semantic_text — one-field semantic search #

Semantic search used to mean wiring two fields: a text field for the source, a dense_vector field with vector_options pointing back at it. The ES-style semantic_text type (Elasticsearch 8.15+ alignment) collapses all of that into one declaration:

PUT /my-collection
{
  "schema": {
    "properties": {
      "body": {
        "type": "semantic_text",
        "service": "zai",
        "model": "GLM-4.6V",
        "dims": 3072
      }
    }
  }
}

From there everything derives. Writes carry plain text — the vector is embedded server-side at write time, in bulk and after partial updates alike. Search needs no field names and no vectors:

POST /my-collection/_search
{"query": {"semantic": "documents about search engines"}}

# hybrid without the retriever DSL
{"query": {"bool": {"should": [
  {"semantic": "search engines"},
  {"match": {"field": "title", "query": "engines"}}
]}}}

The field is pure semantic, like its ES namesake: the text stays in _source but is never inverted-indexed — match does not hit it, the vectors do. Keep a sibling text field for lexical legs.

The binding to the embedding backend is strict, so a mapping can never silently point at the wrong service:

• model must be listed by an AI service — exactly one, or the mapping names service explicitly. An unlisted model, or one listed by several services, is a 400 at mapping time, with the candidates named.
• dims may be omitted when the service declares it on its models entry ({"id": "GLM-4.6V", "dims": 3072}); a conflicting hand-written value is caught at mapping time, and a backend answering with a different width fails the write loudly — no silently dropped vectors.

GET /_schema presents the field back as semantic_text with its binding, so the declaration round-trips.

Honest gap: no chunking yet — the whole field text embeds in one call (fine for titles, names and log lines; long documents are bounded by the model’s input limit). The engine keeps one vector per field, so ES-style chunking is on the roadmap; until then, chunk into child documents and collapse by parent id.

See Vector types, the semantic query and AI Services for the full contract.