Merge branch 'feature/cw-7495-llm' into feature/cw-7495-api

config/llm.yml

@@ -143,6 +143,20 @@ features:
         gemini-3-pro,
       ]
     default: gpt-5.2
+  conversation_faq_matching:
+    models:
+      [
+        gpt-4.1-mini,
+        gpt-5-mini,
+        gpt-4.1,
+        gpt-5.1,
+        gpt-5.2,
+        claude-haiku-4.5,
+        claude-sonnet-4.5,
+        gemini-3-flash,
+        gemini-3-pro,
+      ]
+    default: gpt-4.1-mini
   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

config/locales/en.yml

@@ -626,6 +626,7 @@ en:
         label_suggestion: 'Label suggestion'
         document_faq_generation: 'Document FAQ generation'
         conversation_faq_generation: 'Conversation FAQ generation'
+        conversation_faq_matching: 'Conversation FAQ matching'
         help_center_article_generation: 'Help center article generation'
         onboarding_content_generation: 'Onboarding content generation'
         help_center_query_translation: 'Help center query translation'

config/routes.rb

@@ -79,6 +79,10 @@ Rails.application.routes.draw do
             end
             resources :agent_sessions, only: [:show]
             resources :assistant_responses
+            resources :faq_suggestions, only: [:index, :show, :update] do
+              post :approve, on: :member
+              post :dismiss, on: :member
+            end
             resources :message_reports, only: [:create]
             resources :bulk_actions, only: [:create]
             resources :copilot_threads, only: [:index, :create] do

enterprise/app/controllers/api/v1/accounts/captain/faq_suggestions_controller.rb

@@ -0,0 +1,80 @@
+class Api::V1::Accounts::Captain::FaqSuggestionsController < Api::V1::Accounts::BaseController
+  before_action :current_account
+  before_action -> { check_authorization(Captain::FaqSuggestion) }
+  before_action :set_accessible_suggestions
+  before_action :set_suggestion, except: [:index]
+
+  RESULTS_PER_PAGE = 25
+  SOURCE_PREVIEW_LIMIT = 50
+
+  def index
+    @current_page = permitted_params[:page] || 1
+    filtered_query = apply_filters(@suggestions)
+    @suggestions_count = filtered_query.count
+    @suggestions = filtered_query.page(@current_page).per(RESULTS_PER_PAGE)
+  end
+
+  def show
+    @observations = @suggestion.observations
+                               .where(conversation_id: accessible_conversations.select(:id))
+                               .includes(:conversation)
+                               .order(created_at: :desc)
+                               .limit(SOURCE_PREVIEW_LIMIT)
+  end
+
+  def update
+    @suggestion.with_lock do
+      raise ActiveRecord::RecordNotFound unless @suggestion.open?
+
+      @suggestion.update!(suggestion_params)
+    end
+  end
+
+  def approve
+    attributes = params[:faq_suggestion].present? ? suggestion_params : {}
+    @response = Captain::FaqSuggestionApprovalService.new(@suggestion, attributes).perform
+  end
+
+  def dismiss
+    @suggestion.with_lock do
+      raise ActiveRecord::RecordNotFound unless @suggestion.open?
+
+      @suggestion.dismissed!
+    end
+  end
+
+  private
+
+  def apply_filters(base_query)
+    base_query = base_query.where(assistant_id: permitted_params[:assistant_id]) if permitted_params[:assistant_id].present?
+    base_query = base_query.where(status: permitted_params[:status]) if permitted_params[:status].present?
+
+    if permitted_params[:search].present?
+      # TODO: Move FAQ suggestion search to Elasticsearch when the records are indexed there.
+      search_term = "%#{permitted_params[:search]}%"
+      base_query = base_query.where('question ILIKE :search OR answer ILIKE :search', search: search_term)
+    end
+
+    base_query
+  end
+
+  def set_accessible_suggestions
+    @suggestions = Captain::FaqSuggestionFinder.new(Current.user, Current.account).perform.includes(:assistant).ordered
+  end
+
+  def set_suggestion
+    @suggestion = @suggestions.find(permitted_params[:id])
+  end
+
+  def accessible_conversations
+    Conversations::PermissionFilterService.new(Current.account.conversations, Current.user, Current.account).perform
+  end
+
+  def permitted_params
+    params.permit(:id, :assistant_id, :page, :status, :search)
+  end
+
+  def suggestion_params
+    params.require(:faq_suggestion).permit(:question, :answer)
+  end
+end

