feat(articles): use Responses API for search terms

config/llm.yml

@@ -183,8 +183,8 @@ features:
     default: gpt-5.6-terra
     reasoning_effort: medium
   pdf_faq_generation:
-    models: [gpt-4.1-mini, gpt-5-mini, gpt-4.1, gpt-5.1, gpt-5.2]
-    default: gpt-4.1-mini
+    models: [gpt-4.1-mini, gpt-5-mini, gpt-4.1, gpt-5.1, gpt-5.2, gpt-5.6-luna, gpt-5.6-terra, gpt-5.6-sol]
+    default: gpt-5.6-terra
     reasoning_effort: medium
   help_center_article_generation:
     models:

enterprise/app/models/concerns/agentable.rb

@@ -10,7 +10,8 @@ module Concerns::Agentable
       tools: agent_tools,
       model: agent_model,
       temperature: temperature.presence&.to_f || DEFAULT_TEMPERATURE,
-      response_schema: agent_response_schema
+      response_schema: agent_response_schema,
+      params: agent_params
     )
   end
 
@@ -52,6 +53,13 @@ module Concerns::Agentable
     installation_model.presence || route[:model]
   end
 
+  def agent_params
+    route = Llm::FeatureRouter.resolve(feature: 'assistant', account: account)
+    return {} if route[:reasoning_effort].blank?
+
+    { reasoning_effort: route[:reasoning_effort] }
+  end
+
   def installation_model
     InstallationConfig.find_by(name: 'CAPTAIN_OPEN_AI_MODEL')&.value
   end

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

@@ -1,6 +1,23 @@
 module Enterprise::Concerns::Article
   extend ActiveSupport::Concern
 
+  SEARCH_TERMS_FEATURE = 'help_center_article_generation'
+  SEARCH_TERMS_SCHEMA = {
+    name: 'article_search_terms',
+    schema: {
+      type: 'object',
+      properties: {
+        search_terms: {
+          type: 'array',
+          items: { type: 'string' }
+        }
+      },
+      required: %w[search_terms],
+      additionalProperties: false
+    },
+    strict: true
+  }.freeze
+
   included do
     after_save :add_article_embedding, if: -> { saved_change_to_title? || saved_change_to_description? || saved_change_to_content? }
 
@@ -67,23 +84,38 @@ module Enterprise::Concerns::Article
       { 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 #{openai_api_key}" }
-    body = { model: 'gpt-4o', messages: messages, response_format: { type: 'json_object' } }.to_json
-    Rails.logger.info "Requesting Chat GPT with body: #{body}"
-    response = HTTParty.post(openai_api_url, headers: headers, body: body)
-    Rails.logger.info "Chat GPT response: #{response.body}"
-    JSON.parse(response.parsed_response['choices'][0]['message']['content'])['search_terms']
+
+    response = responses_client.create(
+      model: search_terms_route[:model],
+      messages: messages,
+      schema: SEARCH_TERMS_SCHEMA,
+      reasoning_effort: search_terms_route[:reasoning_effort],
+      metadata: {
+        account_id: account_id,
+        article_id: id,
+        feature: 'article_search_terms'
+      }
+    )
+
+    JSON.parse(response[:message])['search_terms']
   end
 
   private
 
+  def search_terms_route
+    @search_terms_route ||= Llm::FeatureRouter.resolve(feature: SEARCH_TERMS_FEATURE, account: account)
+  end
+
+  def responses_client
+    @responses_client ||= Llm::ResponsesClient.new(api_key: openai_api_key, api_base: openai_api_base)
+  end
+
   def openai_api_key
     InstallationConfig.find_by(name: 'CAPTAIN_OPEN_AI_API_KEY')&.value.presence || raise(I18n.t('captain.api_key_missing'))
   end
 
-  def openai_api_url
+  def openai_api_base
     endpoint = InstallationConfig.find_by(name: 'CAPTAIN_OPEN_AI_ENDPOINT')&.value.presence || 'https://api.openai.com/'
-    endpoint = endpoint.chomp('/')
-    "#{endpoint}/v1/chat/completions"
+    endpoint.chomp('/')
   end
 end

