Search suggestions

GET /assets/suggest?query=par feeds the search bar autocomplete. It returns, in this order:

1. distinct attribute values matching the query, most frequent first (t is the attribute definition ID, tName its display name, locale the locale of the values when the attribute is localized);
2. collection and asset names (t is collection or asset, tId the item ID).

The locale of the user comes from the request headers (X-Data-Locale, then Accept-Language), resolved per workspace like the facets do (LocaleContext::getBestWorkspaceLocale()).

Attribute values

Every document of the asset index carries a suggestions nested field, built by AssetPostTransformListener from the asset attributes:

"suggestions": [
  {"definitionId": "<city>", "locale": "fr", "value": "Paris"},
  {"definitionId": "<city>", "locale": "_", "value": "Parc"},
  {"definitionId": "<country>", "entityId": "<e1>", "locale": "_", "value": "Germany"},
  {"definitionId": "<country>", "entityId": "<e1>", "locale": "fr", "value": "Allemagne"},
  {"definitionId": "<country>", "entityId": "<e1>", "locale": "en", "value": "Germany"}
]

There is one entry per distinct (definition, locale, value) of the asset, for the attribute types supporting suggestions (text, keyword, entity, IP) and values up to 300 characters:

• translatable text and keyword values keep the workspace locale of their attribute, _ when untranslated; non-translatable definitions are always stored under _;
• entities get one label per suggestion locale of the workspace (enabled locales, fallback locales and _): their translation for the locale, falling back to the base label. The attribute entity handlers keep these labels in sync when an entity is renamed, translated, merged or deleted.

SuggestionSearch runs, on the asset index, a nested aggregation with a filter and a multi_terms (definitionId + value) sub-aggregation:

• permissions: the query carries the same filters as the asset search (ACL, tag filter rules, not deleted, accepted status), and only the definitions flagged suggest that the user is allowed to read are aggregated;
• locale: per definition, the aggregated nested documents are restricted to the user's best workspace locale plus _ for translatable text and keyword attributes, to the best workspace locale only for entities (hence exactly one label per entity), and to _ for the other definitions. Values in other locales are never suggested;
• deduplication: one bucket per (definition, value), whose doc_count is the number of visible assets holding the value; the fr and _ copies of the same value merge into one suggestion;
• the filter sub-aggregation, inside the nested one, keeps only the nested documents matching the query, so a multi-valued attribute does not leak its other values.

The highlight of each value comes from a top_hits sub-aggregation.

Caveats

• Changing the enabledLocales or localeFallbacks of a workspace requires re-populating its assets (as for the attrs field): the entity labels are indexed for the locales known at indexing time.
• An asset holding the same value both translated and untranslated counts twice in the ordering of that value; the count is never displayed.
• Elasticsearch caps the nested documents of an asset (index.mapping.nested_objects.limit, 10 000 by default): each entity value costs one entry per workspace locale.

Upgrading from the attribute index

Attribute values used to be indexed one by one in a dedicated attribute index, with the asset permissions copied on every document. That index does not exist anymore:

1. re-populate the asset index so that the documents get their suggestions field:

   bin/console fos:elastica:populate --index=asset

2. delete the orphan attribute indices (<ELASTICSEARCH_INDEX_PREFIX>attribute_<env>*), for instance:

   curl -X DELETE "${ELASTICSEARCH_URL}/${ELASTICSEARCH_INDEX_PREFIX}attribute_prod*"