Vector field getting "appended" instead of replacing

[mattkallo]

Describe the bug
Updating a document with knn_vector setup through ingest pipline re-computes the vector and tries to "append" (doubles the dimension) to the vector field. I am not sure if this is bug or expected behavior.

Steps to reproduce the behavior:

1. Create an index with knn_vector field (eg. content_vector)
2. Add a ingest pipeline that uses a model to generate the embedding from field "content" and update the "content_vector" with the value.
3. Index few documents
4. Try to update one documents (any field of that document)
5. It throws the below error

      "cause": {
        "type": "mapper_parsing_exception",
        "reason": "failed to parse field [content_vector] of type [knn_vector] in document with id 'eb10dee8-0e07-47fa-a338-26865cecf585'. Preview of field's value: 'null'",
        "caused_by": {
          "type": "illegal_argument_exception",
          "reason": "Vector dimension mismatch. Expected: 768, Given: 1536"
        }
      },

Expected behavior
It should update the source and return

Plugins
Default plugins only. No additional ones installed.

Host/Environment (please complete the following information):

• OS: Linux
• Version 2.5.0

Additional context
Pipeline

PUT _ingest/pipeline/nlp-pipeline
{
  "description": "Pipeline",
  "processors" : [
    {
      "text_embedding": {
        "model_id": "gJCndYgBjZUvSFOP3Cwb",
        "field_map": {
          "content": "content_vector"
        },
      "ignore_failure": true
      }
    }
  ]
}

Update document

POST /content-index/_update_by_query
{
  "script": {
    "source": "ctx._source.title='Autonomous planning'",
    "lang": "painless"
  },
  "query": {
    "match": {
      "_id": "eb10dee8-0e07-47fa-a338-26865cecf585"
    }
  }
}

[minalsha]

@vamshin would you help triage this issue? Thanks

[vamshin]

This seem to be neural search plugin ingest issue.

[navneet1v]

@mattkallo the code fix has been deployed in 2.9 branch. Once we have the release for 2.9 this bug will be fixed.

[navneet1v]

Resolving the 2.9 is released.