feat: Use embeddings in help center search (#9227)

app/controllers/public/api/v1/portals/articles_controller.rb

@@ -7,7 +7,7 @@ class Public::Api::V1::Portals::ArticlesController < Public::Api::V1::Portals::B
 
   def index
     @articles = @portal.articles
-    @articles = @articles.search(list_params) if list_params.present?
+    search_articles
     order_by_sort_param
     @articles.page(list_params[:page]) if list_params[:page].present?
   end
@@ -16,6 +16,10 @@ class Public::Api::V1::Portals::ArticlesController < Public::Api::V1::Portals::B
 
   private
 
+  def search_articles
+    @articles = @articles.search(list_params) if list_params.present?
+  end
+
   def order_by_sort_param
     @articles = if list_params[:sort].present? && list_params[:sort] == 'views'
                   @articles.order_by_views
@@ -51,3 +55,5 @@ class Public::Api::V1::Portals::ArticlesController < Public::Api::V1::Portals::B
     ChatwootMarkdownRenderer.new(content).render_article
   end
 end
+
+Public::Api::V1::Portals::ArticlesController.prepend_mod_with('Public::Api::V1::Portals::ArticlesController')

app/models/article.rb

@@ -170,3 +170,4 @@ class Article < ApplicationRecord
     self.slug ||= "#{Time.now.utc.to_i}-#{title.underscore.parameterize(separator: '-')}" if title.present?
   end
 end
+Article.include_mod_with('Concerns::Article')

config/features.yml

@@ -80,3 +80,6 @@
 - name: sla
   enabled: false
   premium: true
+- name: help_center_embedding_search
+  enabled: false
+  premium: true

config/initializers/monkey_patches/schema_dumper.rb

@@ -34,3 +34,4 @@ ActiveRecord::SchemaDumper.ignore_tables << 'responses'
 ActiveRecord::SchemaDumper.ignore_tables << 'response_sources'
 ActiveRecord::SchemaDumper.ignore_tables << 'response_documents'
 ActiveRecord::SchemaDumper.ignore_tables << 'inbox_response_sources'
+ActiveRecord::SchemaDumper.ignore_tables << 'article_embeddings'

enterprise/app/controllers/enterprise/public/api/v1/portals/articles_controller.rb

@@ -0,0 +1,11 @@
+module Enterprise::Public::Api::V1::Portals::ArticlesController
+  private
+
+  def search_articles
+    if @portal.account.feature_enabled?('help_center_embedding_search')
+      @articles = @articles.vector_search(list_params) if list_params.present?
+    else
+      super
+    end
+  end
+end

enterprise/app/models/article_embedding.rb

@@ -0,0 +1,23 @@
+# == Schema Information
+#
+# Table name: article_embeddings
+#
+#  id         :bigint           not null, primary key
+#  embedding  :vector(1536)
+#  term       :text             not null
+#  created_at :datetime         not null
+#  updated_at :datetime         not null
+#  article_id :bigint           not null
+#
+class ArticleEmbedding < ApplicationRecord
+  belongs_to :article
+  has_neighbors :embedding, normalize: true
+
+  before_save :update_response_embedding
+
+  private
+
+  def update_response_embedding
+    self.embedding = Openai::EmbeddingsService.new.get_embedding(term, 'text-embedding-3-small')
+  end
+end

enterprise/app/models/enterprise/concerns/article.rb

@@ -0,0 +1,71 @@
+module Enterprise::Concerns::Article
+  extend ActiveSupport::Concern
+
+  included do
+    after_save :add_article_embedding, if: -> { saved_change_to_title? || saved_change_to_description? || saved_change_to_content? }
+
+    def self.add_article_embedding_association
+      has_many :article_embeddings, dependent: :destroy_async
+    end
+
+    add_article_embedding_association if Features::HelpcenterEmbeddingSearchService.new.feature_enabled?
+
+    def self.vector_search(params)
+      embedding = Openai::EmbeddingsService.new.get_embedding(params['query'], 'text-embedding-3-small')
+      records = joins(
+        :category
+      ).search_by_category_slug(
+        params[:category_slug]
+      ).search_by_category_locale(params[:locale]).search_by_author(params[:author_id]).search_by_status(params[:status])
+      filtered_article_ids = records.pluck(:id)
+
+      # Fetch nearest neighbors and their distances, then filter directly
+
+      # experimenting with filtering results based on result threshold
+      # distance_threshold = 0.2
+      # if using add the filter block to the below query
+      # .filter { |ae| ae.neighbor_distance <= distance_threshold }
+
+      article_ids = ArticleEmbedding.where(article_id: filtered_article_ids)
+                                    .nearest_neighbors(:embedding, embedding, distance: 'cosine')
+                                    .limit(5)
+                                    .pluck(:article_id)
+
+      # Fetch the articles by the IDs obtained from the nearest neighbors search
+      where(id: article_ids)
+    end
+  end
+
+  def add_article_embedding
+    return unless account.feature_enabled?('help_center_embedding_search')
+
+    terms = generate_article_search_terms
+    article_embeddings.destroy_all
+    terms.each { |term| article_embeddings.create!(term: term) }
+  end
+
+  def article_to_search_terms_prompt
+    <<~SYSTEM_PROMPT_MESSAGE
+      For the provided article content, generate potential search query keywords and snippets that can be used to generate the embeddings.
+      Ensure the search terms are as diverse as possible but capture the essence of the article and are super related to the articles.
+      Don't return any terms if there aren't any terms of relevance.
+      Always return results in valid JSON of the following format
+      {
+        "search_terms": []
+      }
+    SYSTEM_PROMPT_MESSAGE
+  end
+
+  def generate_article_search_terms
+    messages = [
+      { role: 'system', content: article_to_search_terms_prompt },
+      { role: 'user', content: "title: #{title} \n description: #{description} \n content: #{content}" }
+    ]
+    headers = { 'Content-Type' => 'application/json', 'Authorization' => "Bearer #{ENV.fetch('OPENAI_API_KEY', nil)}" }
+    body = { model: 'gpt-4-turbo', messages: messages, response_format: { type: 'json_object' } }.to_json
+    Rails.logger.info "Requesting Chat GPT with body: #{body}"
+    response = HTTParty.post('https://api.openai.com/v1/chat/completions', headers: headers, body: body)
+    Rails.logger.info "Chat GPT response: #{response.body}"
+    JSON.parse(response.parsed_response['choices'][0]['message']['content'])['search_terms']
+  end
+end

