WARNING: The 2.x versions of Elasticsearch have passed their EOL dates. If you are running a 2.x version, we strongly advise you to upgrade.
This documentation is no longer maintained and may be removed. For the latest information, see the current Elasticsearch documentation.
Most Fields
editMost Fields
editFull-text search is a battle between recall—returning all the documents that are relevant—and precision—not returning irrelevant documents. The goal is to present the user with the most relevant documents on the first page of results.
To improve recall, we cast the net wide—we include not only
documents that match the user’s search terms exactly, but also
documents that we believe to be pertinent to the query. If a user searches
for “quick brown fox,” a document that contains fast foxes
may well be
a reasonable result to return.
If the only pertinent document that we have is the one containing fast
foxes
, it will appear at the top of the results list. But of course, if
we have 100 documents that contain the words quick brown fox
, then the
fast foxes
document may be considered less relevant, and we would want to
push it further down the list. After including many potential matches, we
need to ensure that the best ones rise to the top.
A common technique for fine-tuning full-text relevance is to index the same text in multiple ways, each of which provides a different relevance signal. The main field would contain terms in their broadest-matching form to match as many documents as possible. For instance, we could do the following:
-
Use a stemmer to index
jumps
,jumping
, andjumped
as their root form:jump
. Then it doesn’t matter if the user searches forjumped
; we could still match documents containingjumping
. -
Include synonyms like
jump
,leap
, andhop
. -
Remove diacritics, or accents: for example,
ésta
,está
, andesta
would all be indexed without accents asesta
.
However, if we have two documents, one of which contains jumped
and the
other jumping
, the user would probably expect the first document to rank
higher, as it contains exactly what was typed in.
We can achieve this by indexing the same text in other fields to provide more-precise matching. One field may contain the unstemmed version, another the original word with diacritics, and a third might use shingles to provide information about word proximity. These other fields act as signals that increase the relevance score of each matching document. The more fields that match, the better.
A document is included in the results list if it matches the broad-matching main field. If it also matches the signal fields, it gets extra points and is pushed up the results list.
We discuss synonyms, word proximity, partial-matching and other potential signals later in the book, but we will use the simple example of stemmed and unstemmed fields to illustrate this technique.
Multifield Mapping
editThe first thing to do is to set up our field to be indexed twice: once in a stemmed form and once in an unstemmed form. To do this, we will use multifields, which we introduced in String Sorting and Multifields:
DELETE /my_index PUT /my_index { "settings": { "number_of_shards": 1 }, "mappings": { "my_type": { "properties": { "title": { "type": "string", "analyzer": "english", "fields": { "std": { "type": "string", "analyzer": "standard" } } } } } } }
See Relevance Is Broken!. |
|
The |
|
The |
Next we index some documents:
PUT /my_index/my_type/1 { "title": "My rabbit jumps" } PUT /my_index/my_type/2 { "title": "Jumping jack rabbits" }
Here is a simple match
query on the title
field for jumping rabbits
:
GET /my_index/_search { "query": { "match": { "title": "jumping rabbits" } } }
This becomes a query for the two stemmed terms jump
and rabbit
, thanks to the
english
analyzer. The title
field of both documents contains both of those
terms, so both documents receive the same score:
{ "hits": [ { "_id": "1", "_score": 0.42039964, "_source": { "title": "My rabbit jumps" } }, { "_id": "2", "_score": 0.42039964, "_source": { "title": "Jumping jack rabbits" } } ] }
If we were to query just the title.std
field, then only document 2 would
match. However, if we were to query both fields and to combine their scores
by using the bool
query, then both documents would match (thanks to the title
field) and document 2 would score higher (thanks to the title.std
field):
GET /my_index/_search { "query": { "multi_match": { "query": "jumping rabbits", "type": "most_fields", "fields": [ "title", "title.std" ] } } }
We want to combine the scores from all matching fields, so we use the
|
{ "hits": [ { "_id": "2", "_score": 0.8226396, "_source": { "title": "Jumping jack rabbits" } }, { "_id": "1", "_score": 0.10741998, "_source": { "title": "My rabbit jumps" } } ] }
We are using the broad-matching title
field to include as many documents as
possible—to increase recall—but we use the title.std
field as a
signal to push the most relevant results to the top.
The contribution of each field to the final score can be controlled by
specifying custom boost
values. For instance, we could boost the title
field to make it the most important field, thus reducing the effect of any
other signal fields: