It is also possible to expand each collapsed top hits with the inner_hits option.

GET /twitter/_search
{
    "query": {
        "match": {
            "message": "elasticsearch"
        }
    },
    "collapse" : {
        "field" : "user", 
        "inner_hits": {
            "name": "last_tweets", 
            "size": 5, 
            "sort": [{ "date": "asc" }] 
        },
        "max_concurrent_group_searches": 4 
    },
    "sort": ["likes"]
}

See inner hits for the complete list of supported options and the format of the response.

It is also possible to request multiple inner_hits for each collapsed hit. This can be useful when you want to get multiple representations of the collapsed hits.

GET /twitter/_search
{
    "query": {
        "match": {
            "message": "elasticsearch"
        }
    },
    "collapse" : {
        "field" : "user", 
        "inner_hits": [
            {
                "name": "most_liked",  
                "size": 3,
                "sort": ["likes"]
            },
            {
                "name": "most_recent", 
                "size": 3,
                "sort": [{ "date": "asc" }]
            }
        ]
    },
    "sort": ["likes"]
}

The expansion of the group is done by sending an additional query for each inner_hit request for each collapsed hit returned in the response. This can significantly slow things down if you have too many groups and/or inner_hit requests.

The max_concurrent_group_searches request parameter can be used to control the maximum number of concurrent searches allowed in this phase. The default is based on the number of data nodes and the default search thread pool size.

Second level of collapsing is also supported and is applied to inner_hits. For example, the following request finds the top scored tweets for each country, and within each country finds the top scored tweets for each user.

GET /twitter/_search
{
    "query": {
        "match": {
            "message": "elasticsearch"
        }
    },
    "collapse" : {
        "field" : "country",
        "inner_hits" : {
            "name": "by_location",
            "collapse" : {"field" : "user"},
            "size": 3
        }
    }
}

Response:

{
    ...
    "hits": [
        {
            "_index": "twitter",
            "_type": "_doc",
            "_id": "9",
            "_score": ...,
            "_source": {...},
            "fields": {"country": ["UK"]},
            "inner_hits":{
                "by_location": {
                    "hits": {
                       ...,
                       "hits": [
                          {
                            ...
                            "fields": {"user" : ["user124"]}
                          },
                          {
                            ...
                            "fields": {"user" : ["user589"]}
                          },
                          {
                            ...
                             "fields": {"user" : ["user001"]}
                          }
                       ]
                    }
                 }
            }
        },
        {
            "_index": "twitter",
            "_type": "_doc",
            "_id": "1",
            "_score": ..,
            "_source": {...},
            "fields": {"country": ["Canada"]},
            "inner_hits":{
                "by_location": {
                    "hits": {
                       ...,
                       "hits": [
                          {
                            ...
                            "fields": {"user" : ["user444"]}
                          },
                          {
                            ...
                            "fields": {"user" : ["user1111"]}
                          },
                          {
                            ...
                             "fields": {"user" : ["user999"]}
                          }
                       ]
                    }
                 }
            }

        },
        ....
    ]
}

The unified highlighter uses the Lucene Unified Highlighter. This highlighter breaks the text into sentences and uses the BM25 algorithm to score individual sentences as if they were documents in the corpus. It also supports accurate phrase and multi-term (fuzzy, prefix, regex) highlighting. This is the default highlighter.

The plain highlighter uses the standard Lucene highlighter. It attempts to reflect the query matching logic in terms of understanding word importance and any word positioning criteria in phrase queries.

The fvh highlighter uses the Lucene Fast Vector highlighter. This highlighter can be used on fields with term_vector set to with_positions_offsets in the mapping. The fast vector highlighter:

To create meaningful search snippets from the terms being queried, the highlighter needs to know the start and end character offsets of each word in the original text. These offsets can be obtained from:

Highlighting settings can be set on a global level and overridden at the field level.

Override global settingsedit

You can specify highlighter settings globally and selectively override them for individual fields.

GET /_search
{
    "query" : {
        "match": { "user": "kimchy" }
    },
    "highlight" : {
        "number_of_fragments" : 3,
        "fragment_size" : 150,
        "fields" : {
            "body" : { "pre_tags" : ["<em>"], "post_tags" : ["</em>"] },
            "blog.title" : { "number_of_fragments" : 0 },
            "blog.author" : { "number_of_fragments" : 0 },
            "blog.comment" : { "number_of_fragments" : 5, "order" : "score" }
        }
    }
}

Specify a highlight queryedit

You can specify a highlight_query to take additional information into account when highlighting. For example, the following query includes both the search query and rescore query in the highlight_query. Without the highlight_query, highlighting would only take the search query into account.

GET /_search
{
    "query" : {
        "match": {
            "comment": {
                "query": "foo bar"
            }
        }
    },
    "rescore": {
        "window_size": 50,
        "query": {
            "rescore_query" : {
                "match_phrase": {
                    "comment": {
                        "query": "foo bar",
                        "slop": 1
                    }
                }
            },
            "rescore_query_weight" : 10
        }
    },
    "_source": false,
    "highlight" : {
        "order" : "score",
        "fields" : {
            "comment" : {
                "fragment_size" : 150,
                "number_of_fragments" : 3,
                "highlight_query": {
                    "bool": {
                        "must": {
                            "match": {
                                "comment": {
                                    "query": "foo bar"
                                }
                            }
                        },
                        "should": {
                            "match_phrase": {
                                "comment": {
                                    "query": "foo bar",
                                    "slop": 1,
                                    "boost": 10.0
                                }
                            }
                        },
                        "minimum_should_match": 0
                    }
                }
            }
        }
    }
}

Set highlighter typeedit

The type field allows to force a specific highlighter type. The allowed values are: unified, plain and fvh. The following is an example that forces the use of the plain highlighter:

GET /_search
{
    "query" : {
        "match": { "user": "kimchy" }
    },
    "highlight" : {
        "fields" : {
            "comment" : {"type" : "plain"}
        }
    }
}

Configure highlighting tagsedit

By default, the highlighting will wrap highlighted text in <em> and </em>. This can be controlled by setting pre_tags and post_tags, for example:

GET /_search
{
    "query" : {
        "match": { "user": "kimchy" }
    },
    "highlight" : {
        "pre_tags" : ["<tag1>"],
        "post_tags" : ["</tag1>"],
        "fields" : {
            "body" : {}
        }
    }
}

When using the fast vector highlighter, you can specify additional tags and the "importance" is ordered.

GET /_search
{
    "query" : {
        "match": { "user": "kimchy" }
    },
    "highlight" : {
        "pre_tags" : ["<tag1>", "<tag2>"],
        "post_tags" : ["</tag1>", "</tag2>"],
        "fields" : {
            "body" : {}
        }
    }
}

You can also use the built-in styled tag schema:

GET /_search
{
    "query" : {
        "match": { "user": "kimchy" }
    },
    "highlight" : {
        "tags_schema" : "styled",
        "fields" : {
            "comment" : {}
        }
    }
}

Highlight on sourceedit

Forces the highlighting to highlight fields based on the source even if fields are stored separately. Defaults to false.

GET /_search
{
    "query" : {
        "match": { "user": "kimchy" }
    },
    "highlight" : {
        "fields" : {
            "comment" : {"force_source" : true}
        }
    }
}

Highlight in all fieldsedit

By default, only fields that contains a query match are highlighted. Set require_field_match to false to highlight all fields.

GET /_search
{
    "query" : {
        "match": { "user": "kimchy" }
    },
    "highlight" : {
        "require_field_match": false,
        "fields": {
                "body" : { "pre_tags" : ["<em>"], "post_tags" : ["</em>"] }
        }
    }
}

Combine matches on multiple fieldsedit

The Fast Vector Highlighter can combine matches on multiple fields to highlight a single field. This is most intuitive for multifields that analyze the same string in different ways. All matched_fields must have term_vector set to with_positions_offsets but only the field to which the matches are combined is loaded so only that field would benefit from having store set to yes.

In the following examples, comment is analyzed by the english analyzer and comment.plain is analyzed by the standard analyzer.

GET /_search
{
    "query": {
        "query_string": {
            "query": "comment.plain:running scissors",
            "fields": ["comment"]
        }
    },
    "highlight": {
        "order": "score",
        "fields": {
            "comment": {
                "matched_fields": ["comment", "comment.plain"],
                "type" : "fvh"
            }
        }
    }
}

The above matches both "run with scissors" and "running with scissors" and would highlight "running" and "scissors" but not "run". If both phrases appear in a large document then "running with scissors" is sorted above "run with scissors" in the fragments list because there are more matches in that fragment.

GET /_search
{
    "query": {
        "query_string": {
            "query": "running scissors",
            "fields": ["comment", "comment.plain^10"]
        }
    },
    "highlight": {
        "order": "score",
        "fields": {
            "comment": {
                "matched_fields": ["comment", "comment.plain"],
                "type" : "fvh"
            }
        }
    }
}

The above highlights "run" as well as "running" and "scissors" but still sorts "running with scissors" above "run with scissors" because the plain match ("running") is boosted.

GET /_search
{
    "query": {
        "query_string": {
            "query": "running scissors",
            "fields": ["comment", "comment.plain^10"]
        }
    },
    "highlight": {
        "order": "score",
        "fields": {
            "comment": {
                "matched_fields": ["comment.plain"],
                "type" : "fvh"
            }
        }
    }
}

The above query wouldn’t highlight "run" or "scissor" but shows that it is just fine not to list the field to which the matches are combined (comment) in the matched fields.

    "highlight": {
        "fields": {
            "comment": {}
        }
    }

    "highlight": {
        "fields": {
            "comment": {
                "matched_fields": ["comment"],
                "type" : "fvh"
            }
        }
    }

Explicitly order highlighted fieldsedit

Elasticsearch highlights the fields in the order that they are sent, but per the JSON spec, objects are unordered. If you need to be explicit about the order in which fields are highlighted specify the fields as an array:

GET /_search
{
    "highlight": {
        "fields": [
            { "title": {} },
            { "text": {} }
        ]
    }
}

None of the highlighters built into Elasticsearch care about the order that the fields are highlighted but a plugin might.

Control highlighted fragmentsedit

Each field highlighted can control the size of the highlighted fragment in characters (defaults to 100), and the maximum number of fragments to return (defaults to 5). For example:

GET /_search
{
    "query" : {
        "match": { "user": "kimchy" }
    },
    "highlight" : {
        "fields" : {
            "comment" : {"fragment_size" : 150, "number_of_fragments" : 3}
        }
    }
}

On top of this it is possible to specify that highlighted fragments need to be sorted by score:

GET /_search
{
    "query" : {
        "match": { "user": "kimchy" }
    },
    "highlight" : {
        "order" : "score",
        "fields" : {
            "comment" : {"fragment_size" : 150, "number_of_fragments" : 3}
        }
    }
}

If the number_of_fragments value is set to 0 then no fragments are produced, instead the whole content of the field is returned, and of course it is highlighted. This can be very handy if short texts (like document title or address) need to be highlighted but no fragmentation is required. Note that fragment_size is ignored in this case.