enterprise/app/services/captain/llm/paginated_faq_generator_service.rb

@@ -4,6 +4,30 @@ class Captain::Llm::PaginatedFaqGeneratorService < Llm::LegacyBaseOpenAiService
   # Default pages per chunk - easily configurable
   DEFAULT_PAGES_PER_CHUNK = 10
   MAX_ITERATIONS = 20 # Safety limit to prevent infinite loops
+  FAQ_RESPONSE_SCHEMA = {
+    name: 'pdf_faq_generation',
+    schema: {
+      type: 'object',
+      properties: {
+        faqs: {
+          type: 'array',
+          items: {
+            type: 'object',
+            properties: {
+              question: { type: 'string' },
+              answer: { type: 'string' }
+            },
+            required: %w[question answer],
+            additionalProperties: false
+          }
+        },
+        has_content: { type: 'boolean' }
+      },
+      required: %w[faqs has_content],
+      additionalProperties: false
+    },
+    strict: true
+  }.freeze
 
   attr_reader :total_pages_processed, :iterations_completed
 
@@ -15,7 +39,8 @@ class Captain::Llm::PaginatedFaqGeneratorService < Llm::LegacyBaseOpenAiService
     @max_pages = options[:max_pages] # Optional limit from UI
     @total_pages_processed = 0
     @iterations_completed = 0
-    @model = Llm::FeatureRouter.resolve(feature: 'pdf_faq_generation', account: document.account)[:model]
+    @route = Llm::FeatureRouter.resolve(feature: 'pdf_faq_generation', account: document.account)
+    @model = @route[:model]
   end
 
   def generate
@@ -98,12 +123,17 @@ class Captain::Llm::PaginatedFaqGeneratorService < Llm::LegacyBaseOpenAiService
   end
 
   def process_page_chunk(start_page, end_page)
-    params = build_chunk_parameters(start_page, end_page)
-
-    instrumentation_params = build_instrumentation_params(params, start_page, end_page)
+    messages = build_chunk_messages(start_page, end_page)
+    instrumentation_params = build_instrumentation_params(messages, start_page, end_page)
 
     response = instrument_llm_call(instrumentation_params) do
-      @client.chat(parameters: params)
+      responses_client.create(
+        model: @model,
+        messages: messages,
+        schema: FAQ_RESPONSE_SCHEMA,
+        reasoning_effort: @route[:reasoning_effort],
+        metadata: document_metadata.merge(start_page: start_page, end_page: end_page)
+      )
     end
 
     result = parse_chunk_response(response)
@@ -113,17 +143,13 @@ class Captain::Llm::PaginatedFaqGeneratorService < Llm::LegacyBaseOpenAiService
     { faqs: [], has_content: false }
   end
 
-  def build_chunk_parameters(start_page, end_page)
-    {
-      model: @model,
-      response_format: { type: 'json_object' },
-      messages: [
-        {
-          role: 'user',
-          content: build_user_content(start_page, end_page)
-        }
-      ]
-    }
+  def build_chunk_messages(start_page, end_page)
+    [
+      {
+        role: 'user',
+        content: build_user_content(start_page, end_page)
+      }
+    ]
   end
 
   def build_user_content(start_page, end_page)
@@ -171,7 +197,7 @@ class Captain::Llm::PaginatedFaqGeneratorService < Llm::LegacyBaseOpenAiService
   end
 
   def parse_chunk_response(response)
-    content = response.dig('choices', 0, 'message', 'content')
+    content = response[:message]
     return { 'faqs' => [], 'has_content' => false } if content.nil?
 
     JSON.parse(sanitize_json_response(content))
@@ -208,13 +234,13 @@ class Captain::Llm::PaginatedFaqGeneratorService < Llm::LegacyBaseOpenAiService
     common_words.size.to_f / total_words
   end
 
-  def build_instrumentation_params(params, start_page, end_page)
+  def build_instrumentation_params(messages, start_page, end_page)
     {
       span_name: 'llm.paginated_faq_generation',
       account_id: @document&.account_id,
       feature_name: 'paginated_faq_generation',
       model: @model,
-      messages: params[:messages],
+      messages: messages,
       metadata: document_metadata.merge(start_page: start_page, end_page: end_page, iteration: @iterations_completed + 1)
     }
   end
@@ -222,4 +248,8 @@ class Captain::Llm::PaginatedFaqGeneratorService < Llm::LegacyBaseOpenAiService
   def document_metadata
     @document&.to_llm_metadata || {}
   end
+
+  def responses_client
+    @responses_client ||= Llm::ResponsesClient.new(api_key: nil, client: @client)
+  end
 end

lib/llm/responses_client.rb

@@ -129,12 +129,24 @@ class Llm::ResponsesClient
         { type: 'input_text', text: normalized_part[:text] }
       when 'image_url'
         { type: 'input_image', image_url: normalized_part.dig(:image_url, :url) || normalized_part[:image_url] }
+      when 'file'
+        input_file_part(normalized_part)
       else
         normalized_part
       end
     end
   end
 
