Description
Description
When querying _field_caps, semantic_text fields are returned as a text field type. The intention is that these fields should "just be text fields", but there are differences, such as the multi_match query not working for semantic_text fields, or semantic_text not being queryable cross-cluster. Getting the index mappings directly is suggested as an alternative but this is very bad for performance (because of the payload size) and does not work cross-cluster.
Either semantic_text should be recognizable as such in _field_caps or the feature gap needs to be closed.
Use cases
For knowledge bases in the Obs AI Assistant, we leverage search connector indices. The user types in a prompt, and then we want to use the prompt to query these indices. For that to work we need to understand what fields should be queried as text. We can do this by calling _field_caps on those indices with type: text. However because we don't understand what fields are semantic_text we cannot use all text queries. This currently means that we have to create match queries for each text field instead of creating a multi_match query. There are probably other queries as well that do not work on semantic_text fields. On a case-by-case basis we can work around this but generally we need the capability to determine upfront what kind of queries can be executed and which cannot in order to prevent internal errors.