Metadata and attribute filtering for AI search

Starting with 17.4, the AI assistant can now constrain content search by document attributes — for example content type, language, or folder path — in addition to keywords. Editors can ask natural-language questions such as “find news articles under products” and get results filtered and constrained by those attributes.

How it works

  1. Metadata generation. Metadata are generated upon the ingestion of a content item into the RAG. See below for the default attributes that get stored in the RAG, see also here on how to alter what goes in your RAG.

  2. Filterable attributes catalog. The platform informs the AI which metadata attributes are available for using in searches.

  3. Attribute-aware search. The assistant uses those attributes to narrow results, alone or together with keyword search, when the user requests it.

Attribute filtering is under active development. Exact-value filters work, for example, you can ask the assistant to find docs with locale 'en'. Friendly names such as "English docs" are not mapped to locale codes yet. Date ranges can't be used as filters in 17.4. More metadata will become filterable in future releases

Default metadata attributes

These attributes are indexed by default and can be used when filtering search results:

FieldMeaningHow to use in assistant
idDocument identifier (uuid)Literally, "find document with id [uuid]"
nodeNameRepository node nameLiterally, "find document with name [name]"
contentCategoryContent category"find Documents [Pages, Resource bundles"]
displayNameDisplay name shown in the CMS"find document named [any term in the display name]
createdAtCreation timeNot supported for filtering in 17.4
lastModifiedAtLast modification timeNot supported for filtering in 17.4
publishedAtPublication timeNot supported for filtering in 17.4
translationIdTranslation group identifier (uuid)Reserved: will be used with translation support
localeLanguage / localeCurrently use literally, "find docs with locale 'de'"
contentTypeDocument type (namespace-qualified name)Use content type friendly name, e.g "find Banners"
pathFolder path in the repositoryFull or partial path match, e.g. "find News under 'archive'

Optional attributes with no value are left out of the index rather than stored as empty strings, in an effort to treat “missing” values in a consistent way across all supported vector stores.

Note that a re-indexing of the RAG is needed if you are on pre 17.4 version. See here for more information.

Share Feedback
Page: /content-ai/metadata-and-attribute-filtering-for-ai-search
Section: Content AI
Category *
Metadata and attribute filtering for AI search | Bloomreach Content Documentation