How to Order Completion Suggester with Fuzziness-CodePudding

When using a Completion Suggester with Fuzziness defined the ordering of results for suggestions are alphabetical instead of most relevant. It seems that whatever the fuzzines is set to is removed from the search/query term at the end of the term. This is not what I expected from reading the Completion Suggester Fuzziness docs which state:

Suggestions that share the longest prefix to the query prefix will be scored higher.

But that is not true. Here is a use case that proves this:

PUT test/
{
  "mappings":{
    "properties":{
      "id":{
        "type":"integer"
      },
      "title":{
        "type":"keyword",
        "fields": {
          "suggest": {
            "type": "completion"
          }
        }
      }
    }
  }
}

POST test/_bulk
{ "index" : {"_id": "1"}}
{ "title": "HOLARAT" }
{ "index" : {"_id": "2"}}
{ "title": "HOLBROOK" }
{ "index" : {"_id": "3"}}
{ "title": "HOLCONNEN" }
{ "index" : {"_id": "4"}}
{ "title": "HOLDEN" }
{ "index" : {"_id": "5"}}
{ "title": "HOLLAND" }

The above creates an index and adds some data.

If a suggestion query is done on said data:

POST test/_search
{
  "_source": {
    "includes": [
      "title"
    ]
  },
  "suggest": {
    "title-suggestion": {
      "completion": {
        "fuzzy": {
          "fuzziness": "1"
        },
        "field": "title.suggest",
        "size": 3
      },
      "prefix": "HOLL"
    }
  }
}

It returns the first 3 results in alphabetical order of the last matching character, instead of the longest prefix (which would be HOLLAND):

{
  ...
  "suggest" : {
    "title-suggestion" : [
      {
        "text" : "HOLL",
        "offset" : 0,
        "length" : 4,
        "options" : [
          {
            "text" : "HOLARAT",
            "_index" : "test",
            "_type" : "_doc",
            "_id" : "1",
            "_score" : 3.0,
            "_source" : {
              "title" : "HOLARAT"
            }
          },
          {
            "text" : "HOLBROOK",
            "_index" : "test",
            "_type" : "_doc",
            "_id" : "2",
            "_score" : 3.0,
            "_source" : {
              "title" : "HOLBROOK"
            }
          },
          {
            "text" : "HOLCONNEN",
            "_index" : "test",
            "_type" : "_doc",
            "_id" : "3",
            "_score" : 3.0,
            "_source" : {
              "title" : "HOLCONNEN"
            }
          }
        ]
      }
    ]
  }
}

If the size param is removed then we can see that the score is the same for all entries, instead of the longest prefix being higher as stated.

With this being the case, how can results from Completion Suggesters with Fuzziness defined be ordered with the longest prefix at the top?

CodePudding user response：

This has been reported in the past and this behavior is actually by design.

What I usually do in this case is to send two suggest queries (similar to what has been suggested here), one for exact match and another for fuzzy match. If the exact match contains a suggestion, I use it, otherwise I resort to using the fuzzy ones.

With the suggest query below, you'll get HOLLAND as exact-suggestion and then the fuzzy matches in fuzzy-suggestion:

POST test/_search
{
  "_source": {
    "includes": [
      "title"
    ]
  },
  "suggest": {
    "fuzzy-suggestion": {
      "completion": {
        "fuzzy": {
          "fuzziness": "1"
        },
        "field": "title.suggest",
        "size": 3
      },
      "prefix": "HOLL"
    },
    "exact-suggestion": {
      "completion": {
        "field": "title.suggest",
        "size": 3
      },
      "prefix": "HOLL"
    }
  }
}