query can include a facets array, such as a field name for collection, and return @search.facets with values and counts alongside matching documents. The counts describe documents associated with each bucket for that query context. They are not global inventory totals, and the same record may contribute to multiple buckets in a collection-valued field. A visitor choosing one collection can change which other buckets are relevant.
The application should display a selected facet as an active filter that can be removed. If a filter yields no results, offer a clear reset instead of silently changing the user’s query. Review whether the UI needs a static set of available categories or only current-result categories; Microsoft documents approaches using separate queries for stable navigation. The process image distinguishes index configuration, bucket query and actual filtering.
Treat counts as query-dependent estimates
In distributed search, shard-level top-N facet collection can undercount or omit less frequent values. Microsoft documents count options and the possibility of increasing the requested bucket count to improve completeness, with a latency tradeoff. Therefore a sidebar badge of 17 is not necessarily an auditable museum inventory count. The article’s example figures are invented to show the relationship between visible buckets and result sets.
If the catalog requires exact administrative totals, calculate them through a separate validated reporting method rather than presenting UI facet counts as financial or legal evidence. Test category fields with more distinct values than the default bucket limit and with multilingual synonyms. Consider avoiding a facet for hundreds of unique accession numbers: the sidebar would be hard to use and shard behavior could be surprising.
Compare a facet click with a permission check
Imagine the public search finds 24 synthetic textile-related records. The user selects the photograph facet, narrowing to nine candidate matches; the server then enforces a public-access filter and displays only authorized items. The comparison figure shows visibly different lists at all three stages. The precise values are illustrative and should not be described as Azure test output.
Do not interpret “9 photographs” as “9 public records” before permission criteria have been applied. In a properly designed application, security filtering is included in the actual query so restricted records never travel to the browser. The diagram makes the permission boundary explicit for readers, but the implementation should not create an unsafe unfiltered client response as an intermediate production step.
Design useful facets for mixed historical metadata
A catalog record can belong to both textile and education collections. In that case, both buckets may count it, and summing the visible facet counts can exceed the number of distinct search results. Museum staff should decide whether that is helpful or confusing for visitors. Display a short explanation when multi-valued categories overlap. Use human-readable decade labels based on verified catalog dates rather than inferring them from automatically generated descriptions.
Avoid faceting on sensitive fields such as acquisition restrictions, contact details or volunteer identity. If a collection contains embargoed materials, consult the access team about whether aggregate counts themselves are sensitive. Accessible keyboard controls, readable selected-state labels and a no-JavaScript fallback may be more valuable than an elaborate animated filter bar. Test screen readers against the actual result count announcement.
Test edge cases before a public launch
Build a synthetic index with single-value and multi-value fields, missing dates, empty collections, duplicate tags, and documents with access restrictions. Run a query for a broad topic and one for a narrow title. Compare the facet buckets, selected filters, document IDs and permissions. The checker should catch an unauthorized record even if its descriptive title and thumbnail look harmless.
Check that switching facets does not break pagination or leave stale result counts on screen. Test changes in catalog records and search index refresh delays. If no buckets appear, diagnose field configuration separately from query filters and missing content. Preserve a request identifier and index version in operational logs without logging private text unnecessarily. A clear review trail helps staff distinguish a usability error from a permission defect.
Publish a catalog sidebar that tells the truth
Launch with a few high-value categories, explain what counts represent, and keep a visible reset path. Revisit bucket labels when curators merge or rename collections, and keep old search links working through documented redirects. Permission checks must continue to function if a visitor manually edits a URL or calls the API without the interface. A search facet is a navigation aid, never the sole authorization system.
The durable gain is a searchable collection that can be explored without overstating the meaning of counts. Follow Microsoft’s documentation for syntax and preview status before enabling advanced hierarchical or aggregated facets. When in doubt, favor simple categories with reliable public records and make unsupported totals explicit instead of inventing precision.
How We Research
For our November 2026 editorial review, we consulted the named official developer sources: Azure AI Search faceted navigation; Microsoft faceted navigation examples. We compared documented feature scope, request shapes, output limitations, and failure cases with a clearly hypothetical nonproduction museum or community-history scenario. We did not run paid API calls, process private recordings, access donor or visitor records, benchmark vendors, measure recognition accuracy, or capture actual product screenshots. All illustrations are original explanatory drawings rather than real output or proof of performance. Prices, availability, privacy terms, model support, and preview features can change; check the current official documentation and an authorized local test before deploying.
Frequently Asked Questions
No. Apply trusted access filters on the server or query layer independently.
No. Microsoft describes facets for supported nonvector fields.
A document may have multiple facet values, and shard-level collection can affect returned bucket counts.
Not in every configuration. Check shard and bucket-limit behavior before treating them as administrative totals.
Microsoft describes additional hierarchical capabilities as preview, so verify support and preview conditions.