+  def input_file_part(part)
+    file = part[:file] || {}
+    {
+      type: 'input_file',
+      file_id: file[:file_id] || part[:file_id],
+      file_url: file[:file_url] || part[:file_url],
+      detail: part[:detail]
+    }.compact
+  end
+
   def text_format(schema)
     return DEFAULT_TEXT_FORMAT unless schema
 

spec/enterprise/models/concerns/agentable_spec.rb

@@ -50,7 +50,8 @@ RSpec.describe Concerns::Agentable do
         tools: [],
         model: Llm::Models.default_model_for('assistant'),
         temperature: 0.8,
-        response_schema: Captain::ResponseSchema
+        response_schema: Captain::ResponseSchema,
+        params: { reasoning_effort: Llm::Models.reasoning_effort_for('assistant') }
       )
 
       dummy_instance.agent
@@ -194,6 +195,14 @@ RSpec.describe Concerns::Agentable do
     end
   end
 
+  describe '#agent_params' do
+    it 'returns reasoning effort for the assistant feature' do
+      expect(dummy_instance.send(:agent_params)).to eq(
+        reasoning_effort: Llm::Models.reasoning_effort_for('assistant')
+      )
+    end
+  end
+
   describe '#agent_response_schema' do
     it 'returns Captain::ResponseSchema' do
       expect(dummy_instance.send(:agent_response_schema)).to eq(Captain::ResponseSchema)

spec/enterprise/services/captain/llm/paginated_faq_generator_service_spec.rb

@@ -35,28 +35,38 @@ RSpec.describe Captain::Llm::PaginatedFaqGeneratorService do
     context 'when generating FAQs from PDF pages' do
       let(:faq_response) do
         {
-          'choices' => [{
-            'message' => {
-              'content' => JSON.generate({
-                                           'faqs' => [
-                                             { 'question' => 'What is this document about?', 'answer' => 'It explains key concepts.' }
-                                           ],
-                                           'has_content' => true
-                                         })
-            }
+          'id' => 'resp-123',
+          'status' => 'completed',
+          'model' => service.model,
+          'output' => [{
+            'type' => 'message',
+            'content' => [{
+              'type' => 'output_text',
+              'text' => JSON.generate({
+                                        'faqs' => [
+                                          { 'question' => 'What is this document about?', 'answer' => 'It explains key concepts.' }
+                                        ],
+                                        'has_content' => true
+                                      })
+            }]
           }]
         }
       end
 
       let(:empty_response) do
         {
-          'choices' => [{
-            'message' => {
-              'content' => JSON.generate({
-                                           'faqs' => [],
-                                           'has_content' => false
-                                         })
-            }
+          'id' => 'resp-456',
+          'status' => 'completed',
+          'model' => service.model,
+          'output' => [{
+            'type' => 'message',
+            'content' => [{
+              'type' => 'output_text',
+              'text' => JSON.generate({
+                                        'faqs' => [],
+                                        'has_content' => false
+                                      })
+            }]
           }]
         }
       end
@@ -66,16 +76,38 @@ RSpec.describe Captain::Llm::PaginatedFaqGeneratorService do
       end
 
       it 'generates FAQs from paginated content' do
-        allow(openai_client).to receive(:chat).and_return(faq_response, empty_response)
+        allow(openai_client).to receive(:json_post).and_return(faq_response, empty_response)
 
         faqs = service.generate
 
         expect(faqs).to have_attributes(size: 1)
         expect(faqs.first['question']).to eq('What is this document about?')
+        expect(openai_client).to have_received(:json_post).with(
+          path: '/responses',
+          parameters: hash_including(
+            model: service.model,
+            reasoning: { effort: Llm::Models.reasoning_effort_for('pdf_faq_generation') },
+            text: {
+              format: hash_including(
+                type: 'json_schema',
+                name: 'pdf_faq_generation',
+                strict: true
+              )
+            },
+            input: [
+              hash_including(
+                content: array_including(
+                  { type: 'input_file', file_id: 'file-123' },
+                  hash_including(type: 'input_text')
+                )
+              )
+            ]
+          )
+        ).at_least(:once)
       end
 
       it 'stops when no more content' do
-        allow(openai_client).to receive(:chat).and_return(empty_response)
+        allow(openai_client).to receive(:json_post).and_return(empty_response)
 
         faqs = service.generate
 
@@ -83,7 +115,7 @@ RSpec.describe Captain::Llm::PaginatedFaqGeneratorService do
       end
 
       it 'respects max iterations limit' do
-        allow(openai_client).to receive(:chat).and_return(faq_response)
+        allow(openai_client).to receive(:json_post).and_return(faq_response)
 
