ankane/searchkick

Searchkick

:rocket: Intelligent search made easy

Searchkick learns what your users are looking for. As more people search, it gets smarter and the results get better. It’s friendly for developers - and magical for your users.

Searchkick handles:

• stemming - tomatoes matches tomato
• special characters - jalapeno matches jalapeño
• extra whitespace - dishwasher matches dish washer
• misspellings - zuchini matches zucchini
• custom synonyms - pop matches soda

Plus:

• query like SQL - no need to learn a new query language
• reindex without downtime
• easily personalize results for each user
• autocomplete
• “Did you mean” suggestions
• supports many languages
• works with Active Record and Mongoid

Check out Searchjoy for analytics and Autosuggest for query suggestions

:tangerine: Battle-tested at Instacart

Contents

• Getting Started
• Querying
• Indexing
• Intelligent Search
• Instant Search / Autocomplete
• Aggregations
• Testing
• Deployment
• Performance
• Advanced Search
• Reference
• Contributing

Searchkick 6.0 was recently released! See how to upgrade

Getting Started

Install Elasticsearch or OpenSearch. For Homebrew, use:

brew install opensearch
brew services start opensearch

Add these lines to your application’s Gemfile:

gem "searchkick"

gem "elasticsearch"   # select one
gem "opensearch-ruby" # select one

The latest version works with Elasticsearch 8 and 9 and OpenSearch 2 and 3. For Elasticsearch 7 and OpenSearch 1, use version 5.5.2 and this readme.

Add searchkick to models you want to search.

class Product < ApplicationRecord
  searchkick
end

Add data to the search index.

Product.reindex

And to query, use:

products = Product.search("apples")
products.each do |product|
  puts product.name
end

Searchkick supports the complete Elasticsearch Search API and OpenSearch Search API. As your search becomes more advanced, we recommend you use the search server DSL for maximum flexibility.

Querying

Query like SQL

Product.search("apples").where(in_stock: true).limit(10).offset(50)

Search specific fields

fields(:name, :brand)

Where

where(store_id: 1, expires_at: Time.now..)

These types of filters are supported

Order

order(_score: :desc) # most relevant first - default

All of these sort options are supported

Limit / offset

limit(20).offset(40)

Select

select(:name)

These source filtering options are supported

Results

Searches return a Searchkick::Relation object. This responds like an array to most methods.

results = Product.search("milk")
results.size
results.any?
results.each { |result| ... }

By default, ids are fetched from the search server and records are fetched from your database. To fetch everything from the search server, use:

Product.search("apples").load(false)

Get total results

results.total_count

Get the time the search took (in milliseconds)

results.took

Get the full response from the search server

results.response

Note: By default, Elasticsearch and OpenSearch limit paging to the first 10,000 results for performance. This applies to the total count as well.

Filtering

Equal

where(store_id: 1)

Not equal

where.not(store_id: 2)

Greater than (gt), less than (lt), greater than or equal (gte), less than or equal (lte)

where(expires_at: {gt: Time.now})

Range

where(orders_count: 1..10)

In

where(aisle_id: [25, 30])

Not in

where.not(aisle_id: [25, 30])

Contains all

where(user_ids: {all: [1, 3]})

Like

where(category: {like: "%frozen%"})

Case-insensitive like

where(category: {ilike: "%frozen%"})

Regular expression

where(category: /frozen .+/)

Prefix

where(category: {prefix: "frozen"})

Exists

where(store_id: {exists: true})

Combine filters with OR

where(_or: [{in_stock: true}, {backordered: true}])

Boosting

Boost important fields

fields("title^10", "description")

Boost by the value of a field (field must be numeric)

boost_by(:orders_count) # give popular documents a little boost
boost_by(orders_count: {factor: 10}) # default factor is 1

Boost matching documents

boost_where(user_id: 1)
boost_where(user_id: {value: 1, factor: 100}) # default factor is 1000
boost_where(user_id: [{value: 1, factor: 100}, {value: 2, factor: 200}])

Boost by recency

boost_by_recency(created_at: {scale: "7d", decay: 0.5})

You can also boost by:

• Conversions
• Distance

Get Everything

Use a * for the query.

Product.search("*")

Pagination

Plays nicely with kaminari and will_paginate.

# controller
@products = Product.search("milk").page(params[:page]).per_page(20)

View with kaminari

<%= paginate @products %>

View with will_paginate

<%= will_paginate @products %>

Partial Matches

By default, results must match all words in the query.

Product.search("fresh honey") # fresh AND honey

To change this, use:

Product.search("fresh honey").operator("or") # fresh OR honey

By default, results must match the entire word - back will not match backpack. You can change this behavior with:

class Product < ApplicationRecord
  searchkick word_start: [:name]
end

And to search (after you reindex):

Product.search("back").fields(:name).match(:word_start)

Available options are:

The default is :word. The most matches will happen with :word_middle.

To specify different matching for different fields, use:

Product.search(query).fields({name: :word_start}, {brand: :word_middle})

Exact Matches

To match a field exactly (case-sensitive), use:

Product.search(query).fields({name: :exact})

Phrase Matches

To only match the exact order, use:

Product.search("fresh honey").match(:phrase)

Stemming and Language

Searchkick stems words by default for better matching. apple and apples both stem to appl, so searches for either term will have the same matches.

Searchkick defaults to English for stemming. To change this, use:

class Product < ApplicationRecord
  searchkick language: "german"
end

See the list of languages. A few languages require plugins:

• chinese - analysis-ik plugin
• chinese2 - analysis-smartcn plugin
• japanese - analysis-kuromoji plugin
• korean - analysis-openkoreantext plugin
• korean2 - analysis-nori plugin
• polish - analysis-stempel plugin
• ukrainian - analysis-ukrainian plugin
• vietnamese - analysis-vietnamese plugin

You can also use a Hunspell dictionary for stemming.

class Product < ApplicationRecord
  searchkick stemmer: {type: "hunspell", locale: "en_US"}
end

Disable stemming with:

class Image < ApplicationRecord
  searchkick stem: false
end

Exclude certain words from stemming with:

class Image < ApplicationRecord
  searchkick stem_exclusion: ["apples"]
end

Or change how words are stemmed:

class Image < ApplicationRecord
  searchkick stemmer_override: ["apples => other"]
end

Synonyms

class Product < ApplicationRecord
  searchkick search_synonyms: [["pop", "soda"], ["burger", "hamburger"]]
end

Call Product.reindex after changing synonyms. Synonyms are applied at search time before stemming, and can be a single word or multiple words.

For directional synonyms, use:

search_synonyms: ["lightbulb => halogenlamp"]

Dynamic Synonyms

The above approach works well when your synonym list is static, but in practice, this is often not the case. When you analyze search conversions, you often want to add new synonyms without a full reindex. We recommend placing synonyms in a file on the search server (in the config directory). This allows you to reload synonyms without reindexing.

pop, soda
burger, hamburger

Then use:

class Product < ApplicationRecord
  searchkick search_synonyms: "synonyms.txt"
end

And reload with:

Product.search_index.reload_synonyms

Misspellings

By default, Searchkick handles misspelled queries by returning results with an edit distance of one.

You can change this with:

Product.search("zucini").misspellings(edit_distance: 2) # zucchini

To prevent poor precision and improve performance for correctly spelled queries (which should be a majority for most applications), Searchkick can first perform a search without misspellings, and if there are too few results, perform another with them.

Product.search("zuchini").misspellings(below: 5)

If there are fewer than 5 results, a 2nd search is performed with misspellings enabled. The result of this query is returned.

Turn off misspellings with:

Product.search("zuchini").misspellings(false) # no zucchini

Specify which fields can include misspellings with:

Product.search("zucini").fields(:name, :color).misspellings(fields: [:name])

When doing this, you must also specify fields to search

Bad Matches

If a user searches butter, they may also get results for peanut butter. To prevent this, use:

Product.search("butter").exclude("peanut butter")

You can map queries and terms to exclude with:

exclude_queries = {
  "butter" => ["peanut butter"],
  "cream" => ["ice cream", "whipped cream"]
}

Product.search(query).exclude(exclude_queries[query])

You can demote results by boosting by a factor less than one:

Product.search("butter").boost_where(category: {value: "pantry", factor: 0.5})

Emoji

Search :ice_cream::cake: and get ice cream cake!

Add this line to your application’s Gemfile:

gem "gemoji-parser"

And use:

Product.search("🍨🍰").emoji

Indexing

Control what data is indexed with the search_data method. Call Product.reindex after changing this method.

class Product < ApplicationRecord
  belongs_to :department

  def search_data
    {
      name: name,
      department_name: department.name,
      on_sale: sale_price.present?
    }
  end
end

Searchkick uses find_in_batches to import documents. To eager load associations, use the search_import scope.

class Product < ApplicationRecord
  scope :search_import, -> { includes(:department) }
end

By default, all records are indexed. To control which records are indexed, use the should_index? method.

class Product < ApplicationRecord
  def should_index?
    active # only index active records
  end
end

If a reindex is interrupted, you can resume it with:

Product.reindex(resume: true)

For large data sets, try parallel reindexing.

To Reindex, or Not to Reindex

Reindex

• when you install or upgrade searchkick
• change the search_data method
• change the searchkick method

No need to reindex

• app starts

Strategies

There are four strategies for keeping the index synced with your database.

1. Inline (default)

Anytime a record is inserted, updated, or deleted

2. Asynchronous

Use background jobs for better performance

class Product < ApplicationRecord
  searchkick callbacks: :async
end

Jobs are added to a queue named searchkick.

3. Queuing

Push ids of records that need updated to a queue and reindex in the background in batches. This is more performant than the asynchronous method, which updates records individually. See how to set up.

4. Manual

Turn off automatic syncing

class Product < ApplicationRecord
  searchkick callbacks: false
end

And reindex a record or relation manually.

product.reindex
# or
store.products.reindex(mode: :async)

You can also do bulk updates.

Searchkick.callbacks(:bulk) do
  Product.find_each(&:update_fields)
end

Or temporarily skip updates.