GET /_search
{
    "query" : {
        "match": { "user": "kimchy" }
    },
    "highlight" : {
        "fields" : {
            "body" : {},
            "blog.title" : {"number_of_fragments" : 0}
        }
    }
}

When using fvh one can use fragment_offset parameter to control the margin to start highlighting from.

In the case where there is no matching fragment to highlight, the default is to not return anything. Instead, we can return a snippet of text from the beginning of the field by setting no_match_size (default 0) to the length of the text that you want returned. The actual length may be shorter or longer than specified as it tries to break on a word boundary.

GET /_search
{
    "query" : {
        "match": { "user": "kimchy" }
    },
    "highlight" : {
        "fields" : {
            "comment" : {
                "fragment_size" : 150,
                "number_of_fragments" : 3,
                "no_match_size": 150
            }
        }
    }
}

Highlight using the postings listedit

Here is an example of setting the comment field in the index mapping to allow for highlighting using the postings:

PUT /example
{
  "mappings": {
    "properties": {
      "comment" : {
        "type": "text",
        "index_options" : "offsets"
      }
    }
  }
}

Here is an example of setting the comment field to allow for highlighting using the term_vectors (this will cause the index to be bigger):

PUT /example
{
  "mappings": {
    "properties": {
      "comment" : {
        "type": "text",
        "term_vector" : "with_positions_offsets"
      }
    }
  }
}

Specify a fragmenter for the plain highlighteredit

When using the plain highlighter, you can choose between the simple and span fragmenters:

GET twitter/_search
{
    "query" : {
        "match_phrase": { "message": "number 1" }
    },
    "highlight" : {
        "fields" : {
            "message" : {
                "type": "plain",
                "fragment_size" : 15,
                "number_of_fragments" : 3,
                "fragmenter": "simple"
            }
        }
    }
}

Response:

{
    ...
    "hits": {
        "total" : {
            "value": 1,
            "relation": "eq"
        },
        "max_score": 1.6011951,
        "hits": [
            {
                "_index": "twitter",
                "_type": "_doc",
                "_id": "1",
                "_score": 1.6011951,
                "_source": {
                    "user": "test",
                    "message": "some message with the number 1",
                    "date": "2009-11-15T14:12:12",
                    "likes": 1
                },
                "highlight": {
                    "message": [
                        " with the <em>number</em>",
                        " <em>1</em>"
                    ]
                }
            }
        ]
    }
}

GET twitter/_search
{
    "query" : {
        "match_phrase": { "message": "number 1" }
    },
    "highlight" : {
        "fields" : {
            "message" : {
                "type": "plain",
                "fragment_size" : 15,
                "number_of_fragments" : 3,
                "fragmenter": "span"
            }
        }
    }
}

Response:

{
    ...
    "hits": {
        "total" : {
            "value": 1,
            "relation": "eq"
        },
        "max_score": 1.6011951,
        "hits": [
            {
                "_index": "twitter",
                "_type": "_doc",
                "_id": "1",
                "_score": 1.6011951,
                "_source": {
                    "user": "test",
                    "message": "some message with the number 1",
                    "date": "2009-11-15T14:12:12",
                    "likes": 1
                },
                "highlight": {
                    "message": [
                        " with the <em>number</em> <em>1</em>"
                    ]
                }
            }
        ]
    }
}

If the number_of_fragments option is set to 0, NullFragmenter is used which does not fragment the text at all. This is useful for highlighting the entire contents of a document or field.

Relevant settings: fragment_size, fragmenter, type of highlighter, boundary_chars, boundary_max_scan, boundary_scanner, boundary_scanner_locale.

Plain highlighter begins with analyzing the text using the given analyzer, and creating a token stream from it. Plain highlighter uses a very simple algorithm to break the token stream into fragments. It loops through terms in the token stream, and every time the current term’s end_offset exceeds fragment_size multiplied by the number of created fragments, a new fragment is created. A little more computation is done with using span fragmenter to avoid breaking up text between highlighted terms. But overall, since the breaking is done only by fragment_size, some fragments can be quite odd, e.g. beginning with a punctuation mark.

Unified or FVH highlighters do a better job of breaking up a text into fragments by utilizing Java’s BreakIterator. This ensures that a fragment is a valid sentence as long as fragment_size allows for this.

Relevant settings: number_of_fragments.

To find the best, most relevant, fragments, a highlighter needs to score each fragment in respect to the given query. The goal is to score only those terms that participated in generating the hit on the document. For some complex queries, this is still work in progress.

The plain highlighter creates an in-memory index from the current token stream, and re-runs the original query criteria through Lucene’s query execution planner to get access to low-level match information for the current text. For more complex queries the original query could be converted to a span query, as span queries can handle phrases more accurately. Then this obtained low-level match information is used to score each individual fragment. The scoring method of the plain highlighter is quite simple. Each fragment is scored by the number of unique query terms found in this fragment. The score of individual term is equal to its boost, which is by default is 1. Thus, by default, a fragment that contains one unique query term, will get a score of 1; and a fragment that contains two unique query terms, will get a score of 2 and so on. The fragments are then sorted by their scores, so the highest scored fragments will be output first.

