Query fails on old indices when returning a semantic text field and source is disabled

[Mikep86]

Elasticsearch Version

9.2

Installed Plugins

No response

Java Version

bundled

OS Version

n/a

Problem Description

In 9.2, we made a breaking change to exclude vectors from _source by default. Additionally, we tied that index setting to control if semantic field embeddings are included in _source. These two changes have combined to create a problem when:

• Querying an index created before 9.2 (i.e. it includes vectors in _source by default)
• A semantic_text field is in _source
• _source is disabled by the request

The failure mode is indicated by:

      {
        "shard": 4,
        "index": "properties",
        "node": "HoRCRAz6Sr2Awtz2BaKorw",
        "reason": {
          "type": "illegal_argument_exception",
          "reason": "can not write type [class org.elasticsearch.xpack.inference.mapper.SemanticTextField]"
        }
      }

Steps to Reproduce

1. Create an index with a semantic_text field in a version prior to 9.2
2. Add some docs to the index
3. Upgrade to 9.2
4. Query the index with _source disabled such that a doc with a populated semantic_text field is returned

GET my-index/_search
{
  "_source": false,
  "query": {
    <some query that returns a doc with a semantic text field>
  }
}

Logs (if relevant)

No response

[elasticsearchmachine]

Pinging @elastic/search-relevance (Team:Search - Relevance)

[Mikep86]

This has been fixed in #136312. It's also worth noting that as part of the fix, we decoupled the index.mapping.exclude_source_vectors index setting from _inference_fields metafield behavior. _inference_fields inclusion is now completely controlled by request settings.