Searchkick.callbacks(false) do
  Product.find_each(&:update_fields)
end

Or override the model’s strategy.

product.reindex(mode: :async) # :inline or :queue

Associations

Data is not automatically synced when an association is updated. If this is desired, add a callback to reindex:

class Image < ApplicationRecord
  belongs_to :product

  after_commit :reindex_product

  def reindex_product
    product.reindex
  end
end

Default Scopes

If you have a default scope that filters records, use the should_index? method to exclude them from indexing:

class Product < ApplicationRecord
  default_scope { where(deleted_at: nil) }

  def should_index?
    deleted_at.nil?
  end
end

If you want to index and search filtered records, set:

class Product < ApplicationRecord
  searchkick unscope: true
end

Intelligent Search

The best starting point to improve your search by far is to track searches and conversions. Searchjoy makes it easy.

Product.search("apple").track(user_id: current_user.id)

See the docs for how to install and use. Focus on top searches with a low conversion rate.

Searchkick can then use the conversion data to learn what users are looking for. If a user searches for “ice cream” and adds Ben & Jerry’s Chunky Monkey to the cart (our conversion metric at Instacart), that item gets a little more weight for similar searches. This can make a huge difference on the quality of your search.

Add conversion data with:

class Product < ApplicationRecord
  has_many :conversions, class_name: "Searchjoy::Conversion", as: :convertable
  has_many :searches, class_name: "Searchjoy::Search", through: :conversions

  searchkick conversions_v2: [:conversions] # name of field

  def search_data
    {
      name: name,
      conversions: searches.group(:query).distinct.count(:user_id)
      # {"ice cream" => 234, "chocolate" => 67, "cream" => 2}
    }
  end
end

Reindex and set up a cron job to add new conversions daily. For zero downtime deployment, temporarily set conversions_v2(false) in your search calls until the data is reindexed.

Performant Conversions

A performant way to do conversions is to cache them to prevent N+1 queries. For Postgres, create a migration with:

add_column :products, :search_conversions, :jsonb

For MySQL, use :json, and for others, use :text with a JSON serializer.

Next, update your model. Create a separate method for conversion data so you can use partial reindexing.

class Product < ApplicationRecord
  searchkick conversions_v2: [:conversions]

  def search_data
    {
      name: name,
      category: category
    }.merge(conversions_data)
  end

  def conversions_data
    {
      conversions: search_conversions || {}
    }
  end
end

Deploy and reindex your data. For zero downtime deployment, temporarily set conversions_v2(false) in your search calls until the data is reindexed.

Product.reindex

Then, create a job to update the conversions column and reindex records with new conversions. Here’s one you can use for Searchjoy:

class UpdateConversionsJob < ApplicationJob
  def perform(class_name, since: nil, update: true, reindex: true)
    model = Searchkick.load_model(class_name)

    # get records that have a recent conversion
    recently_converted_ids =
      Searchjoy::Conversion.where(convertable_type: class_name, created_at: since..)
        .order(:convertable_id).distinct.pluck(:convertable_id)

    # split into batches
    recently_converted_ids.in_groups_of(1000, false) do |ids|
      if update
        # fetch conversions
        conversions =
          Searchjoy::Conversion.where(convertable_id: ids, convertable_type: class_name)
            .joins(:search).where.not(searchjoy_searches: {user_id: nil})
            .group(:convertable_id, :query).distinct.count(:user_id)

        # group by record
        conversions_by_record = {}
        conversions.each do |(id, query), count|
          (conversions_by_record[id] ||= {})[query] = count
        end

        # update conversions column
        model.transaction do
          conversions_by_record.each do |id, conversions|
            model.where(id: id).update_all(search_conversions: conversions)
          end
        end
      end

      if reindex
        # reindex conversions data
        model.where(id: ids).reindex(:conversions_data, ignore_missing: true)
      end
    end
  end
end

Run the job:

UpdateConversionsJob.perform_now("Product")

And set it up to run daily.

UpdateConversionsJob.perform_later("Product", since: 1.day.ago)

Personalized Results

Order results differently for each user. For example, show a user’s previously purchased products before other results.

class Product < ApplicationRecord
  def search_data
    {
      name: name,
      orderer_ids: orders.pluck(:user_id) # boost this product for these users
    }
  end
end

Reindex and search with:

Product.search("milk").boost_where(orderer_ids: current_user.id)

Instant Search / Autocomplete

Autocomplete predicts what a user will type, making the search experience faster and easier.

Note: To autocomplete on search terms rather than results, check out Autosuggest.

Note 2: If you only have a few thousand records, don’t use Searchkick for autocomplete. It’s much faster to load all records into JavaScript and autocomplete there (eliminates network requests).

First, specify which fields use this feature. This is necessary since autocomplete can increase the index size significantly, but don’t worry - this gives you blazing fast queries.

class Movie < ApplicationRecord
  searchkick word_start: [:title, :director]
end

Reindex and search with:

Movie.search("jurassic pa").fields(:title).match(:word_start)

Use a front-end library like typeahead.js to show the results.

Here’s how to make it work with Rails

First, add a route and controller action.

class MoviesController < ApplicationController
  def autocomplete
    render json: Movie.search(params[:query]).fields("title^5", "director")
      .match(:word_start).limit(10).load(false).misspellings(below: 5).map(&:title)
  end
end

Note: Use load(false) and misspellings(below: n) (or misspellings(false)) for best performance.

Then add the search box and JavaScript code to a view.