FVH doesn’t need to analyze the text and build an in-memory index, as it uses pre-indexed document term vectors, and finds among them terms that correspond to the query. FVH scores each fragment by the number of query terms found in this fragment. Similarly to plain highlighter, score of individual term is equal to its boost value. In contrast to plain highlighter, all query terms are counted, not only unique terms.

Unified highlighter can use pre-indexed term vectors or pre-indexed terms offsets, if they are available. Otherwise, similar to Plain Highlighter, it has to create an in-memory index from the text. Unified highlighter uses the BM25 scoring model to score fragments.

Relevant settings: pre-tags, post-tags.

The goal is to highlight only those terms that participated in generating the hit on the document. For some complex boolean queries, this is still work in progress, as highlighters don’t reflect the boolean logic of a query and only extract leaf (terms, phrases, prefix etc) queries.

Plain highlighter given the token stream and the original text, recomposes the original text to highlight only terms from the token stream that are contained in the low-level match information structure from the previous step.

FVH and unified highlighter use intermediate data structures to represent fragments in some raw form, and then populate them with actual text.

A highlighter uses pre-tags, post-tags to encode highlighted terms.

Let’s look in more details how unified highlighter works.

First, we create a index with a text field content, that will be indexed using english analyzer, and will be indexed without offsets or term vectors.

PUT test_index
{
    "mappings": {
        "properties": {
            "content" : {
                "type" : "text",
                "analyzer" : "english"
            }
        }
    }
}

We put the following document into the index:

PUT test_index/_doc/doc1
{
  "content" : "For you I'm only a fox like a hundred thousand other foxes. But if you tame me, we'll need each other. You'll be the only boy in the world for me. I'll be the only fox in the world for you."
}

And we ran the following query with a highlight request:

GET test_index/_search
{
    "query": {
        "match_phrase" : {"content" : "only fox"}
    },
    "highlight": {
        "type" : "unified",
        "number_of_fragments" : 3,
        "fields": {
            "content": {}
        }
    }
}

After doc1 is found as a hit for this query, this hit will be passed to the unified highlighter for highlighting the field content of the document. Since the field content was not indexed either with offsets or term vectors, its raw field value will be analyzed, and in-memory index will be built from the terms that match the query:

{"token":"onli","start_offset":12,"end_offset":16,"position":3},
{"token":"fox","start_offset":19,"end_offset":22,"position":5},
{"token":"fox","start_offset":53,"end_offset":58,"position":11},
{"token":"onli","start_offset":117,"end_offset":121,"position":24},
{"token":"onli","start_offset":159,"end_offset":163,"position":34},
{"token":"fox","start_offset":164,"end_offset":167,"position":35}

Our complex phrase query will be converted to the span query: spanNear([text:onli, text:fox], 0, true), meaning that we are looking for terms "onli: and "fox" within 0 distance from each other, and in the given order. The span query will be run against the created before in-memory index, to find the following match:

{"term":"onli", "start_offset":159, "end_offset":163},
{"term":"fox", "start_offset":164, "end_offset":167}

In our example, we have got a single match, but there could be several matches. Given the matches, the unified highlighter breaks the text of the field into so called "passages". Each passage must contain at least one match. The unified highlighter with the use of Java’s BreakIterator ensures that each passage represents a full sentence as long as it doesn’t exceed fragment_size. For our example, we have got a single passage with the following properties (showing only a subset of the properties here):

Passage:
    startOffset: 147
    endOffset: 189
    score: 3.7158387
    matchStarts: [159, 164]
    matchEnds: [163, 167]
    numMatches: 2

Notice how a passage has a score, calculated using the BM25 scoring formula adapted for passages. Scores allow us to choose the best scoring passages if there are more passages available than the requested by the user number_of_fragments. Scores also let us to sort passages by order: "score" if requested by the user.

As the final step, the unified highlighter will extract from the field’s text a string corresponding to each passage:

"I'll be the only fox in the world for you."

and will format with the tags <em> and </em> all matches in this string using the passages’s matchStarts and matchEnds information:

I'll be the <em>only</em> <em>fox</em> in the world for you.

This kind of formatted strings are the final result of the highlighter returned to the user.

Inner hits support the following options:

Inner hits also supports the following per document features:

The nested inner_hits can be used to include nested inner objects as inner hits to a search hit.

PUT test
{
  "mappings": {
    "properties": {
      "comments": {
        "type": "nested"
      }
    }
  }
}

PUT test/_doc/1?refresh
{
  "title": "Test title",
  "comments": [
    {
      "author": "kimchy",
      "number": 1
    },
    {
      "author": "nik9000",
      "number": 2
    }
  ]
}

POST test/_search
{
  "query": {
    "nested": {
      "path": "comments",
      "query": {
        "match": {"comments.number" : 2}
      },
      "inner_hits": {} 
    }
  }
}

An example of a response snippet that could be generated from the above search request:

{
  ...,
  "hits": {
    "total" : {
        "value": 1,
        "relation": "eq"
    },
    "max_score": 1.0,
    "hits": [
      {
        "_index": "test",
        "_type": "_doc",
        "_id": "1",
        "_score": 1.0,
        "_source": ...,
        "inner_hits": {
          "comments": { 
            "hits": {
              "total" : {
                  "value": 1,
                  "relation": "eq"
              },
              "max_score": 1.0,
              "hits": [
                {
                  "_index": "test",
                  "_type": "_doc",
                  "_id": "1",
                  "_nested": {
                    "field": "comments",
                    "offset": 1
                  },
                  "_score": 1.0,
                  "_source": {
                    "author": "nik9000",
                    "number": 2
                  }
                }
              ]
            }
          }
        }
      }
    ]
  }
}

The _nested metadata is crucial in the above example, because it defines from what inner nested object this inner hit came from. The field defines the object array field the nested hit is from and the offset relative to its location in the _source. Due to sorting and scoring the actual location of the hit objects in the inner_hits is usually different than the location a nested inner object was defined.