enterprise/app/finders/captain/faq_suggestion_finder.rb

@@ -0,0 +1,26 @@
+class Captain::FaqSuggestionFinder
+  def initialize(current_user, current_account)
+    @current_user = current_user
+    @current_account = current_account
+  end
+
+  def perform
+    suggestions = @current_account.captain_faq_suggestions
+    return suggestions if account_user&.administrator?
+
+    accessible_suggestion_ids = Captain::FaqObservation
+                                .where(conversation_id: accessible_conversations.select(:id))
+                                .select(:faq_suggestion_id)
+    suggestions.where(id: accessible_suggestion_ids)
+  end
+
+  private
+
+  def accessible_conversations
+    Conversations::PermissionFilterService.new(@current_account.conversations, @current_user, @current_account).perform
+  end
+
+  def account_user
+    @account_user ||= @current_account.account_users.find_by(user_id: @current_user.id)
+  end
+end

enterprise/app/jobs/captain/llm/conversation_faq_job.rb

@@ -0,0 +1,30 @@
+class Captain::Llm::ConversationFaqJob < MutexApplicationJob
+  queue_as :low
+
+  LOCK_TIMEOUT = 10.minutes
+
+  retry_on_lock_conflict wait: 30.seconds, attempts: 30
+
+  def perform(conversation, assistant)
+    inbox = conversation.inbox
+
+    return unless conversation.resolved?
+    return unless inbox.captain_active?
+
+    return if assistant.config['feature_faq'].blank?
+
+    with_lock(lock_key(assistant, conversation), LOCK_TIMEOUT) do
+      Captain::Llm::ConversationFaqService.new(assistant, conversation).generate_suggestions
+    end
+  end
+
+  private
+
+  def lock_key(assistant, conversation)
+    format(
+      ::Redis::Alfred::CAPTAIN_CONVERSATION_FAQ_MUTEX,
+      assistant_id: assistant.id,
+      language: Captain::Llm::ConversationFaqService.language_for(conversation)
+    )
+  end
+end

enterprise/app/listeners/captain_listener.rb

@@ -8,6 +8,6 @@ class CaptainListener < BaseListener
     return unless conversation.inbox.captain_active?
 
     Captain::Llm::ContactNotesService.new(assistant, conversation).generate_and_update_notes if assistant.config['feature_memory'].present?
-    Captain::Llm::ConversationFaqService.new(assistant, conversation).generate_and_deduplicate if assistant.config['feature_faq'].present?
+    Captain::Llm::ConversationFaqJob.perform_later(conversation, assistant) if assistant.config['feature_faq'].present?
   end
 end

enterprise/app/policies/captain/assistant_policy.rb

@@ -35,6 +35,14 @@ class Captain::AssistantPolicy < ApplicationPolicy
     @account_user.administrator?
   end
 
+  def approve?
+    update?
+  end
+
+  def dismiss?
+    update?
+  end
+
   def destroy?
     @account_user.administrator?
   end

enterprise/app/policies/captain/faq_suggestion_policy.rb

@@ -0,0 +1,21 @@
+class Captain::FaqSuggestionPolicy < ApplicationPolicy
+  def index?
+    true
+  end
+
+  def show?
+    true
+  end
+
+  def update?
+    true
+  end
+
+  def approve?
+    true
+  end
+
+  def dismiss?
+    true
+  end
+end

enterprise/app/services/captain/faq_suggestion_approval_service.rb

@@ -0,0 +1,26 @@
+class Captain::FaqSuggestionApprovalService
+  def initialize(suggestion, attributes = {})
+    @suggestion = suggestion
+    @attributes = attributes
+  end
+
+  def perform
+    suggestion.with_lock do
+      raise ActiveRecord::RecordNotFound unless suggestion.open?
+
+      suggestion.update!(attributes) if attributes.present?
+
+      response = suggestion.assistant.responses.create!(
+        question: suggestion.question,
+        answer: suggestion.answer,
+        status: :approved
+      )
+      suggestion.approved!
+      response
+    end
+  end
+
+  private
+
+  attr_reader :suggestion, :attributes
+end