<input type="text" id="query" name="query" />

<script src="jquery.js"></script>
<script src="typeahead.bundle.js"></script>
<script>
  var movies = new Bloodhound({
    datumTokenizer: Bloodhound.tokenizers.whitespace,
    queryTokenizer: Bloodhound.tokenizers.whitespace,
    remote: {
      url: '/movies/autocomplete?query=%QUERY',
      wildcard: '%QUERY'
    }
  });
  $('#query').typeahead(null, {
    source: movies
  });
</script>

Suggestions

class Product < ApplicationRecord
  searchkick suggest: [:name] # fields to generate suggestions
end

Reindex and search with:

products = Product.search("peantu butta").suggest
products.suggestions # ["peanut butter"]

Aggregations

Aggregations provide aggregated search data.

products = Product.search("chuck taylor").aggs(:product_type, :gender, :brand)
products.aggs

By default, where conditions apply to aggregations.

Product.search("wingtips").where(color: "brandy").aggs(:size)
# aggregations for brandy wingtips are returned

Change this with:

Product.search("wingtips").where(color: "brandy").aggs(:size).smart_aggs(false)
# aggregations for all wingtips are returned

Set where conditions for each aggregation separately with:

Product.search("wingtips").aggs(size: {where: {color: "brandy"}})

Limit

Product.search("apples").aggs(store_id: {limit: 10})

Order

Product.search("wingtips").aggs(color: {order: {"_key" => "asc"}}) # alphabetically

All of these options are supported

Ranges

price_ranges = [{to: 20}, {from: 20, to: 50}, {from: 50}]
Product.search("*").aggs(price: {ranges: price_ranges})

Minimum document count

Product.search("apples").aggs(store_id: {min_doc_count: 2})

Script support

Product.search("*").aggs(color: {script: {source: "'Color: ' + _value"}})

Date histogram

Product.search("pear").aggs(products_per_year: {date_histogram: {field: :created_at, interval: :year}})

For other aggregation types, including sub-aggregations, use body_options:

Product.search("orange").body_options(aggs: {price: {histogram: {field: :price, interval: 10}}})

Highlight

Specify which fields to index with highlighting.

class Band < ApplicationRecord
  searchkick highlight: [:name]
end

Highlight the search query in the results.

bands = Band.search("cinema").highlight

View the highlighted fields with:

bands.with_highlights.each do |band, highlights|
  highlights[:name] # "Two Door <em>Cinema</em> Club"
end

To change the tag, use:

Band.search("cinema").highlight(tag: "<strong>")

To highlight and search different fields, use:

Band.search("cinema").fields(:name).highlight(fields: [:description])

By default, the entire field is highlighted. To get small snippets instead, use:

bands = Band.search("cinema").highlight(fragment_size: 20)
bands.with_highlights(multiple: true).each do |band, highlights|
  highlights[:name].join(" and ")
end

Additional options can be specified for each field:

Band.search("cinema").fields(:name).highlight(fields: {name: {fragment_size: 200}})

You can find available highlight options in the Elasticsearch or OpenSearch reference.

Similar Items

Find similar items

product = Product.first
product.similar.fields(:name).where(size: "12 oz")

Geospatial Searches

class Restaurant < ApplicationRecord
  searchkick locations: [:location]

  def search_data
    attributes.merge(location: {lat: latitude, lon: longitude})
  end
end

Reindex and search with:

Restaurant.search("pizza").where(location: {near: {lat: 37, lon: -114}, within: "100mi"}) # or 160km

Bounded by a box

Restaurant.search("sushi").where(location: {top_left: {lat: 38, lon: -123}, bottom_right: {lat: 37, lon: -122}})

Note: top_right and bottom_left also work

Bounded by a polygon

Restaurant.search("dessert").where(location: {geo_polygon: {points: [{lat: 38, lon: -123}, {lat: 39, lon: -123}, {lat: 37, lon: 122}]}})

Boost By Distance

Boost results by distance - closer results are boosted more

Restaurant.search("noodles").boost_by_distance(location: {origin: {lat: 37, lon: -122}})

Also supports additional options

Restaurant.search("wings").boost_by_distance(location: {origin: {lat: 37, lon: -122}, function: "linear", scale: "30mi", decay: 0.5})

Geo Shapes

You can also index and search geo shapes.

class Restaurant < ApplicationRecord
  searchkick geo_shape: [:bounds]

  def search_data
    attributes.merge(
      bounds: {
        type: "envelope",
        coordinates: [{lat: 4, lon: 1}, {lat: 2, lon: 3}]
      }
    )
  end
end

See the Elasticsearch documentation for details.

Find shapes intersecting with the query shape

Restaurant.search("soup").where(bounds: {geo_shape: {type: "polygon", coordinates: [[{lat: 38, lon: -123}, ...]]}})

Falling entirely within the query shape

Restaurant.search("salad").where(bounds: {geo_shape: {type: "circle", relation: "within", coordinates: {lat: 38, lon: -123}, radius: "1km"}})

Not touching the query shape

Restaurant.search("burger").where(bounds: {geo_shape: {type: "envelope", relation: "disjoint", coordinates: [{lat: 38, lon: -123}, {lat: 37, lon: -122}]}})

Inheritance

Searchkick supports single table inheritance.

class Dog < Animal
end

In your parent model, set:

class Animal < ApplicationRecord
  searchkick inheritance: true
end

The parent and child model can both reindex.