By default the _source is returned also for the hit objects in inner_hits, but this can be changed. Either via _source filtering feature part of the source can be returned or be disabled. If stored fields are defined on the nested level these can also be returned via the fields feature.

An important default is that the _source returned in hits inside inner_hits is relative to the _nested metadata. So in the above example only the comment part is returned per nested hit and not the entire source of the top level document that contained the comment.

Nested document don’t have a _source field, because the entire source of document is stored with the root document under its _source field. To include the source of just the nested document, the source of the root document is parsed and just the relevant bit for the nested document is included as source in the inner hit. Doing this for each matching nested document has an impact on the time it takes to execute the entire search request, especially when size and the inner hits' size are set higher than the default. To avoid the relatively expensive source extraction for nested inner hits, one can disable including the source and solely rely on doc values fields. Like this:

PUT test
{
  "mappings": {
    "properties": {
      "comments": {
        "type": "nested"
      }
    }
  }
}

PUT test/_doc/1?refresh
{
  "title": "Test title",
  "comments": [
    {
      "author": "kimchy",
      "text": "comment text"
    },
    {
      "author": "nik9000",
      "text": "words words words"
    }
  ]
}

POST test/_search
{
  "query": {
    "nested": {
      "path": "comments",
      "query": {
        "match": {"comments.text" : "words"}
      },
      "inner_hits": {
        "_source" : false,
        "docvalue_fields" : [
          "comments.text.keyword"
        ]
      }
    }
  }
}

If a mapping has multiple levels of hierarchical nested object fields each level can be accessed via dot notated path. For example if there is a comments nested field that contains a votes nested field and votes should directly be returned with the root hits then the following path can be defined:

PUT test
{
  "mappings": {
    "properties": {
      "comments": {
        "type": "nested",
        "properties": {
          "votes": {
            "type": "nested"
          }
        }
      }
    }
  }
}

PUT test/_doc/1?refresh
{
  "title": "Test title",
  "comments": [
    {
      "author": "kimchy",
      "text": "comment text",
      "votes": []
    },
    {
      "author": "nik9000",
      "text": "words words words",
      "votes": [
        {"value": 1 , "voter": "kimchy"},
        {"value": -1, "voter": "other"}
      ]
    }
  ]
}

POST test/_search
{
  "query": {
    "nested": {
      "path": "comments.votes",
        "query": {
          "match": {
            "comments.votes.voter": "kimchy"
          }
        },
        "inner_hits" : {}
    }
  }
}

Which would look like:

{
  ...,
  "hits": {
    "total" : {
        "value": 1,
        "relation": "eq"
    },
    "max_score": 0.6931471,
    "hits": [
      {
        "_index": "test",
        "_type": "_doc",
        "_id": "1",
        "_score": 0.6931471,
        "_source": ...,
        "inner_hits": {
          "comments.votes": { 
            "hits": {
              "total" : {
                  "value": 1,
                  "relation": "eq"
              },
              "max_score": 0.6931471,
              "hits": [
                {
                  "_index": "test",
                  "_type": "_doc",
                  "_id": "1",
                  "_nested": {
                    "field": "comments",
                    "offset": 1,
                    "_nested": {
                      "field": "votes",
                      "offset": 0
                    }
                  },
                  "_score": 0.6931471,
                  "_source": {
                    "value": 1,
                    "voter": "kimchy"
                  }
                }
              ]
            }
          }
        }
      }
    ]
  }
}

This indirect referencing is only supported for nested inner hits.

The parent/child inner_hits can be used to include parent or child:

PUT test
{
  "mappings": {
    "properties": {
      "my_join_field": {
        "type": "join",
        "relations": {
          "my_parent": "my_child"
        }
      }
    }
  }
}

PUT test/_doc/1?refresh
{
  "number": 1,
  "my_join_field": "my_parent"
}

PUT test/_doc/2?routing=1&refresh
{
  "number": 1,
  "my_join_field": {
    "name": "my_child",
    "parent": "1"
  }
}

POST test/_search
{
  "query": {
    "has_child": {
      "type": "my_child",
      "query": {
        "match": {
          "number": 1
        }
      },
      "inner_hits": {}    
    }
  }
}

An example of a response snippet that could be generated from the above search request:

{
    ...,
    "hits": {
        "total" : {
            "value": 1,
            "relation": "eq"
        },
        "max_score": 1.0,
        "hits": [
            {
                "_index": "test",
                "_type": "_doc",
                "_id": "1",
                "_score": 1.0,
                "_source": {
                    "number": 1,
                    "my_join_field": "my_parent"
                },
                "inner_hits": {
                    "my_child": {
                        "hits": {
                            "total" : {
                                "value": 1,
                                "relation": "eq"
                            },
                            "max_score": 1.0,
                            "hits": [
                                {
                                    "_index": "test",
                                    "_type": "_doc",
                                    "_id": "2",
                                    "_score": 1.0,
                                    "_routing": "1",
                                    "_source": {
                                        "number": 1,
                                        "my_join_field": {
                                            "name": "my_child",
                                            "parent": "1"
                                        }
                                    }
                                }
                            ]
                        }
                    }
                }
            }
        ]
    }
}

The query rescorer executes a second query only on the Top-K results returned by the query and post_filter phases. The number of docs which will be examined on each shard can be controlled by the window_size parameter, which defaults to 10.