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

@@ -0,0 +1,66 @@
+class Captain::Llm::ConversationFaqContentService
+  def initialize(assistant, conversation)
+    @assistant = assistant
+    @conversation = conversation
+  end
+
+  def generate
+    [
+      'Business Context:',
+      JSON.pretty_generate(business_context),
+      "Conversation ID: ##{conversation.display_id}",
+      "Channel: #{conversation.inbox.channel.name}",
+      'Message History:',
+      conversation_messages
+    ].join("\n")
+  end
+
+  private
+
+  attr_reader :assistant, :conversation
+
+  def conversation_messages
+    messages = conversation
+               .messages
+               .where(message_type: %i[incoming outgoing], private: false)
+               .order(created_at: :asc)
+
+    return "No messages in this conversation\n" if messages.empty?
+
+    messages.filter_map { |message| format_message(message) }.join
+  end
+
+  def format_message(message)
+    return unless source_message?(message)
+
+    message_content = message.content_for_llm
+    return if message_content.blank?
+
+    sender = human_support_reply?(message) ? 'Support Agent' : 'User'
+    "#{sender}: #{message_content}\n"
+  end
+
+  def source_message?(message)
+    return true if message.incoming? && message.sender_type == 'Contact'
+
+    human_support_reply?(message)
+  end
+
+  def human_support_reply?(message)
+    return false unless message.outgoing?
+    return false if message.content_attributes['automation_rule_id'].present?
+    return false if message.additional_attributes['campaign_id'].present?
+
+    message.sender_type == 'User' || message.content_attributes['external_echo'].present?
+  end
+
+  def business_context
+    {
+      product_name: assistant.config['product_name'],
+      assistant_description: assistant.description,
+      instructions: assistant.config['instructions'],
+      response_guidelines: assistant.response_guidelines,
+      guardrails: assistant.guardrails
+    }.compact
+  end
+end