Animal.reindex
Dog.reindex # equivalent, all animals reindexed

And to search, use:

Animal.search("*")                # all animals
Dog.search("*")                   # just dogs
Animal.search("*").type(Cat, Dog) # just cats and dogs

Notes:

1. The suggest option retrieves suggestions from the parent at the moment.

   Dog.search("airbudd").suggest # suggestions for all animals
   

2. This relies on a type field that is automatically added to the indexed document. Be wary of defining your own type field in search_data, as it will take precedence.

Debugging Queries

To help with debugging queries, you can use:

Product.search("soap").debug

This prints useful info to stdout.

See how the search server scores your queries with:

Product.search("soap").explain.response

See how the search server tokenizes your queries with:

Product.search_index.tokens("Dish Washer Soap", analyzer: "searchkick_index")
# ["dish", "dishwash", "washer", "washersoap", "soap"]

Product.search_index.tokens("dishwasher soap", analyzer: "searchkick_search")
# ["dishwashersoap"] - no match

Product.search_index.tokens("dishwasher soap", analyzer: "searchkick_search2")
# ["dishwash", "soap"] - match!!

Partial matches

Product.search_index.tokens("San Diego", analyzer: "searchkick_word_start_index")
# ["s", "sa", "san", "d", "di", "die", "dieg", "diego"]

Product.search_index.tokens("dieg", analyzer: "searchkick_word_search")
# ["dieg"] - match!!

See the complete list of analyzers.

Testing

As you iterate on your search, it’s a good idea to add tests.

For performance, only enable Searchkick callbacks for the tests that need it.

Rails

Add to your test/test_helper.rb:

module ActiveSupport
  class TestCase
    parallelize_setup do |worker|
      Searchkick.index_suffix = worker

      # reindex models for parallel tests
      Product.reindex
    end
  end
end

# reindex models for non-parallel tests
Product.reindex

# and disable callbacks
Searchkick.disable_callbacks

And use:

class ProductTest < ActiveSupport::TestCase
  setup do
    Searchkick.enable_callbacks
  end

  teardown do
    Searchkick.disable_callbacks
  end

  test "search" do
    Product.create!(name: "Apple")
    Product.search_index.refresh
    assert_equal ["Apple"], Product.search("apple").map(&:name)
  end
end

Minitest

Add to your test/test_helper.rb:

# reindex models
Product.reindex

# and disable callbacks
Searchkick.disable_callbacks

And use:

class ProductTest < Minitest::Test
  def setup
    Searchkick.enable_callbacks
  end

  def teardown
    Searchkick.disable_callbacks
  end

  def test_search
    Product.create!(name: "Apple")
    Product.search_index.refresh
    assert_equal ["Apple"], Product.search("apple").map(&:name)
  end
end

RSpec

Add to your spec/spec_helper.rb:

RSpec.configure do |config|
  config.before(:suite) do
    # reindex models
    Product.reindex

    # and disable callbacks
    Searchkick.disable_callbacks
  end

  config.around(:each, search: true) do |example|
    Searchkick.callbacks(nil) do
      example.run
    end
  end
end

And use:

describe Product, search: true do
  it "searches" do
    Product.create!(name: "Apple")
    Product.search_index.refresh
    assert_equal ["Apple"], Product.search("apple").map(&:name)
  end
end

Factory Bot

Define a trait for each model:

FactoryBot.define do
  factory :product do
    trait :reindex do
      after(:create) do |product, _|
        product.reindex(refresh: true)
      end
    end
  end
end

And use:

FactoryBot.create(:product, :reindex)

GitHub Actions

Check out setup-elasticsearch for an easy way to install Elasticsearch:

    - uses: ankane/setup-elasticsearch@v1

And setup-opensearch for an easy way to install OpenSearch:

    - uses: ankane/setup-opensearch@v1

Deployment

For the search server, Searchkick uses ENV["ELASTICSEARCH_URL"] for Elasticsearch and ENV["OPENSEARCH_URL"] for OpenSearch. This defaults to http://localhost:9200.

• Elastic Cloud
• Amazon OpenSearch Service
• Heroku
• Self-Hosted and Other

Elastic Cloud

Create an initializer config/initializers/elasticsearch.rb with:

ENV["ELASTICSEARCH_URL"] = "https://user:password@host:port"

Then deploy and reindex:

rake searchkick:reindex:all

Amazon OpenSearch Service

Create an initializer config/initializers/opensearch.rb with:

ENV["OPENSEARCH_URL"] = "https://es-domain-1234.us-east-1.es.amazonaws.com:443"

To use signed requests, include in your Gemfile:

gem "faraday_middleware-aws-sigv4"

and add to your initializer:

Searchkick.aws_credentials = {
  access_key_id: ENV["AWS_ACCESS_KEY_ID"],
  secret_access_key: ENV["AWS_SECRET_ACCESS_KEY"],
  region: "us-east-1"
}

Then deploy and reindex:

rake searchkick:reindex:all

Heroku

Choose an add-on: Bonsai, SearchBox, or Elastic Cloud.

