How-To Guides

You can delete the documents matching a search by calling delete on the Search object instead of execute like this:

s = Search(index='i').query("match", title="python")
response = s.delete()

The library provides classes for all Elasticsearch query types. Pass all the parameters as keyword arguments. The classes accept any keyword arguments, the dsl then takes all arguments passed to the constructor and serializes them as top-level keys in the resulting dictionary (and thus the resulting json being sent to elasticsearch). This means that there is a clear one-to-one mapping between the raw query and its equivalent in the DSL:

from elasticsearch.dsl.query import MultiMatch, Match

# {"multi_match": {"query": "python django", "fields": ["title", "body"]}}
MultiMatch(query='python django', fields=['title', 'body'])

# {"match": {"title": {"query": "web framework", "type": "phrase"}}}
Match(title={"query": "web framework", "type": "phrase"})

You can use the Q shortcut to construct the instance using a name with parameters or the raw dict:

from elasticsearch.dsl import Q

Q("multi_match", query='python django', fields=['title', 'body'])
Q({"multi_match": {"query": "python django", "fields": ["title", "body"]}})

To add the query to the Search object, use the .query() method:

q = Q("multi_match", query='python django', fields=['title', 'body'])
s = s.query(q)

The method also accepts all the parameters as the Q shortcut:

s = s.query("multi_match", query='python django', fields=['title', 'body'])

If you already have a query object, or a dict representing one, you can just override the query used in the Search object:

s.query = Q('bool', must=[Q('match', title='python'), Q('match', body='best')])

Sometimes you want to refer to a field within another field, either as a multi-field (title.keyword) or in a structured json document like address.city. To make it easier, the Q shortcut (as well as the query, filter, and exclude methods on Search class) allows you to use _+ (double underscore) in place of a dot in a keyword argument:

s = Search()
s = s.filter('term', category__keyword='Python')
s = s.query('match', address__city='prague')

Alternatively you can always fall back to python’s kwarg unpacking if you prefer:

s = Search()
s = s.filter('term', **{'category.keyword': 'Python'})
s = s.query('match', **{'address.city': 'prague'})

Query objects can be combined using logical operators:

Q("match", title='python') | Q("match", title='django')
# {"bool": {"should": [...]}}

Q("match", title='python') & Q("match", title='django')
# {"bool": {"must": [...]}}

~Q("match", title="python")
# {"bool": {"must_not": [...]}}

When you call the .query() method multiple times, the & operator will be used internally:

s = s.query().query()
print(s.to_dict())
# {"query": {"bool": {...}}}

If you want to have precise control over the query form, use the Q shortcut to directly construct the combined query:

q = Q('bool',
    must=[Q('match', title='python')],
    should=[Q(...), Q(...)],
    minimum_should_match=1
)
s = Search().query(q)

If you want to add a query in a filter context you can use the filter() method to make things easier:

s = Search()
s = s.filter('terms', tags=['search', 'python'])

Behind the scenes this will produce a Bool query and place the specified terms query into its filter branch, making it equivalent to:

s = Search()
s = s.query('bool', filter=[Q('terms', tags=['search', 'python'])])

If you want to use the post_filter element for faceted navigation, use the .post_filter() method.

You can also exclude() items from your query like this:

s = Search()
s = s.exclude('terms', tags=['search', 'python'])

which is shorthand for: s = s.query('bool', filter=[~Q('terms', tags=['search', 'python'])])

To define an aggregation, you can use the A shortcut:

from elasticsearch.dsl import A

A('terms', field='tags')
# {"terms": {"field": "tags"}}

To nest aggregations, you can use the .bucket(), .metric() and .pipeline() methods:

a = A('terms', field='category')
# {'terms': {'field': 'category'}}

a.metric('clicks_per_category', 'sum', field='clicks')\
    .bucket('tags_per_category', 'terms', field='tags')
# {
#   'terms': {'field': 'category'},
#   'aggs': {
#     'clicks_per_category': {'sum': {'field': 'clicks'}},
#     'tags_per_category': {'terms': {'field': 'tags'}}
#   }
# }