By default the scores from the original query and the rescore query are combined linearly to produce the final _score for each document. The relative importance of the original query and of the rescore query can be controlled with the query_weight and rescore_query_weight respectively. Both default to 1.

For example:

POST /_search
{
   "query" : {
      "match" : {
         "message" : {
            "operator" : "or",
            "query" : "the quick brown"
         }
      }
   },
   "rescore" : {
      "window_size" : 50,
      "query" : {
         "rescore_query" : {
            "match_phrase" : {
               "message" : {
                  "query" : "the quick brown",
                  "slop" : 2
               }
            }
         },
         "query_weight" : 0.7,
         "rescore_query_weight" : 1.2
      }
   }
}

The way the scores are combined can be controlled with the score_mode:

It is also possible to execute multiple rescores in sequence:

POST /_search
{
   "query" : {
      "match" : {
         "message" : {
            "operator" : "or",
            "query" : "the quick brown"
         }
      }
   },
   "rescore" : [ {
      "window_size" : 100,
      "query" : {
         "rescore_query" : {
            "match_phrase" : {
               "message" : {
                  "query" : "the quick brown",
                  "slop" : 2
               }
            }
         },
         "query_weight" : 0.7,
         "rescore_query_weight" : 1.2
      }
   }, {
      "window_size" : 10,
      "query" : {
         "score_mode": "multiply",
         "rescore_query" : {
            "function_score" : {
               "script_score": {
                  "script": {
                    "source": "Math.log10(doc.likes.value + 2)"
                  }
               }
            }
         }
      }
   } ]
}

The first one gets the results of the query then the second one gets the results of the first, etc. The second rescore will "see" the sorting done by the first rescore so it is possible to use a large window on the first rescore to pull documents into a smaller window for the second rescore.

A scroll returns all the documents which matched the search at the time of the initial search request. It ignores any subsequent changes to these documents. The scroll_id identifies a search context which keeps track of everything that Elasticsearch needs to return the correct documents. The search context is created by the initial request and kept alive by subsequent requests.

The scroll parameter (passed to the search request and to every scroll request) tells Elasticsearch how long it should keep the search context alive. Its value (e.g. 1m, see Time units) does not need to be long enough to process all data — it just needs to be long enough to process the previous batch of results. Each scroll request (with the scroll parameter) sets a new expiry time. If a scroll request doesn’t pass in the scroll parameter, then the search context will be freed as part of that scroll request.

Normally, the background merge process optimizes the index by merging together smaller segments to create new, bigger segments. Once the smaller segments are no longer needed they are deleted. This process continues during scrolling, but an open search context prevents the old segments from being deleted since they are still in use.

Additionally, if a segment contains deleted or updated documents then the search context must keep track of whether each document in the segment was live at the time of the initial search request. Ensure that your nodes have sufficient heap space if you have many open scrolls on an index that is subject to ongoing deletes or updates.

You can check how many search contexts are open with the nodes stats API:

GET /_nodes/stats/indices/search

Search context are automatically removed when the scroll timeout has been exceeded. However keeping scrolls open has a cost, as discussed in the previous section so scrolls should be explicitly cleared as soon as the scroll is not being used anymore using the clear-scroll API:

DELETE /_search/scroll
{
    "scroll_id" : "DXF1ZXJ5QW5kRmV0Y2gBAAAAAAAAAD4WYm9laVYtZndUQlNsdDcwakFMNjU1QQ=="
}

Multiple scroll IDs can be passed as array:

DELETE /_search/scroll
{
    "scroll_id" : [
      "DXF1ZXJ5QW5kRmV0Y2gBAAAAAAAAAD4WYm9laVYtZndUQlNsdDcwakFMNjU1QQ==",
      "DnF1ZXJ5VGhlbkZldGNoBQAAAAAAAAABFmtSWWRRWUJrU2o2ZExpSGJCVmQxYUEAAAAAAAAAAxZrUllkUVlCa1NqNmRMaUhiQlZkMWFBAAAAAAAAAAIWa1JZZFFZQmtTajZkTGlIYkJWZDFhQQAAAAAAAAAFFmtSWWRRWUJrU2o2ZExpSGJCVmQxYUEAAAAAAAAABBZrUllkUVlCa1NqNmRMaUhiQlZkMWFB"
    ]
}

All search contexts can be cleared with the _all parameter:

DELETE /_search/scroll/_all

The scroll_id can also be passed as a query string parameter or in the request body. Multiple scroll IDs can be passed as comma separated values:

DELETE /_search/scroll/DXF1ZXJ5QW5kRmV0Y2gBAAAAAAAAAD4WYm9laVYtZndUQlNsdDcwakFMNjU1QQ==,DnF1ZXJ5VGhlbkZldGNoBQAAAAAAAAABFmtSWWRRWUJrU2o2ZExpSGJCVmQxYUEAAAAAAAAAAxZrUllkUVlCa1NqNmRMaUhiQlZkMWFBAAAAAAAAAAIWa1JZZFFZQmtTajZkTGlIYkJWZDFhQQAAAAAAAAAFFmtSWWRRWUJrU2o2ZExpSGJCVmQxYUEAAAAAAAAABBZrUllkUVlCa1NqNmRMaUhiQlZkMWFB

For scroll queries that return a lot of documents it is possible to split the scroll in multiple slices which can be consumed independently:

GET /twitter/_search?scroll=1m
{
    "slice": {
        "id": 0, 
        "max": 2 
    },
    "query": {
        "match" : {
            "title" : "elasticsearch"
        }
    }
}
GET /twitter/_search?scroll=1m
{
    "slice": {
        "id": 1,
        "max": 2
    },
    "query": {
        "match" : {
            "title" : "elasticsearch"
        }
    }
}

The result from the first request returned documents that belong to the first slice (id: 0) and the result from the second request returned documents that belong to the second slice. Since the maximum number of slices is set to 2 the union of the results of the two requests is equivalent to the results of a scroll query without slicing. By default the splitting is done on the shards first and then locally on each shard using the _id field with the following formula: slice(doc) = floorMod(hashCode(doc._id), max) For instance if the number of shards is equal to 2 and the user requested 4 slices then the slices 0 and 2 are assigned to the first shard and the slices 1 and 3 are assigned to the second shard.

Each scroll is independent and can be processed in parallel like any scroll request.

To avoid this cost entirely it is possible to use the doc_values of another field to do the slicing but the user must ensure that the field has the following properties:

GET /twitter/_search?scroll=1m
{
    "slice": {
        "field": "date",
        "id": 0,
        "max": 10
    },
    "query": {
        "match" : {
            "title" : "elasticsearch"
        }
    }
}

For append only time-based indices, the timestamp field can be used safely.

Parameter value: query_then_fetch.

The request is processed in two phases. In the first phase, the query is forwarded to all involved shards. Each shard executes the search and generates a sorted list of results, local to that shard. Each shard returns just enough information to the coordinating node to allow it to merge and re-sort the shard level results into a globally sorted set of results, of maximum length size.

During the second phase, the coordinating node requests the document content (and highlighted snippets, if any) from only the relevant shards.

GET twitter/_search?search_type=query_then_fetch

Parameter value: dfs_query_then_fetch.

Same as "Query Then Fetch", except for an initial scatter phase which goes and computes the distributed term frequencies for more accurate scoring.

GET twitter/_search?search_type=dfs_query_then_fetch

The sort values for each document returned are also returned as part of the response.

The order option can have the following values:

The order defaults to desc when sorting on the _score, and defaults to asc when sorting on anything else.

Elasticsearch supports sorting by array or multi-valued fields. The mode option controls what array value is picked for sorting the document it belongs to. The mode option can have the following values:

The default sort mode in the ascending sort order is min — the lowest value is picked. The default sort mode in the descending order is max — the highest value is picked.

PUT /my_index/_doc/1?refresh
{
   "product": "chocolate",
   "price": [20, 4]
}

POST /_search
{
   "query" : {
      "term" : { "product" : "chocolate" }
   },
   "sort" : [
      {"price" : {"order" : "asc", "mode" : "avg"}}
   ]
}

For numeric fields it is also possible to cast the values from one type to another using the numeric_type option. This option accepts the following values: ["double", "long", "date", "date_nanos"] and can be useful for cross-index search if the sort field is mapped differently on some indices.

Consider for instance these two indices:

PUT /index_double
{
    "mappings": {
        "properties": {
            "field": { "type": "double" }
        }
    }
}

PUT /index_long
{
    "mappings": {
        "properties": {
            "field": { "type": "long" }
        }
    }
}

Since field is mapped as a double in the first index and as a long in the second index, it is not possible to use this field to sort requests that query both indices by default. However you can force the type to one or the other with the numeric_type option in order to force a specific type for all indices:

POST /index_long,index_double/_search
{
   "sort" : [
      {
        "field" : {
            "numeric_type" : "double"
        }
      }
   ]
}

In the example above, values for the index_long index are casted to a double in order to be compatible with the values produced by the index_double index. It is also possible to transform a floating point field into a long but note that in this case floating points are replaced by the largest value that is less than or equal (greater than or equal if the value is negative) to the argument and is equal to a mathematical integer.

This option can also be used to convert a date field that uses millisecond resolution to a date_nanos field with nanosecond resolution. Consider for instance these two indices:

PUT /index_double
{
    "mappings": {
        "properties": {
            "field": { "type": "date" }
        }
    }
}

PUT /index_long
{
    "mappings": {
        "properties": {
            "field": { "type": "date_nanos" }
        }
    }
}

Values in these indices are stored with different resolutions so sorting on these fields will always sort the date before the date_nanos (ascending order). With the numeric_type type option it is possible to set a single resolution for the sort, setting to date will convert the date_nanos to the millisecond resolution while date_nanos will convert the values in the date field to the nanoseconds resolution:

POST /index_long,index_double/_search
{
   "sort" : [
      {
        "field" : {
            "numeric_type" : "date_nanos"
        }
      }
   ]
}