For Elasticsearch on Bonsai:

heroku addons:create bonsai
heroku config:set ELASTICSEARCH_URL=`heroku config:get BONSAI_URL`

For OpenSearch on Bonsai:

heroku addons:create bonsai --engine=opensearch
heroku config:set OPENSEARCH_URL=`heroku config:get BONSAI_URL`

For SearchBox:

heroku addons:create searchbox:starter
heroku config:set ELASTICSEARCH_URL=`heroku config:get SEARCHBOX_URL`

For Elastic Cloud (previously Found):

heroku addons:create foundelasticsearch
heroku addons:open foundelasticsearch

Visit the Shield page and reset your password. You’ll need to add the username and password to your url. Get the existing url with:

heroku config:get FOUNDELASTICSEARCH_URL

And add elastic:password@ right after https:// and add port 9243 at the end:

heroku config:set ELASTICSEARCH_URL=https://elastic:[email protected]:9243

Then deploy and reindex:

heroku run rake searchkick:reindex:all

Self-Hosted and Other

Create an initializer with:

ENV["ELASTICSEARCH_URL"] = "https://user:password@host:port"
# or
ENV["OPENSEARCH_URL"] = "https://user:password@host:port"

Then deploy and reindex:

rake searchkick:reindex:all

Data Protection

We recommend encrypting data at rest and in transit (even inside your own network). This is especially important if you send personal data of your users to the search server.

Bonsai, Elastic Cloud, and Amazon OpenSearch Service all support encryption at rest and HTTPS.

Automatic Failover

Create an initializer with multiple hosts:

ENV["ELASTICSEARCH_URL"] = "https://user:password@host1,https://user:password@host2"
# or
ENV["OPENSEARCH_URL"] = "https://user:password@host1,https://user:password@host2"

Client Options

Create an initializer with:

Searchkick.client_options[:reload_connections] = true

See the docs for Elasticsearch or Opensearch for a complete list of options.

Lograge

Add the following to config/environments/production.rb:

config.lograge.custom_options = lambda do |event|
  options = {}
  options[:search] = event.payload[:searchkick_runtime] if event.payload[:searchkick_runtime].to_f > 0
  options
end

See Production Rails for other good practices.

Performance

Persistent HTTP Connections

Significantly increase performance with persistent HTTP connections. Add Typhoeus to your Gemfile and it’ll automatically be used.

gem "typhoeus"

To reduce log noise, create an initializer with:

Ethon.logger = Logger.new(nil)

Searchable Fields

By default, all string fields are searchable (can be used in fields option). Speed up indexing and reduce index size by only making some fields searchable.

class Product < ApplicationRecord
  searchkick searchable: [:name]
end

Filterable Fields

By default, all string fields are filterable (can be used in where option). Speed up indexing and reduce index size by only making some fields filterable.

class Product < ApplicationRecord
  searchkick filterable: [:brand]
end

Note: Non-string fields are always filterable and should not be passed to this option.

Parallel Reindexing

For large data sets, you can use background jobs to parallelize reindexing.

Product.reindex(mode: :async)
# {index_name: "products_production_20260111210018065"}

Once the jobs complete, promote the new index with:

Product.search_index.promote(index_name)

You can optionally track the status with Redis:

Searchkick.redis = Redis.new

And use:

Searchkick.reindex_status(index_name)

You can also have Searchkick wait for reindexing to complete

Product.reindex(mode: :async, wait: true)

You can use your background job framework to control concurrency. For Solid Queue, create an initializer with:

module SearchkickBulkReindexConcurrency
  extend ActiveSupport::Concern

  included do
    limits_concurrency to: 3, key: ""
  end
end

Rails.application.config.after_initialize do
  Searchkick::BulkReindexJob.include(SearchkickBulkReindexConcurrency)
end

This will allow only 3 jobs to run at once.

Refresh Interval

You can specify a longer refresh interval while reindexing to increase performance.

Product.reindex(mode: :async, refresh_interval: "30s")

Note: This only makes a noticeable difference with parallel reindexing.

When promoting, have it restored to the value in your mapping (defaults to 1s).

Product.search_index.promote(index_name, update_refresh_interval: true)

Queuing

Push ids of records needing reindexing to a queue and reindex in bulk for better performance. First, set up Redis in an initializer. We recommend using connection_pool.

Searchkick.redis = ConnectionPool.new { Redis.new }

And ask your models to queue updates.

class Product < ApplicationRecord
  searchkick callbacks: :queue
end

Then, set up a background job to run.

Searchkick::ProcessQueueJob.perform_later(class_name: "Product")

You can check the queue length with:

Product.search_index.reindex_queue.length

For more tips, check out Keeping Elasticsearch in Sync.

Routing

Searchkick supports routing, which can significantly speed up searches.

class Business < ApplicationRecord
  searchkick routing: true

  def search_routing
    city_id
  end
end

Reindex and search with:

Business.search("ice cream").routing(params[:city_id])

Partial Reindexing

Reindex a subset of attributes to reduce time spent generating search data and cut down on network traffic.

class Product < ApplicationRecord
  def search_data
    {
      name: name,
      category: category
    }.merge(prices_data)
  end

  def prices_data
    {
      price: price,
      sale_price: sale_price
    }
  end
end

And use:

Product.reindex(:prices_data)

Ignore errors for missing documents with:

Product.reindex(:prices_data, ignore_missing: true)

Advanced

Searchkick makes it easy to use the Elasticsearch or OpenSearch DSL on its own.

