Page Search

For page searches on consumerfinance.gov, we use Elasticsearch and the django-opensearch-dsl library, which is a lightweight wrapper around opensearch-dsl-py.

Indexing

For any of our Django apps that need to implement search for their Django models or Wagtail page types, we include a documents.py file that defines the Elasticsearch documents that will map the model or page to the Elasticsearch index.

The Document class includes three things:

  1. The Elasticsearch index configuration
  2. The Django model information
  3. Custom fields to index, and any preparation that they require

We'll use our Ask CFPB answer search document as an example for each of these:

from django_opensearch_dsl import Document
from django_opensearch_dsl.registries import registry

@registry.register_document
class AnswerPageDocument(Document):
    pass

Elasticsearch index configuration

The index configuration is provided by an Index subclass on the document class that defines the django-opensearch-dsl index options.

from search.elasticsearch_helpers import environment_specific_index

class AnswerPageDocument(Document):

    class Index:
        name = environment_specific_index('ask-cfpb')
        settings = {'number_of_shards': 1,
                    'number_of_replicas': 0}

For index naming, we have a helper function, environment_specific_index, that will generate the index name specific to the deployment environment. This allows each index to be isolated to a deployment environment within an Elasticsearch cluster.

Django model information

The Django model information is provided by a Django class on the document class that defines the model and any fields names to be indexed directly from the model. A get_queryset method can be overriden to perform any filtering on the model's queryset before content is indexed.

from ask_cfpb.models.answer_page import AnswerPage

class AnswerPageDocument(Document):

    def get_queryset(self):
        query_set = super().get_queryset()
        return query_set.filter(live=True, redirect_to_page=None)

    class Django:
        model = AnswerPage

        fields = [
            'search_tags',
            'language',
        ]

The fields in fields will be indexed without any preparation/manipulation, directly as are stored on the model.

Custom fields

Sometimes it might be desirable to index a field as an alternative type — say, the string that matches an integer for a Django field that species choices.

It might also be desirable to construct a field to index from multiple fields on the model, particularly for Wagtail pages with stream fields.

We may also want to specify Elasticsearch-specific field properties, like a custom analyzer.

To do so, we specify custom fields as attributes on the document class, with an attr argument that specifies the field on the model to reference.

from django_opensearch_dsl import fields

from search.elasticsearch_helpers import synonym_analyzer

class AnswerPageDocument(Document):
    text = fields.TextField(attr="text", analyzer=synonym_analyzer)

The attr on the model can be a @property or a Django model field.

We can also do any data preparation/manipulation for fields using prepare_-prefixed methods.

from django_opensearch_dsl import fields

class AnswerPageDocument(Document):
    portal_topics = fields.KeywordField()

    def prepare_portal_topics(self, instance):
        return [topic.heading for topic in instance.portal_topic.all()]

Helpers

We provide a few common helpers in search.elasticsearch_helpers for use in creating document classes:

  • environment_specific_index(base_name): Generate the index name for the base_name that is specific to the deployment environment. This allows each index to be isolated to a deployment environment within an Elasticsearch cluster.
  • ngram_tokenizer: A reusable ngram analyzer for creating fields that autocomplete terms. This is used for type-ahead search boxes.
  • synonym_analyzer: A reusable analyzer for creating fields that will match synonyms of a search term.
from search.elasticsearch_helpers import (
    environment_specific_index,
    ngram_tokenizer,
    synonym_analyzer,
)

Building the index

With the Document class created for your model in a documents.py module within a Django app listed in INSTALLED_APPS, all that is left to do is to use the django-opensearch-dsl management commands to rebuild the index:

./cfgov/manage.py opensearch index --force rebuild [INDEX]
./cfgov/manage.py opensearch document --force --indices [INDEX] --refresh --parallel index

The above commands can also be run to rebuild the index for an app's models at any time.

Finally, the indexes for all apps can be rebuilt using:

./cfgov/manage.py opensearch index --force rebuild
./cfgov/manage.py opensearch document --force --refresh --parallel index

Conveniently, we have a shell script that invokes the above two commands to rebuild all indexes:

./index.sh

Searching

The document class provides a search() class method that returns a Search object. The Search object is opensearch-dsl-py's representation of Elasticsearch search requests.

To query for a specific term, for example:

from ask_cfpb.documents import AnswerPageDocument

AnswerPageDocument.search().query(
    "match", text={"query": search_term, "operator": "AND"}
)

We can also add a filter context before querying, which we do to limit results to a specific language in Ask CFPB:

from ask_cfpb.documents import AnswerPageDocument

search = AnswerPageDocument.search().filter("term", language=language)
search.query(
    "match", text={"query": search_term, "operator": "AND"}
)

Autocomplete

For search box autocomplete, we use a field with our ngram_tokenizer analyzer and then issue a "match" search query for that field.

Using the Ask CFPB document search above, with its language filter context, this looks like:

from search.elasticsearch_helpers import ngram_tokenizer

class AnswerPageDocument(Document):
    autocomplete = fields.TextField(analyzer=ngram_tokenizer)

search = AnswerPageDocument.search().filter("term", language=language)
search.query('match', autocomplete=search_term)

Suggestions

For suggested spelling corrections for search terms, the Search object has a suggest() method that provides spelling suggestions for a given term on a given field.

Using the Ask CFPB document search above, with its language filter context, this looks like:

from ask_cfpb.documents import AnswerPageDocument

search = AnswerPageDocument.search().filter("term", language=language)
search.suggest('suggestion', search_term, term={'field': 'text'})

References