         # Force max iterations
         service.instance_variable_set(:@iterations_completed, 19)

spec/lib/llm/models_spec.rb

@@ -45,7 +45,7 @@ RSpec.describe Llm::Models do
 
     it 'exposes GPT-5.6 models for Captain V2 workflows without changing the legacy PDF path' do
       expect(described_class.models_for('assistant')).to include('gpt-5.6-luna', 'gpt-5.6-terra', 'gpt-5.6-sol')
-      expect(described_class.models_for('pdf_faq_generation')).not_to include('gpt-5.6-luna', 'gpt-5.6-terra', 'gpt-5.6-sol')
+      expect(described_class.models_for('pdf_faq_generation')).to include('gpt-5.6-luna', 'gpt-5.6-terra', 'gpt-5.6-sol')
     end
   end
 

spec/lib/llm/responses_client_spec.rb

@@ -193,7 +193,8 @@ RSpec.describe Llm::ResponsesClient do
             role: 'user',
             content: [
               { type: 'text', text: 'What is in this image?' },
-              { type: 'image_url', image_url: { url: 'https://example.com/image.png' } }
+              { type: 'image_url', image_url: { url: 'https://example.com/image.png' } },
+              { type: 'file', file: { file_id: 'file-123' } }
             ]
           }
         ]
@@ -205,7 +206,8 @@ RSpec.describe Llm::ResponsesClient do
             role: 'user',
             content: [
               { type: 'input_text', text: 'What is in this image?' },
-              { type: 'input_image', image_url: 'https://example.com/image.png' }
+              { type: 'input_image', image_url: 'https://example.com/image.png' },
+              { type: 'input_file', file_id: 'file-123' }
             ]
           }
         ]

spec/models/article_spec.rb

@@ -208,6 +208,58 @@ RSpec.describe Article do
     end
   end
 
+  describe '#generate_article_search_terms' do
+    let(:responses_client) { instance_double(Llm::ResponsesClient) }
+    let(:article) do
+      create(
+        :article,
+        account: account,
+        category: category_1,
+        content: 'How to configure billing invoices and payment reminders',
+        description: 'Billing setup guide',
+        portal: portal_1,
+        author: user,
+        title: 'Configure billing'
+      )
+    end
+
+    before do
+      InstallationConfig.find_or_initialize_by(name: 'CAPTAIN_OPEN_AI_API_KEY').tap do |config|
+        config.value = 'sk-test'
+        config.save!
+      end
+      InstallationConfig.find_or_initialize_by(name: 'CAPTAIN_OPEN_AI_ENDPOINT').tap do |config|
+        config.value = 'https://api.openai.test/v1'
+        config.save!
+      end
+      allow(Llm::ResponsesClient).to receive(:new).and_return(responses_client)
+      allow(responses_client).to receive(:create).and_return(
+        { message: { search_terms: ['billing setup', 'invoice reminders'] }.to_json }
+      )
+    end
+
+    it 'uses Responses API with the help center article model route' do
+      expect(article.generate_article_search_terms).to eq(['billing setup', 'invoice reminders'])
+
+      expect(Llm::ResponsesClient).to have_received(:new).with(
+        api_key: 'sk-test',
+        api_base: 'https://api.openai.test/v1'
+      )
+      expect(responses_client).to have_received(:create).with(
+        hash_including(
+          model: Llm::Models.default_model_for('help_center_article_generation'),
+          reasoning_effort: Llm::Models.reasoning_effort_for('help_center_article_generation'),
+          schema: Enterprise::Concerns::Article::SEARCH_TERMS_SCHEMA,
+          metadata: hash_including(
+            account_id: account.id,
+            article_id: article.id,
+            feature: 'article_search_terms'
+          )
+        )
+      )
+    end
+  end
+
   describe '.update_positions' do
     let!(:article_a) { create(:article, portal: portal_1, category: category_1, author: user, position: 10) }
     let!(:article_b) { create(:article, portal: portal_1, category: category_1, author: user, position: 11) }