enterprise/app/services/features/base_service.rb

@@ -0,0 +1,7 @@
+class Features::BaseService
+  MIGRATION_VERSION = ActiveRecord::Migration[7.0]
+
+  def vector_extension_enabled?
+    ActiveRecord::Base.connection.extension_enabled?('vector')
+  end
+end

enterprise/app/services/features/helpcenter_embedding_search_service.rb

@@ -0,0 +1,42 @@
+# ensure vector extension is enabled via response bot service
+class Features::HelpcenterEmbeddingSearchService < Features::BaseService
+  def enable_in_installation
+    create_tables
+  end
+
+  def disable_in_installation
+    drop_tables
+  end
+
+  def feature_enabled?
+    vector_extension_enabled? && MIGRATION_VERSION.table_exists?(:article_embeddings)
+  end
+
+  def create_tables
+    return unless vector_extension_enabled?
+
+    %i[article_embeddings].each do |table|
+      send("create_#{table}_table")
+    end
+  end
+
+  def drop_tables
+    %i[article_embeddings].each do |table|
+      MIGRATION_VERSION.drop_table table if MIGRATION_VERSION.table_exists?(table)
+    end
+  end
+
+  private
+
+  def create_article_embeddings_table
+    return if MIGRATION_VERSION.table_exists?(:article_embeddings)
+
+    MIGRATION_VERSION.create_table :article_embeddings do |t|
+      t.bigint :article_id, null: false
+      t.text :term, null: false
+      t.vector :embedding, limit: 1536
+      t.timestamps
+    end
+    MIGRATION_VERSION.add_index :article_embeddingsk, :embedding, using: :ivfflat, opclass: :vector_l2_ops
+  end
+end

enterprise/app/services/features/response_bot_service.rb

@@ -1,6 +1,4 @@
-class Features::ResponseBotService
+class Features::ResponseBotService < Features::BaseService
-  MIGRATION_VERSION = ActiveRecord::Migration[7.0]
-
   def enable_in_installation
     enable_vector_extension
     create_tables
@@ -21,10 +19,6 @@ class Features::ResponseBotService
     MIGRATION_VERSION.disable_extension 'vector'
   end
 
-  def vector_extension_enabled?
-    ActiveRecord::Base.connection.extension_enabled?('vector')
-  end
-
   def create_tables
     return unless vector_extension_enabled?
 

enterprise/app/services/openai/embeddings_service.rb

@@ -1,11 +1,11 @@
 class Openai::EmbeddingsService
-  def get_embedding(content)
+  def get_embedding(content, model = 'text-embedding-ada-002')
-    fetch_embeddings(content)
+    fetch_embeddings(content, model)
   end
 
   private
 
-  def fetch_embeddings(input)
+  def fetch_embeddings(input, model)
     url = 'https://api.openai.com/v1/embeddings'
     headers = {
       'Authorization' => "Bearer #{ENV.fetch('OPENAI_API_KEY', '')}",
@@ -13,7 +13,7 @@ class Openai::EmbeddingsService
     }
     data = {
       input: input,
-      model: 'text-embedding-ada-002'
+      model: model
     }
 
     response = Net::HTTP.post(URI(url), data.to_json, headers)

spec/enterprise/controllers/enterprise/public/api/v1/portals/articles_controller_spec.rb

@@ -0,0 +1,19 @@
+require 'rails_helper'
+
+RSpec.describe 'Public Articles API', type: :request do
+  let!(:portal) { create(:portal, slug: 'test-portal', config: { allowed_locales: %w[en es] }, custom_domain: 'www.example.com') }
+
+  describe 'GET /public/api/v1/portals/:slug/articles' do
+    before do
+      portal.account.enable_features!(:help_center_embedding_search)
+    end
+
+    context 'with help_center_embedding_search feature' do
+      it 'get all articles with searched text query using vector search if enabled' do
+        allow(Article).to receive(:vector_search)
+        get "/hc/#{portal.slug}/en/articles.json", params: { query: 'funny' }
+        expect(Article).to have_received(:vector_search)
+      end
+    end
+  end
+end