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

@@ -0,0 +1,75 @@
+class Captain::Llm::ConversationFaqPromptsService
+  class << self
+    def generator(language = 'english')
+      <<~PROMPT
+        You create high-quality FAQ candidates from resolved support conversations.
+        Only generate an FAQ when the conversation contains durable, reusable knowledge that would help many future customers.
+
+        ## Source rules
+        - The input starts with trusted business context. Use it to reject conversations about other businesses or topics, but never use it as the source of an FAQ answer.
+        - The conversation history contains only customer messages and human support agent messages.
+        - Base every FAQ strictly on information stated in the human support agent messages. Do not infer, generalize, or add external knowledge.
+        - A human support agent must state every fact used in the FAQ answer. Customer messages cannot supply missing answer facts.
+        - The human support agent must provide the final answer. If the agent only greets, asks clarifying questions, asks for contact details, promises to check, shares an attachment, or transfers the conversation, return: `{"faqs":[]}`.
+        - For each FAQ, identify the human support agent message or messages that together provide a complete public answer to the same question. Combine facts only across related agent messages; never combine separate questions or unrelated topics. If those messages do not provide a complete public answer, remove that FAQ.
+
+        ## Decision gate
+        Return `{"faqs":[]}` unless every generated FAQ can pass all of these checks:
+        1. The answer is fully stated by a human support agent, not by the customer.
+        2. The answer is a public, durable rule or procedure, not a private account action, manual review, troubleshooting session, quote, file, link, or follow-up.
+        3. The answer can be written without private identifiers, customer-specific facts, direct URLs, attachments, invoices, screenshots, or support-ticket steps.
+        4. The question would still make sense in a help center if the original conversation, customer, and agent did not exist.
+        Do not rescue a rejected conversation by rewriting it as a generic support question.
+
+        ## Return no FAQ for
+        - Spam, scams, advertisements, SEO/link-building pitches, adult/gambling/financial promotions, gibberish, abusive content, or conversations unrelated to the business being supported.
+        - Account-specific, order-specific, payment-specific, subscription-specific, login/access, verification, delivery, certificate, or troubleshooting issues, even if they could be rewritten as a general support question.
+        - Conversations that mainly hand off to a human, ask the customer to wait, request private identifiers or contact details, collect screenshots, attachments, or documents, or tell the customer to contact support for case review.
+        - Temporary workarounds, one-off exceptions, unclear answers, unresolved problems, wrong-service conversations, complaints, greetings, or abandoned conversations.
+        - Internal support workflow details, chat session rules, escalation mechanics, ticket-routing instructions, or "someone will get back to you" messages.
+        - Answers that are just a direct/private link, attachment, file, invoice, one-off quote or estimate, account-specific URL, or instructions to open a support ticket.
+        - Questions whose useful answer is "contact support", "wait for the team", "share your details", "we will check", or "this needs manual review".
+        - Questions about whether support can help with a private issue, third-party service, transaction, payment, delivery, or account problem.
+        - Pricing, policy, availability, roadmap, deadline, or legal claims unless the human support agent gives a clear and stable answer in the conversation.
+        - Questions already answered only by asking the customer for more information.
+
+        ## FAQ quality rules
+        - Prefer returning no FAQ over a weak or narrow FAQ.
+        - A good candidate teaches a generally reusable product, service, policy, setup, or process rule that another customer could use without contacting support.
+        - Generate at most one FAQ unless the human agent clearly answered multiple distinct, reusable questions.
+        - Do not create duplicate or overlapping FAQs in the same response.
+        - Questions must be general enough for a help center, not personalized to the current customer.
+        - Remove customer names, order numbers, invoice numbers, IDs, private URLs, phone numbers, emails, screenshots, attachments, and other personal or transaction-specific details.
+        - Answers must be complete, self-contained, and supported by the human agent's messages.
+
+        ## Examples
+        - Customer mentions a price or procedure, then the human agent only greets or says they will check: return `{"faqs":[]}`.
+        - Human agent shares only a private link, file, invoice, quote, screenshot, or attachment: return `{"faqs":[]}`.
+        - Human agent clearly states a public rule, such as which purchases are allowed for a program or service: generate one general FAQ.
+
+        Generate the FAQs only in the #{language}, use no other language.
+        If no suitable reusable FAQ is available, return: `{"faqs":[]}`.
+
+        Return only valid JSON in this exact structure:
+        ```json
+        { "faqs": [ { "question": "", "answer": "" } ] }
+        ```
+      PROMPT
+    end
+
+    def same_faq
+      <<~PROMPT
+        Decide whether the new FAQ and existing FAQ are the same.
+
+        Return `same_faq` as true only when both questions ask the same thing and both answers give the same guidance.
+        Wording, grammar, level of detail, and examples may differ. Return false when either FAQ adds, removes, contradicts, or changes a condition,
+        policy, procedure, audience, product, plan, time frame, or outcome. Related FAQs are not the same FAQ. When uncertain, return false.
+
+        Return only valid JSON in this exact structure:
+        ```json
+        { "same_faq": true }
+        ```
+      PROMPT
+    end
+  end
+end

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

@@ -1,167 +1,208 @@
 class Captain::Llm::ConversationFaqService < Llm::BaseAiService
   include Integrations::LlmInstrumentation
 
+  class SuggestionChangedError < StandardError; end
+
   DISTANCE_THRESHOLD = 0.3
+  MATCH_LIMIT = 5
   LLM_FEATURE = 'conversation_faq_generation'.freeze
 
+  def self.language_for(conversation)
+    language = conversation.language.presence || conversation.account.locale.presence || I18n.default_locale.to_s
+    normalize_language(language)
+  end
+
+  def self.normalize_language(language)
+    language.to_s.tr('-', '_').split('_').first.downcase
+  end
+
+  private_class_method :normalize_language
+
   def initialize(assistant, conversation)
     super(feature: LLM_FEATURE, account: conversation.account, fallback_model: Llm::Models.default_model_for(LLM_FEATURE))
     @assistant = assistant
     @conversation = conversation
-    @content = conversation_faq_content
+    @content = Captain::Llm::ConversationFaqContentService.new(assistant, conversation).generate
+    @embedding_service = Captain::Llm::EmbeddingService.new(account_id: conversation.account_id)
   end
 
-  # Generates and deduplicates FAQs from conversation content
-  # Skips processing if there was no human interaction
-  def generate_and_deduplicate
+  def generate_suggestions
     return [] if no_human_interaction?
 