Advanced Mapping

Create a custom mapping:

class Product < ApplicationRecord
  searchkick mappings: {
    properties: {
      name: {type: "keyword"}
    }
  }
end

Note: If you use a custom mapping, you'll need to use custom searching as well.

To keep the mappings and settings generated by Searchkick, use:

class Product < ApplicationRecord
  searchkick merge_mappings: true, mappings: {...}
end

Advanced Search

And use the body option to search:

products = Product.search.body(query: {match: {name: "milk"}})

View the response with:

products.response

To modify the query generated by Searchkick, use:

products = Product.search("milk").body_options(min_score: 1)

or

products =
  Product.search("apples") do |body|
    body[:min_score] = 1
  end

Client

To access the Elasticsearch::Client or OpenSearch::Client directly, use:

Searchkick.client

Multi Search

To batch search requests for performance, use:

products = Product.search("snacks")
coupons = Coupon.search("snacks")
Searchkick.multi_search([products, coupons])

Then use products and coupons as typical results.

Note: Errors are not raised as with single requests. Use the error method on each query to check for errors.

Multiple Models

Search across multiple models with:

Searchkick.search("milk").models(Product, Category)

Boost specific models with:

indices_boost(Category => 2, Product => 1)

Multi-Tenancy

Check out this great post on the Apartment gem. Follow a similar pattern if you use another gem.

Scroll API

Searchkick also supports the scroll API. Scrolling is not intended for real time user requests, but rather for processing large amounts of data.

Product.search("*").scroll("1m") do |batch|
  # process batch ...
end

You can also scroll batches manually.

products = Product.search("*").scroll("1m")
while products.any?
  # process batch ...

  products = products.scroll
end

products.clear_scroll

Deep Paging

By default, Elasticsearch and OpenSearch limit paging to the first 10,000 results. Here’s why. We don’t recommend changing this, but if you really need all results, you can use:

class Product < ApplicationRecord
  searchkick deep_paging: true
end

If you just need an accurate total count, you can instead use:

Product.search("pears").body_options(track_total_hits: true)

Nested Data

To query nested data, use dot notation.

Product.search("san").fields("store.city").where("store.zip_code" => 12345)

Nearest Neighbor Search

Available for Elasticsearch 8.6+ and OpenSearch 2.4+

class Product < ApplicationRecord
  searchkick knn: {embedding: {dimensions: 3, distance: "cosine"}}
end

Also supports euclidean and inner_product

Reindex and search with:

Product.search.knn(field: :embedding, vector: [1, 2, 3]).limit(10)

HNSW Options

Nearest neighbor search uses HNSW for indexing.

Specify m and ef_construction

class Product < ApplicationRecord
  searchkick knn: {embedding: {dimensions: 3, distance: "cosine", m: 16, ef_construction: 100}}
end

Specify ef_search

Product.search.knn(field: :embedding, vector: [1, 2, 3], ef_search: 40).limit(10)

Semantic Search

First, add nearest neighbor search to your model

class Product < ApplicationRecord
  searchkick knn: {embedding: {dimensions: 768, distance: "cosine"}}
end

Generate an embedding for each record (you can use an external service or a library like Informers)

embed = Informers.pipeline("embedding", "Snowflake/snowflake-arctic-embed-m-v1.5")
embed_options = {model_output: "sentence_embedding", pooling: "none"} # specific to embedding model

Product.find_each do |product|
  embedding = embed.(product.name, **embed_options)
  product.update!(embedding: embedding)
end

For search, generate an embedding for the query (the query prefix is specific to the embedding model)

query_prefix = "Represent this sentence for searching relevant passages: "
query_embedding = embed.(query_prefix + query, **embed_options)

And perform nearest neighbor search

Product.search.knn(field: :embedding, vector: query_embedding).limit(20)

See a full example

Hybrid Search

Perform keyword search and semantic search in parallel

keyword_search = Product.search(query).limit(20)
semantic_search = Product.search.knn(field: :embedding, vector: query_embedding).limit(20)
Searchkick.multi_search([keyword_search, semantic_search])

To combine the results, use Reciprocal Rank Fusion (RRF)

Searchkick::Reranking.rrf(keyword_search, semantic_search).first(5)

Or a reranking model

rerank = Informers.pipeline("reranking", "mixedbread-ai/mxbai-rerank-xsmall-v1")
results = (keyword_search.to_a + semantic_search.to_a).uniq
rerank.(query, results.map(&:name)).first(5).map { |v| results[v[:doc_id]] }

See a full example

Reference

Reindex one record

product = Product.find(1)
product.reindex

Reindex multiple records

Product.where(store_id: 1).reindex

Reindex associations

store.products.reindex

Remove old indices

Product.search_index.clean_indices

Use custom settings

class Product < ApplicationRecord
  searchkick settings: {number_of_shards: 3}
end

Use a different index name

class Product < ApplicationRecord
  searchkick index_name: "products_v2"
end

Use a dynamic index name

class Product < ApplicationRecord
  searchkick index_name: -> { "#{name.tableize}-#{I18n.locale}" }
end

Prefix the index name

class Product < ApplicationRecord
  searchkick index_prefix: "datakick"
end

For all models

Searchkick.index_prefix = "datakick"

Use a different term for boosting by conversions

Product.search("banana").conversions_v2(term: "organic banana")

