a problem about hybrid search

[mengrennwpu]

Support guidelines

• [X] I've read the support guidelines

Background

I installed Elasticsearch using Docker.

elasticsearch version: 8.12.2 elasticsearch plugin versions: analysis-ik-8.12.2, elastiknn-8.12.2.1

The details of the created index are as follows:

{
    "settings": {
        "index": {
            "number_of_shards": 2
        }
    },
    "mappings": {
        "dynamic": false,
        "properties": {
            "content": {
                "type": "keyword"
            },
            "type": {
                "type": "keyword"
            },
            "value": {
                "type": "text",
                "analyzer": "ik_max_word",
                "search_analyzer": "ik_smart"
            },
            "embeddings": {
                "type": "elastiknn_dense_float_vector",
                "elastiknn": {
                    "dims": 768,
                    "model": "lsh",
                    "similarity": "cosine",
                    "L": 99,
                    "k": 3
                }
            }
        }
    }
}

the total number of documents is 1230. For example:

{
    "content": "牛角龙",
    "value": "牛角龙",
    "type": "entity",
    "embeddings": [
        0.049,
        -0.02,
        0.049,
        -0.02,
        0.049,
        -0.02,
        0.049,
        -0.02
    ]
}

Bug

1. When I search using only text, it works correctly.

• search body

  {
  "_source": {
      "excludes": [
          "embeddings"
      ]
  },
  "query": {
      "function_score": {
          "query": {
              "bool": {
                  "must": [
                      {
                          "term": {
                              "type": "entity"
                          }
                      },
                      {
                          "match": {
                              "value": "牛角龙有多长"
                          }
                      }
                  ]
              }
          }
      }
  },
  "size": 10
  }

• search result

  [
  {
      "_index": "align_gjzrbwggpgwxz15266552",
      "_id": "25da696bb0118917f43bc57bdf9c3a47",
      "_score": 6.274896,
      "_source": {
          "content": "牛角龙",
          "value": "牛角龙",
          "type": "entity"
      }
  },
  {
      "_index": "align_gjzrbwggpgwxz15266552",
      "_id": "d1955eeb40b7d590b0431586f0676f0f",
      "_score": 0.1788314,
      "_source": {
          "content": "龙猎龙",
          "value": "龙猎龙",
          "type": "entity"
      }
  }
  ]

2. When I search using only vector, it also works correctly.

• search body

  {
          "_source": {"excludes": ["embeddings"]},
          "query": {
              "elastiknn_nearest_neighbors": {
                             "field": "embeddings",
                             "vec": [0.04262753203511238, -0.06507280468940735, -0.07080274075269699, -0.022111210972070694, 0.03259402513504028],
                             "model": "lsh",
                             "similarity": "cosine",
                             "candidates": 100
                         }
          },
          "size": 10
      }

• search result

  [
  {
      "_index": "align_gjzrbwggpgwxz15266552",
      "_id": "25da696bb0118917f43bc57bdf9c3a47",
      "_score": 1.8340821,
      "_source": {
          "content": "牛角龙",
          "value": "牛角龙",
          "type": "entity"
      }
  },
  {
      "_index": "align_gjzrbwggpgwxz15266552",
      "_id": "7eaa446168ee245d5625a7ebeca9aae9",
      "_score": 1.7581223,
      "_source": {
          "content": "牛头龙",
          "value": "牛头龙",
          "type": "entity"
      }
  }
  ]

  3. When I use the hybrid search, the results are wrong

• search body

  {
  "_source": {
      "excludes": [
          "embeddings"
      ]
  },
  "query": {
      "function_score": {
          "query": {
              "bool": {
                  "must": [
                      {
                          "term": {
                              "type": "entity"
                          }
                      },
                      {
                          "match": {
                              "value": "牛角龙有多长"
                          }
                      }
                  ]
              }
          },
          "functions": [
              {
                  "elastiknn_nearest_neighbors": {
                      "field": "embeddings",
                      "vec": [
                          0.04262753203511238,
                          -0.06507280468940735,
                          -0.07080274075269699
                      ],
                      "model": "lsh",
                      "similarity": "cosine",
                      "candidates": 100
                  }
              }
          ]
      }
  },
  "size": 10
  }

• search result

  [
  {
      "_index": "align_gjzrbwggpgwxz15266552",
      "_id": "e8242ee3cb353a1580f7cfde11c6d88f",
      "_score": 0.08963046,
      "_source": {
          "content": "护甲萨尔塔龙",
          "value": "护甲萨尔塔龙",
          "type": "entity"
      }
  },
  {
      "_index": "align_gjzrbwggpgwxz15266552",
      "_id": "4a867698c11dbfb86e8a4de7bd018c42",
      "_score": 0.073508464,
      "_source": {
          "content": "气腔龙",
          "value": "气腔龙",
          "type": "entity"
      }
  }
  ]

So, where did I go wrong? I'm looking forward to and greatly appreciate your response.

Elastiknn Version

8.12.2.1

Platform

Docker

Steps to reproduce

No response

Additional info

No response

[mengrennwpu]

When I change the model 'lsh' to 'exact' during hybrid search , it works correctly. But as you know, performance will be affected.

[alexklibisz]

Hi @mengrennwpu, there are some known caveats with the function-score query: https://alexklibisz.github.io/elastiknn/api/#using-a-function-score-query

When using "model": "lsh", the "candidates" parameter is ignored and vectors are not re-scored with the exact similarity like they are with a elastiknn_nearest_neighbors query. Instead, the score is: max similarity score * proportion of matching hashes. This is a necessary consequence of the fact that score functions take a doc ID and must immediately return a score.

Perhaps that is happening here?

Also, I'm confused why your example is using size: 10 but only returning 2 results. Are you omitting some results? Or is this the actual complete response?

[mengrennwpu]

@alexklibisz Thank you very much for your response.

The response you referred to regarding the function-score query is correct. And yes, I omitted some results, i'm sorry for causing any misunderstanding.