Include metadata with dense vector chunks in Elasticsearch

nested-metadata.json

PUT audio
{
  "mappings": {
    "properties": {
      "my_long_content": {
        "type": "nested", //because there can be multiple vectors per doc
        "properties": {
          "vector": {
            "type": "dense_vector" //the vector used for ranking
          },
          "text": {
            "type": "text" //the extracted text from which the vector was created
          },
          "time_ms": {
            "type": "long" //offset from the start in ms of the chunk
          }
        }
      }
    }
  }
}

PUT audio/_doc/1
{
  "my_long_content" : [
    {
      "vector" : [23,14,8],
      "text" :  "doc 1 chunk 1",
      "time_ms": 0
    },
    {
      "vector" : [34,95,17],
      "text" :  "doc 1 chunk 2",
      "time_ms": 34
    }
  ]
}

PUT audio/_doc/2
{
  "my_long_content" : [
    {
      "vector" : [3,2,890],
      "text" :  "doc 2 chunk 1",
      "time_ms": 0
    },
    {
      "vector" : [129,765,13],
      "text" :  "doc 2 chunk 2",
      "time_ms": 90
    }
  ]
}


GET audio/_search
{
  "_source": false,
  "size": 1,
  "knn": {
    "field": "my_long_content.vector",
    "query_vector": [
      35,
      90,
      20
    ],
    "inner_hits": {
      "_source": true,
      "size": 1
    }
  }
}