To add aggregations to the Search object, use the .aggs property, which acts as a top-level aggregation:

s = Search()
a = A('terms', field='category')
s.aggs.bucket('category_terms', a)
# {
#   'aggs': {
#     'category_terms': {
#       'terms': {
#         'field': 'category'
#       }
#     }
#   }
# }

or

s = Search()
s.aggs.bucket('articles_per_day', 'date_histogram', field='publish_date', interval='day')\
    .metric('clicks_per_day', 'sum', field='clicks')\
    .pipeline('moving_click_average', 'moving_avg', buckets_path='clicks_per_day')\
    .bucket('tags_per_day', 'terms', field='tags')

s.to_dict()
# {
#   "aggs": {
#     "articles_per_day": {
#       "date_histogram": { "interval": "day", "field": "publish_date" },
#       "aggs": {
#         "clicks_per_day": { "sum": { "field": "clicks" } },
#         "moving_click_average": { "moving_avg": { "buckets_path": "clicks_per_day" } },
#         "tags_per_day": { "terms": { "field": "tags" } }
#       }
#     }
#   }
# }

You can access an existing bucket by its name:

s = Search()

s.aggs.bucket('per_category', 'terms', field='category')
s.aggs['per_category'].metric('clicks_per_category', 'sum', field='clicks')
s.aggs['per_category'].bucket('tags_per_category', 'terms', field='tags')

As opposed to other methods on the Search objects, defining aggregations is done in-place (does not return a copy).

To issue a kNN search, use the .knn() method:

s = Search()
vector = get_embedding("search text")

s = s.knn(
    field="embedding",
    k=5,
    num_candidates=10,
    query_vector=vector
)

The field, k and num_candidates arguments can be given as positional or keyword arguments and are required. In addition to these, query_vector or query_vector_builder must be given as well.

The .knn() method can be invoked multiple times to include multiple kNN searches in the request.

To specify sorting order, use the .sort() method:

s = Search().sort(
    'category',
    '-title',
    {"lines" : {"order" : "asc", "mode" : "avg"}}
)

It accepts positional arguments which can be either strings or dictionaries. String value is a field name, optionally prefixed by the - sign to specify a descending order.

To reset the sorting, just call the method with no arguments:

s = s.sort()

To specify the from/size parameters, use the Python slicing API:

s = s[10:20]
# {"from": 10, "size": 10}

s = s[:20]
# {"size": 20}

s = s[10:]
# {"from": 10}

s = s[10:20][2:]
# {"from": 12, "size": 8}

If you want to access all the documents matched by your query you can use the scan method which uses the scan/scroll elasticsearch API:

for hit in s.scan():
    print(hit.title)

Note that in this case the results won’t be sorted.

To set common attributes for highlighting use the highlight_options method:

s = s.highlight_options(order='score')

Enabling highlighting for individual fields is done using the highlight method:

s = s.highlight('title')
# or, including parameters:
s = s.highlight('title', fragment_size=50)

The fragments in the response will then be available on each Result object as .meta.highlight.FIELD which will contain the list of fragments:

response = s.execute()
for hit in response:
    for fragment in hit.meta.highlight.title:
        print(fragment)

To specify a suggest request on your Search object use the suggest method:

# check for correct spelling
s = s.suggest('my_suggestion', 'pyhton', term={'field': 'title'})

The first argument is the name of the suggestions (name under which it will be returned), second is the actual text you wish the suggester to work on and the keyword arguments will be added to the suggest’s json as-is which means that it should be one of term, phrase or completion to indicate which type of suggester should be used.

To collapse search results use the collapse method on your Search object:

s = Search().query("match", message="GET /search")
# collapse results by user_id
s = s.collapse("user_id")

The top hits will only include one result per user_id. You can also expand each collapsed top hit with the inner_hits parameter, max_concurrent_group_searches being the number of concurrent requests allowed to retrieve the inner hits per group:

inner_hits = {"name": "recent_search", "size": 5, "sort": [{"@timestamp": "desc"}]}
s = s.collapse("user_id", inner_hits=inner_hits, max_concurrent_group_searches=4)

To use Elasticsearch’s more_like_this functionality, you can use the MoreLikeThis query type.

A simple example is below

from elasticsearch.dsl.query import MoreLikeThis
from elasticsearch.dsl import Search

my_text = 'I want to find something similar'

s = Search()
# We're going to match based only on two fields, in this case text and title
s = s.query(MoreLikeThis(like=my_text, fields=['text', 'title']))
# You can also exclude fields from the result to make the response quicker in the normal way
s = s.source(exclude=["text"])
response = s.execute()

for hit in response:
    print(hit.title)

To set extra properties of the search request, use the .extra() method. This can be used to define keys in the body that cannot be defined via a specific API method like explain or search_after:

s = s.extra(explain=True)

To set query parameters, use the .params() method:

s = s.params(routing="42")

If you need to limit the fields being returned by elasticsearch, use the source() method:

# only return the selected fields
s = s.source(['title', 'body'])
# don't return any fields, just the metadata
s = s.source(False)
# explicitly include/exclude fields
s = s.source(includes=["title"], excludes=["user.*"])
# reset the field selection
s = s.source(None)

The search object can be serialized into a dictionary by using the .to_dict() method.

You can also create a Search object from a dict using the from_dict class method. This will create a new Search object and populate it using the data from the dict:

s = Search.from_dict({"query": {"match": {"title": "python"}}})

If you wish to modify an existing Search object, overriding it’s properties, instead use the update_from_dict method that alters an instance in-place:

s = Search(index='i')
s.update_from_dict({"query": {"match": {"title": "python"}}, "size": 42})

To access to the hits returned by the search, access the hits property or just iterate over the Response object:

response = s.execute()
print('Total %d hits found.' % response.hits.total)
for h in response:
    print(h.title, h.body)

The individual hits is wrapped in a convenience class that allows attribute access to the keys in the returned dictionary. All the metadata for the results are accessible via meta (without the leading _):

response = s.execute()
h = response.hits[0]
print('/%s/%s/%s returned with score %f' % (
    h.meta.index, h.meta.doc_type, h.meta.id, h.meta.score))

Aggregations are available through the aggregations property:

for tag in response.aggregations.per_tag.buckets:
    print(tag.key, tag.max_lines.value)

The Document instances use native python types like str and datetime. In case of Object or Nested fields an instance of the InnerDoc subclass is used, as in the add_comment method in the above example where we are creating an instance of the Comment class.

There are some specific types that were created as part of this library to make working with some field types easier, for example the Range object used in any of the range fields:

from elasticsearch.dsl import Document, DateRange, Keyword, Range

class RoomBooking(Document):
    room = Keyword()
    dates = DateRange()


rb = RoomBooking(
  room='Conference Room II',
  dates=Range(
    gte=datetime(2018, 11, 17, 9, 0, 0),
    lt=datetime(2018, 11, 17, 10, 0, 0)
  )
)

# Range supports the in operator correctly:
datetime(2018, 11, 17, 9, 30, 0) in rb.dates # True

# you can also get the limits and whether they are inclusive or exclusive:
rb.dates.lower # datetime(2018, 11, 17, 9, 0, 0), True
rb.dates.upper # datetime(2018, 11, 17, 10, 0, 0), False

# empty range is unbounded
Range().lower # None, False

Document fields can be defined using standard Python type hints if desired. Here are some simple examples:

from typing import Optional

class Post(Document):
    title: str                      # same as title = Text(required=True)
    created_at: Optional[datetime]  # same as created_at = Date(required=False)
    published: bool                 # same as published = Boolean(required=True)

It is important to note that when using Field subclasses such as Text, Date and Boolean, they must be given in the right-side of an assignment, as shown in examples above. Using these classes as type hints will result in errors.

Python types are mapped to their corresponding field type according to the following table:

To type a field as optional, the standard Optional modifier from the Python typing package can be used. When using Python 3.10 or newer, "pipe" syntax can also be used, by adding | None to a type. The List modifier can be added to a field to convert it to an array, similar to using the multi=True argument on the field object.

from typing import Optional, List

class MyDoc(Document):
    pub_date: Optional[datetime]  # same as pub_date = Date()
    middle_name: str | None       # same as middle_name = Text()
    authors: List[str]            # same as authors = Text(multi=True, required=True)
    comments: Optional[List[str]] # same as comments = Text(multi=True)

A field can also be given a type hint of an InnerDoc subclass, in which case it becomes an Object field of that class. When the InnerDoc subclass is wrapped with List, a Nested field is created instead.

from typing import List

class Address(InnerDoc):
    ...

class Comment(InnerDoc):
    ...

class Post(Document):
    address: Address         # same as address = Object(Address, required=True)
    comments: List[Comment]  # same as comments = Nested(Comment, required=True)

Unfortunately it is impossible to have Python type hints that uniquely identify every possible Elasticsearch field type. To choose a field type that is different than the ones in the table above, the field instance can be added explicitly as a right-side assignment in the field declaration. The next example creates a field that is typed as Optional[str], but is mapped to Keyword instead of Text:

class MyDocument(Document):
    category: Optional[str] = Keyword()

This form can also be used when additional options need to be given to initialize the field, such as when using custom analyzer settings or changing the required default:

class Comment(InnerDoc):
    content: str = Text(analyzer='snowball', required=True)

When using type hints as above, subclasses of Document and InnerDoc inherit some of the behaviors associated with Python dataclasses, as defined by PEP 681 and the dataclass_transform decorator. To add per-field dataclass options such as default or default_factory, the mapped_field() wrapper can be used on the right side of a typed field declaration:

class MyDocument(Document):
    title: str = mapped_field(default="no title")
    created_at: datetime = mapped_field(default_factory=datetime.now)
    published: bool = mapped_field(default=False)
    category: str = mapped_field(Keyword(required=True), default="general")

When using the mapped_field() wrapper function, an explicit field type instance can be passed as a first positional argument, as the category field does in the example above.

Static type checkers such as mypy and pyright can use the type hints and the dataclass-specific options added to the mapped_field() function to improve type inference and provide better real-time suggestions in IDEs.

One situation in which type checkers can’t infer the correct type is when using fields as class attributes. Consider the following example:

class MyDocument(Document):
    title: str

doc = MyDocument()
# doc.title is typed as "str" (correct)
# MyDocument.title is also typed as "str" (incorrect)

To help type checkers correctly identify class attributes as such, the M generic must be used as a wrapper to the type hint, as shown in the next examples:

from elasticsearch.dsl import M

class MyDocument(Document):
    title: M[str]
    created_at: M[datetime] = mapped_field(default_factory=datetime.now)

doc = MyDocument()
# doc.title is typed as "str"
# doc.created_at is typed as "datetime"
# MyDocument.title is typed as "InstrumentedField"
# MyDocument.created_at is typed as "InstrumentedField"

Note that the M type hint does not provide any runtime behavior and its use is not required, but it can be useful to eliminate spurious type errors in IDEs or type checking builds.

The InstrumentedField objects returned when fields are accessed as class attributes are proxies for the field instances that can be used anywhere a field needs to be referenced, such as when specifying sort options in a Search object:

# sort by creation date descending, and title ascending
s = MyDocument.search().sort(-MyDocument.created_at, MyDocument.title)

When specifying sorting order, the {plus} and - unary operators can be used on the class field attributes to indicate ascending and descending order.

Finally, the ClassVar annotation can be used to define a regular class attribute that should not be mapped to the Elasticsearch index:

from typing import ClassVar

class MyDoc(Document):
  title: M[str] created_at: M[datetime] =
  mapped_field(default_factory=datetime.now) my_var:
  ClassVar[str] # regular class variable, ignored by Elasticsearch

The DSL module will always respect the timezone information (or lack thereof) on the datetime objects passed in or stored in Elasticsearch. Elasticsearch itself interprets all datetimes with no timezone information as UTC. If you wish to reflect this in your python code, you can specify default_timezone when instantiating a Date field:

class Post(Document):
    created_at = Date(default_timezone='UTC')

In that case any datetime object passed in (or parsed from elasticsearch) will be treated as if it were in UTC timezone.

Before you first use the Post document type, you need to create the mappings in Elasticsearch. For that you can either use the index object or create the mappings directly by calling the init class method:

# create the mappings in Elasticsearch
Post.init()

This code will typically be run in the setup for your application during a code deploy, similar to running database migrations.

To create a new Post document just instantiate the class and pass in any fields you wish to set, you can then use standard attribute setting to change/add more fields. Note that you are not limited to the fields defined explicitly:

# instantiate the document
first = Post(title='My First Blog Post, yay!', published=True)
# assign some field values, can be values or lists of values
first.category = ['everything', 'nothing']
# every document has an id in meta
first.meta.id = 47


# save the document into the cluster
first.save()

All the metadata fields (id, routing, index etc) can be accessed (and set) via a meta attribute or directly using the underscored variant:

post = Post(meta={'id': 42})

# prints 42
print(post.meta.id)

# override default index
post.meta.index = 'my-blog'

To retrieve an existing document use the get class method:

# retrieve the document
first = Post.get(id=42)
# now we can call methods, change fields, ...
first.add_comment('me', 'This is nice!')
# and save the changes into the cluster again
first.save()

The Update API can also be used via the update method. By default any keyword arguments, beyond the parameters of the API, will be considered fields with new values. Those fields will be updated on the local copy of the document and then sent over as partial document to be updated:

# retrieve the document
first = Post.get(id=42)
# you can update just individual fields which will call the update API
# and also update the document in place
first.update(published=True, published_by='me')

In case you wish to use a painless script to perform the update you can pass in the script string as script or the id of a stored script via script_id. All additional keyword arguments to the update method will then be passed in as parameters of the script. The document will not be updated in place.

# retrieve the document
first = Post.get(id=42)
# we execute a script in elasticsearch with additional kwargs being passed
# as params into the script
first.update(script='ctx._source.category.add(params.new_category)',
             new_category='testing')

If the document is not found in elasticsearch an exception (elasticsearch.NotFoundError) will be raised. If you wish to return None instead just pass in ignore=404 to suppress the exception:

p = Post.get(id='not-in-es', ignore=404)
p is None

When you wish to retrieve multiple documents at the same time by their id you can use the mget method:

posts = Post.mget([42, 47, 256])

mget will, by default, raise a NotFoundError if any of the documents wasn’t found and RequestError if any of the document had resulted in error. You can control this behavior by setting parameters:

The index associated with the Document is accessible via the _index class property which gives you access to the index class.

The _index attribute is also home to the load_mappings method which will update the mapping on the Index from elasticsearch. This is very useful if you use dynamic mappings and want the class to be aware of those fields (for example if you wish the Date fields to be properly (de)serialized):

Post._index.load_mappings()

To delete a document just call its delete method:

first = Post.get(id=42)
first.delete()

To specify analyzer values for Text fields you can just use the name of the analyzer (as a string) and either rely on the analyzer being defined (like built-in analyzers) or define the analyzer yourself manually.

Alternatively you can create your own analyzer and have the persistence layer handle its creation, from our example earlier:

from elasticsearch.dsl import analyzer, tokenizer

my_analyzer = analyzer('my_analyzer',
    tokenizer=tokenizer('trigram', 'nGram', min_gram=3, max_gram=3),
    filter=['lowercase']
)

Each analysis object needs to have a name (my_analyzer and trigram in our example) and tokenizers, token filters and char filters also need to specify type (nGram in our example).

Once you have an instance of a custom analyzer you can also call the analyze API on it by using the simulate method:

response = my_analyzer.simulate('Hello World!')

# ['hel', 'ell', 'llo', 'lo ', 'o w', ' wo', 'wor', 'orl', 'rld', 'ld!']
tokens = [t.token for t in response.tokens]

To search for this document type, use the search class method:

# by calling .search we get back a standard Search object
s = Post.search()
# the search is already limited to the index and doc_type of our document
s = s.filter('term', published=True).query('match', title='first')


results = s.execute()

# when you execute the search the results are wrapped in your document class (Post)
for post in results:
    print(post.meta.score, post.title)

Alternatively you can just take a Search object and restrict it to return our document type, wrapped in correct class:

s = Search()
s = s.doc_type(Post)

You can also combine document classes with standard doc types (just strings), which will be treated as before. You can also pass in multiple Document subclasses and each document in the response will be wrapped in it’s class.

If you want to run suggestions, just use the suggest method on the Search object:

s = Post.search()
s = s.suggest('title_suggestions', 'pyth', completion={'field': 'title_suggest'})

response = s.execute()

for result in response.suggest.title_suggestions:
    print('Suggestions for %s:' % result.text)
    for option in result.options:
        print('  %s (%r)' % (option.text, option.payload))

In the Meta class inside your document definition you can define various metadata for your document:

Any attributes on the Meta class that are instance of MetaField will be used to control the mapping of the meta fields (_all, dynamic etc). Just name the parameter (without the leading underscore) as the field you wish to map and pass any parameters to the MetaField class:

class Post(Document):
    title = Text()

    class Meta:
        all = MetaField(enabled=False)
        dynamic = MetaField('strict')

This section of the Document definition can contain any information about the index, its name, settings and other attributes:

You can use standard Python inheritance to extend models, this can be useful in a few scenarios. For example if you want to have a BaseDocument defining some common fields that several different Document classes should share:

class User(InnerDoc):
    username = Text(fields={'keyword': Keyword()})
    email = Text()

class BaseDocument(Document):
    created_by = Object(User)
    created_date = Date()
    last_updated = Date()

    def save(**kwargs):
        if not self.created_date:
            self.created_date = datetime.now()
        self.last_updated = datetime.now()
        return super(BaseDocument, self).save(**kwargs)

class BlogPost(BaseDocument):
    class Index:
        name = 'blog'

Another use case would be using the join type to have multiple different entities in a single index. You can see an example of this approach. Note that in this case, if the subclasses don’t define their own Index classes, the mappings are merged and shared between all the subclasses.

The DSL module also exposes an option to manage index templates in elasticsearch using the IndexTemplate class which has very similar API to Index.

Once an index template is saved in elasticsearch it’s contents will be automatically applied to new indices (existing indices are completely unaffected by templates) that match the template pattern (any index starting with blogs- in our example), even if the index is created automatically upon indexing a document into that index.

Potential workflow for a set of time based indices governed by a single template:

from datetime import datetime

from elasticsearch.dsl import Document, Date, Text


class Log(Document):
    content = Text()
    timestamp = Date()

    class Index:
        name = "logs-*"
        settings = {
          "number_of_shards": 2
        }

    def save(self, **kwargs):
        # assign now if no timestamp given
        if not self.timestamp:
            self.timestamp = datetime.now()

        # override the index to go to the proper timeslot
        kwargs['index'] = self.timestamp.strftime('logs-%Y%m%d')
        return super().save(**kwargs)

# once, as part of application setup, during deploy/migrations:
logs = Log._index.as_template('logs', order=0)
logs.save()

# to perform search across all logs:
search = Log.search()

There are several different facets available:

By default facet results will only calculate document count, if you wish for a different metric you can pass in any single value metric aggregation as the metric kwarg (TermsFacet(field='tags', metric=A('max', field=timestamp))). When specifying metric the results will be, by default, sorted in descending order by that metric. To change it to ascending specify metric_sort="asc" and to just sort by document count use metric_sort=False.

If you require any custom behavior or modifications simply override one or more of the methods responsible for the class' functions:

the response returned from the FacetedSearch object (by calling .execute()) is a subclass of the standard Response class that adds a property called facets which contains a dictionary with lists of buckets -each represented by a tuple of key, document count and a flag indicating whether this value has been filtered on.

The search object can be serialized into a dictionary by using the .to_dict() method.

You can also create a Update By Query object from a dict using the from_dict class method. This will create a new Update By Query object and populate it using the data from the dict:

ubq = UpdateByQuery.from_dict({"query": {"match": {"title": "python"}}})

If you wish to modify an existing Update By Query object, overriding it’s properties, instead use the update_from_dict method that alters an instance in-place:

ubq = UpdateByQuery(index='i')
ubq.update_from_dict({"query": {"match": {"title": "python"}}, "size": 42})

To set extra properties of the search request, use the .extra() method. This can be used to define keys in the body that cannot be defined via a specific API method like explain:

ubq = ubq.extra(explain=True)

To set query parameters, use the .params() method:

ubq = ubq.params(routing="42")

These warnings come from the aiohttp package, which is used internally by the AsyncElasticsearch client. They appear often when the application exits and are caused by HTTP connections that are open when they are garbage collected. To avoid these warnings, make sure that you close your connections.

es = async_connections.get_connection()
await es.close()