Define multiple conversion fields

class Product < ApplicationRecord
  has_many :searches, class_name: "Searchjoy::Search"

  searchkick conversions_v2: ["unique_conversions", "total_conversions"]

  def search_data
    {
      name: name,
      unique_conversions: searches.group(:query).distinct.count(:user_id),
      total_conversions: searches.group(:query).count
    }
  end
end

And specify which to use

Product.search("banana") # boost by both fields (default)
Product.search("banana").conversions_v2("total_conversions") # only boost by total_conversions
Product.search("banana").conversions_v2(false) # no conversion boosting

Change timeout

Searchkick.timeout = 15 # defaults to 10

Set a lower timeout for searches

Searchkick.search_timeout = 3

Change the search method name

Searchkick.search_method_name = :lookup

Change the queue name

Searchkick.queue_name = :search_reindex # defaults to :searchkick

Change the queue name or priority for a model

class Product < ApplicationRecord
  searchkick job_options: {queue: "critical", priority: 10}
end

Change the queue name or priority for a specific call

Product.reindex(mode: :async, job_options: {queue: "critical", priority: 10})

Change the parent job

Searchkick.parent_job = "ApplicationJob" # defaults to "ActiveJob::Base"

Eager load associations

Product.search("milk").includes(:brand, :stores)

Eager load different associations by model

Searchkick.search("*").models(Product, Store).model_includes(Product => [:store], Store => [:product])

Run additional scopes on results

Product.search("milk").scope_results(->(r) { r.with_attached_images })

Set opaque id for slow logs

Product.search("milk").opaque_id("some-id")
# or
Searchkick.multi_search(searches, opaque_id: "some-id")

Specify default fields to search

class Product < ApplicationRecord
  searchkick default_fields: [:name]
end

Turn off special characters

class Product < ApplicationRecord
  # A will not match Ä
  searchkick special_characters: false
end

Turn on stemming for conversions

class Product < ApplicationRecord
  searchkick stem_conversions: true
end

Make search case-sensitive

class Product < ApplicationRecord
  searchkick case_sensitive: true
end

Note: If misspellings are enabled (default), results with a single character case difference will match. Turn off misspellings if this is not desired.

Change import batch size

class Product < ApplicationRecord
  searchkick batch_size: 200 # defaults to 1000
end

Create index without importing

Product.reindex(import: false)

Use a different id

class Product < ApplicationRecord
  def search_document_id
    custom_id
  end
end

Add request parameters like search_type

Product.search("carrots").request_params(search_type: "dfs_query_then_fetch")

Set options across all models

Searchkick.model_options = {
  batch_size: 200
}

Reindex conditionally

class Product < ApplicationRecord
  searchkick callback_options: {if: :search_data_changed?}

  def search_data_changed?
    previous_changes.include?("name")
  end
end

Reindex all models - Rails only

rake searchkick:reindex:all

Turn on misspellings after a certain number of characters

Product.search("api").misspellings(prefix_length: 2) # api, apt, no ahi

BigDecimal values are indexed as floats by default so they can be used for boosting. Convert them to strings to keep full precision.

class Product < ApplicationRecord
  def search_data
    {
      units: units.to_s("F")
    }
  end
end

Gotchas

Consistency

Elasticsearch and OpenSearch are eventually consistent, meaning it can take up to a second for a change to reflect in search. You can use the refresh method to have it show up immediately.

product.save!
Product.search_index.refresh

Inconsistent Scores

Due to the distributed nature of Elasticsearch and OpenSearch, you can get incorrect results when the number of documents in the index is low. You can read more about it here. To fix this, do:

class Product < ApplicationRecord
  searchkick settings: {number_of_shards: 1}
end

For convenience, this is set by default in the test environment.

Upgrading

6.0

Searchkick 6 brings a new query builder API:

Product.search("apples").where(in_stock: true).limit(10).offset(50)

All existing options can be used as methods, or you can continue to use the existing API.

This release also significantly improves the performance of searches when using conversions. To upgrade conversions without downtime, add conversions_v2 to your model and an additional field to search_data:

class Product < ApplicationRecord
  searchkick conversions: [:conversions], conversions_v2: [:conversions_v2]

  def search_data
    conversions = searches.group(:query).distinct.count(:user_id)
    {
      conversions: conversions,
      conversions_v2: conversions
    }
  end
end

Reindex, then remove conversions:

class Product < ApplicationRecord
  searchkick conversions_v2: [:conversions_v2]

  def search_data
    {
      conversions_v2: searches.group(:query).distinct.count(:user_id)
    }
  end
end

Other improvements include the option to ignore errors for missing documents with partial reindexing and more customization for background jobs. Check out the changelog for the full list of changes.

History

View the changelog

Thanks

Thanks to Karel Minarik for Elasticsearch Ruby and Tire, Jaroslav Kalistsuk for zero downtime reindexing, and Alex Leschenko for Elasticsearch autocomplete.

Contributing

Everyone is encouraged to help improve this project. Here are a few ways you can help:

• Report bugs
• Fix bugs and submit pull requests
• Write, clarify, or fix documentation
• Suggest or add new features

To get started with development:

git clone https://github.com/ankane/searchkick.git
cd searchkick
bundle install
bundle exec rake test

Feel free to open an issue to get feedback on your idea before spending too much time on it.