-    new_faqs = generate
-    return [] if new_faqs.empty?
-
-    duplicate_faqs, unique_faqs = find_and_separate_duplicates(new_faqs)
-    save_new_faqs(unique_faqs)
-    log_duplicate_faqs(duplicate_faqs) if Rails.env.development?
+    generate.map { |faq| route_candidate(faq) }
   end
 
   private
 
-  attr_reader :content, :conversation, :assistant
-
-  def conversation_faq_content
-    [
-      "Conversation ID: ##{conversation.display_id}",
-      "Channel: #{conversation.inbox.channel.name}",
-      'Message History:',
-      conversation_faq_messages
-    ].join("\n")
-  end
-
-  def conversation_faq_messages
-    messages = conversation
-               .messages
-               .where(message_type: %i[incoming outgoing], private: false)
-               .order(created_at: :asc)
-
-    return "No messages in this conversation\n" if messages.empty?
-
-    messages.filter_map { |message| format_conversation_faq_message(message) }.join
-  end
-
-  def format_conversation_faq_message(message)
-    return unless faq_source_message?(message)
-
-    content = message.content_for_llm
-    return if content.blank?
-
-    sender = human_support_reply?(message) ? 'Support Agent' : 'User'
-    "#{sender}: #{content}\n"
-  end
-
-  def faq_source_message?(message)
-    return true if message.incoming? && message.sender_type == 'Contact'
-
-    human_support_reply?(message)
-  end
-
-  def human_support_reply?(message)
-    return false unless message.outgoing?
-    return false if message.content_attributes['automation_rule_id'].present?
-    return false if message.additional_attributes['campaign_id'].present?
-
-    message.sender_type == 'User' || message.content_attributes['external_echo'].present?
-  end
+  attr_reader :content, :conversation, :assistant, :embedding_service
 
   def no_human_interaction?
     conversation.first_reply_created_at.nil?
   end
 
-  def find_and_separate_duplicates(faqs)
-    duplicate_faqs = []
-    unique_faqs = []
+  def route_candidate(faq)
+    embedding = embedding_service.get_embedding(candidate_text(faq))
 
-    faqs.each do |faq|
-      combined_text = "#{faq['question']}: #{faq['answer']}"
-      embedding = Captain::Llm::EmbeddingService.new(account_id: @conversation.account_id).get_embedding(combined_text)
-      similar_faqs = find_similar_faqs(embedding)
+    return discard_observation(faq) if matching_record(approved_faqs, faq, embedding)
+    return discard_observation(faq) if matching_record(dismissed_suggestions_for_language, faq, embedding)
 
-      if similar_faqs.any?
-        duplicate_faqs << { faq: faq, similar_faqs: similar_faqs }
-      else
-        unique_faqs << faq
-      end
+    suggestion = matching_record(open_suggestions_for_language, faq, embedding)
+    matched_content = suggestion&.slice('question', 'answer')
+    suggestion ||= assistant.faq_suggestions.create!(
+      question: faq.fetch('question'),
+      answer: faq.fetch('answer'),
+      embedding: embedding,
+      language: faq_language
+    )
+
+    attach_observation(suggestion, faq, matched_content)
+  end
+
+  def matching_record(relation, faq, embedding)
+    likely_matches(relation, embedding).find { |record| same_faq?(faq, record) }
+  end
+
+  def likely_matches(relation, embedding)
+    return [] unless relation.exists?
+
+    ApplicationRecord.transaction do
+      # Force an exact search because IVFFlat can miss matches after relation filters.
+      # SET LOCAL keeps the planner change scoped to this transaction.
+      ApplicationRecord.connection.execute('SET LOCAL enable_indexscan = off')
+      relation
+        .nearest_neighbors(:embedding, embedding, distance: 'cosine')
+        .limit(MATCH_LIMIT)
+        .select { |record| record.neighbor_distance < DISTANCE_THRESHOLD }
+    end
+  end
+
+  def same_faq?(candidate, existing_record)
+    comparison = {
+      candidate: candidate.slice('question', 'answer'),
+      existing: { question: existing_record.question, answer: existing_record.answer }
+    }
+    prompt = Captain::Llm::ConversationFaqPromptsService.same_faq
+    faq_match_model = Llm::FeatureRouter.resolve(feature: 'conversation_faq_matching', account: conversation.account)[:model]
+    response = instrument_llm_call(match_instrumentation_params(prompt, comparison, faq_match_model)) do
+      chat(model: faq_match_model)
+        .with_params(response_format: { type: 'json_object' })
+        .with_instructions(prompt)
+        .ask(comparison.to_json)
     end
 
