[FEATURE] Provide context on Inner hit of multi vector to aid highlighting/debug use cases

[heemin32]

As a follow up item on #1065, I would like to see inner hit of nested field so that I know which item inside nested fields are matched.

Create KNN index

PUT /my-knn-index-1
{
  "settings": {
    "index": {
      "knn": true,
      "knn.algo_param.ef_search": 100
    }
  },
  "mappings": {
    "properties": {
      "nested_field": {
        "type": "nested",
        "properties": {
          "my_vector": {
            "type": "knn_vector",
            "dimension": 3,
            "method": {
              "name": "hnsw",
              "space_type": "l2",
              "engine": "lucene",
              "parameters": {
                "ef_construction": 100,
                "m": 16
              }
            }
          }
        }
      }
    }
  }
}

Insert Data

PUT /_bulk?refresh=true
{ "index": { "_index": "my-knn-index-1", "_id": "1" } }
{"parking": false, "nested_field":[{"my_vector":[1,1,1]},{"my_vector":[2,2,2]},{"my_vector":[3,3,3]}]}
{ "index": { "_index": "my-knn-index-1", "_id": "2" } }
{"parking": true, "nested_field":[{"my_vector":[10,10,10]},{"my_vector":[20,20,20]},{"my_vector":[30,30,30]}]}
{ "index": { "_index": "my-knn-index-1", "_id": "3" } }
{"parking": true, "nested_field":[{"my_vector":[100,100,100]},{"my_vector":[200,200,200]},{"my_vector":[300,300,300]}]}

Query

GET /my-knn-index-1/_search
{
	"query": {
		"nested": {
			"path": "nested_field",
			"query": {
				"knn": {
					"nested_field.my_vector": {
						"vector": [
							1,
							1,
							1
						],
						"k": 2
					}
				}
			},
			"inner_hits": {}
		}
	}
}

Response

It returns two nearest values among entire nested values, but not best one for each document.

{
	"took": 9,
	"timed_out": false,
	"_shards": {
		"total": 1,
		"successful": 1,
		"skipped": 0,
		"failed": 0
	},
	"hits": {
		"total": {
			"value": 2,
			"relation": "eq"
		},
		"max_score": 1.0,
		"hits": [
			{
				"_index": "my-knn-index-1",
				"_id": "1",
				"_score": 1.0,
				"_source": {
					"parking": false,
					"nested_field": [
						{
							"my_vector": [
								1,
								1,
								1
							]
						},
						{
							"my_vector": [
								2,
								2,
								2
							]
						},
						{
							"my_vector": [
								3,
								3,
								3
							]
						}
					]
				},
				"inner_hits": {
					"nested_field": {
						"hits": {
							"total": {
								"value": 2,
								"relation": "eq"
							},
							"max_score": 1.0,
							"hits": [
								{
									"_index": "my-knn-index-1",
									"_id": "1",
									"_nested": {
										"field": "nested_field",
										"offset": 0
									},
									"_score": 1.0,
									"_source": {
										"my_vector": [
											1,
											1,
											1
										]
									}
								},
								{
									"_index": "my-knn-index-1",
									"_id": "1",
									"_nested": {
										"field": "nested_field",
										"offset": 1
									},
									"_score": 0.25,
									"_source": {
										"my_vector": [
											2,
											2,
											2
										]
									}
								}
							]
						}
					}
				}
			},
			{
				"_index": "my-knn-index-1",
				"_id": "3",
				"_score": 3.400898E-5,
				"_source": {
					"parking": true,
					"nested_field": [
						{
							"my_vector": [
								100,
								100,
								100
							]
						},
						{
							"my_vector": [
								200,
								200,
								200
							]
						},
						{
							"my_vector": [
								300,
								300,
								300
							]
						}
					]
				},
				"inner_hits": {
					"nested_field": {
						"hits": {
							"total": {
								"value": 0,
								"relation": "eq"
							},
							"max_score": null,
							"hits": []
						}
					}
				}
			}
		]
	}
}

[navneet1v]

@heemin32 were you able to figure out why the query with filter is not returning inner hits?

Plus is this consistent with Faiss engine too?

[heemin32]

Same for both Faiss and Lucene. Wasn't able to find the why.

[navneet1v]

@heemin32 its pretty interesting that inner hits is working for standard search and not for filtered vector search even when we run almost same code in both the places.

[abdulacs]

@navneet1v We faced the same issue for our use case and due to this we were unable to use on a query with filter

[vamshin]

@heemin32 looks like we are addressing two different issues on this thread. 1) ability to know which inner hit contributed the score for the document and 2)Filter seem to be broken with inner hits. Should we create separate issues for these?

[heemin32]

@vamshin. That is correct. However, in terms of feature completeness, wouldn't it be better to resolve them both together?

[vamshin]

@heemin32 agree both need to be resolved. Just that it might be hard to discover for folks seeing issues with filter and inner hits. Not a major concern.

[heemin32]

@vamshin After some deep dive, it seems they are two completely separate issue in terms of implementation. Let me create a new issue for innerHit with filter and address them one by one.

[heemin32]

Closing in favor of opensearch-project/OpenSearch#13903