Elasticsearch also supports sorting by fields that are inside one or more nested objects. The sorting by nested field support has a nested sort option with the following properties:

POST /_search
{
   "query" : {
      "term" : { "product" : "chocolate" }
   },
   "sort" : [
       {
          "offer.price" : {
             "mode" :  "avg",
             "order" : "asc",
             "nested": {
                "path": "offer",
                "filter": {
                   "term" : { "offer.color" : "blue" }
                }
             }
          }
       }
    ]
}

POST /_search
{
   "query": {
      "nested": {
         "path": "parent",
         "query": {
            "bool": {
                "must": {"range": {"parent.age": {"gte": 21}}},
                "filter": {
                    "nested": {
                        "path": "parent.child",
                        "query": {"match": {"parent.child.name": "matt"}}
                    }
                }
            }
         }
      }
   },
   "sort" : [
      {
         "parent.child.age" : {
            "mode" :  "min",
            "order" : "asc",
            "nested": {
               "path": "parent",
               "filter": {
                  "range": {"parent.age": {"gte": 21}}
               },
               "nested": {
                  "path": "parent.child",
                  "filter": {
                     "match": {"parent.child.name": "matt"}
                  }
               }
            }
         }
      }
   ]
}

The missing parameter specifies how docs which are missing the sort field should be treated: The missing value can be set to _last, _first, or a custom value (that will be used for missing docs as the sort value). The default is _last.

For example:

GET /_search
{
    "sort" : [
        { "price" : {"missing" : "_last"} }
    ],
    "query" : {
        "term" : { "product" : "chocolate" }
    }
}

By default, the search request will fail if there is no mapping associated with a field. The unmapped_type option allows you to ignore fields that have no mapping and not sort by them. The value of this parameter is used to determine what sort values to emit. Here is an example of how it can be used:

GET /_search
{
    "sort" : [
        { "price" : {"unmapped_type" : "long"} }
    ],
    "query" : {
        "term" : { "product" : "chocolate" }
    }
}

If any of the indices that are queried doesn’t have a mapping for price then Elasticsearch will handle it as if there was a mapping of type long, with all documents in this index having no value for this field.

Allow to sort by _geo_distance. Here is an example, assuming pin.location is a field of type geo_point:

GET /_search
{
    "sort" : [
        {
            "_geo_distance" : {
                "pin.location" : [-70, 40],
                "order" : "asc",
                "unit" : "km",
                "mode" : "min",
                "distance_type" : "arc",
                "ignore_unmapped": true
            }
        }
    ],
    "query" : {
        "term" : { "user" : "kimchy" }
    }
}

The following formats are supported in providing the coordinates:

GET /_search
{
    "sort" : [
        {
            "_geo_distance" : {
                "pin.location" : {
                    "lat" : 40,
                    "lon" : -70
                },
                "order" : "asc",
                "unit" : "km"
            }
        }
    ],
    "query" : {
        "term" : { "user" : "kimchy" }
    }
}

GET /_search
{
    "sort" : [
        {
            "_geo_distance" : {
                "pin.location" : "40,-70",
                "order" : "asc",
                "unit" : "km"
            }
        }
    ],
    "query" : {
        "term" : { "user" : "kimchy" }
    }
}

GET /_search
{
    "sort" : [
        {
            "_geo_distance" : {
                "pin.location" : "drm3btev3e86",
                "order" : "asc",
                "unit" : "km"
            }
        }
    ],
    "query" : {
        "term" : { "user" : "kimchy" }
    }
}

GET /_search
{
    "sort" : [
        {
            "_geo_distance" : {
                "pin.location" : [-70, 40],
                "order" : "asc",
                "unit" : "km"
            }
        }
    ],
    "query" : {
        "term" : { "user" : "kimchy" }
    }
}

Multiple geo points can be passed as an array containing any geo_point format, for example

GET /_search
{
    "sort" : [
        {
            "_geo_distance" : {
                "pin.location" : [[-70, 40], [-71, 42]],
                "order" : "asc",
                "unit" : "km"
            }
        }
    ],
    "query" : {
        "term" : { "user" : "kimchy" }
    }
}

and so forth.

The final distance for a document will then be min/max/avg (defined via mode) distance of all points contained in the document to all points given in the sort request.

Allow to sort based on custom scripts, here is an example:

GET /_search
{
    "query" : {
        "term" : { "user" : "kimchy" }
    },
    "sort" : {
        "_script" : {
            "type" : "number",
            "script" : {
                "lang": "painless",
                "source": "doc['field_name'].value * params.factor",
                "params" : {
                    "factor" : 1.1
                }
            },
            "order" : "asc"
        }
    }
}

When sorting on a field, scores are not computed. By setting track_scores to true, scores will still be computed and tracked.

GET /_search
{
    "track_scores": true,
    "sort" : [
        { "post_date" : {"order" : "desc"} },
        { "name" : "desc" },
        { "age" : "desc" }
    ],
    "query" : {
        "term" : { "user" : "kimchy" }
    }
}

When sorting, the relevant sorted field values are loaded into memory. This means that per shard, there should be enough memory to contain them. For string based types, the field sorted on should not be analyzed / tokenized. For numeric types, if possible, it is recommended to explicitly set the type to narrower types (like short, integer and float).

To disable the stored fields (and metadata fields) entirely use: _none_:

GET /_search
{
    "stored_fields": "_none_",
    "query" : {
        "term" : { "user" : "kimchy" }
    }
}