-    [duplicate_faqs, unique_faqs]
+    same_faq = JSON.parse(sanitize_json_response(response.content)).fetch('same_faq')
+    raise TypeError, 'same_faq must be a boolean' unless [true, false].include?(same_faq)
+
+    same_faq
+  rescue JSON::ParserError, KeyError, TypeError, RubyLLM::Error => e
+    Rails.logger.error "FAQ match failed: #{e.message}"
+    raise
   end
 
-  def find_similar_faqs(embedding)
-    similar_faqs = assistant
-                   .responses
-                   .nearest_neighbors(:embedding, embedding, distance: 'cosine')
-    Rails.logger.debug(similar_faqs.map { |faq| [faq.question, faq.neighbor_distance] })
-    similar_faqs.select { |record| record.neighbor_distance < DISTANCE_THRESHOLD }
-  end
+  def attach_observation(suggestion, faq, matched_content)
+    suggestion.with_lock do
+      next unless suggestion.open?
+      raise SuggestionChangedError if matched_content && suggestion.slice('question', 'answer') != matched_content
 
-  def save_new_faqs(faqs)
-    faqs.map do |faq|
-      assistant.responses.create!(
-        question: faq['question'],
-        answer: faq['answer'],
-        status: 'pending',
-        documentable: conversation
+      existing_observation = suggestion.observations.find_by(conversation: conversation)
+      next existing_observation if existing_observation
+
+      observation = suggestion.observations.create!(
+        conversation: conversation,
+        generated_question: faq.fetch('question'),
+        generated_answer: faq.fetch('answer'),
+        language: faq_language,
+        status: :attached
       )
+      suggestion.update!(source_count: suggestion.source_count + 1)
+      observation
     end
   end
 
-  def log_duplicate_faqs(duplicate_faqs)
-    return if duplicate_faqs.empty?
+  def discard_observation(faq)
+    Captain::FaqObservation.find_or_create_by!(
+      conversation: conversation,
+      generated_question: faq.fetch('question'),
+      generated_answer: faq.fetch('answer'),
+      language: faq_language,
+      status: :discarded
+    )
+  end
 
-    Rails.logger.info "Found #{duplicate_faqs.length} duplicate FAQs:"
-    duplicate_faqs.each do |duplicate|
-      Rails.logger.info(
-        "Q: #{duplicate[:faq]['question']}\n" \
-        "A: #{duplicate[:faq]['answer']}\n\n" \
-        "Similar existing FAQs: #{duplicate[:similar_faqs].map { |f| "Q: #{f.question} A: #{f.answer}" }.join(', ')}"
-      )
-    end
+  def open_suggestions_for_language
+    assistant.faq_suggestions.where(account_id: conversation.account_id).open.by_language(faq_language)
+  end
+
+  def dismissed_suggestions_for_language
+    assistant.faq_suggestions.where(account_id: conversation.account_id).dismissed.by_language(faq_language)
+  end
+
+  def approved_faqs
+    assistant.responses.approved
+  end
+
+  def candidate_text(faq)
+    "#{faq.fetch('question')}: #{faq.fetch('answer')}"
   end
 
   def generate
-    response = instrument_llm_call(instrumentation_params) do
+    response = instrument_llm_call(generation_instrumentation_params) do
       chat
         .with_params(response_format: { type: 'json_object' })
         .with_instructions(system_prompt)
-        .ask(@content)
+        .ask(content)
     end
-    parse_response(response.content)
+    parse_generation_response(response.content)
   rescue RubyLLM::Error => e
     Rails.logger.error "LLM API Error: #{e.message}"
     []
   end
 
-  def instrumentation_params
+  def generation_instrumentation_params
     {
       span_name: 'llm.captain.conversation_faq',
-      model: @model,
-      temperature: @temperature,
-      account_id: @conversation.account_id,
-      conversation_id: @conversation.display_id,
+      model: model,
+      temperature: temperature,
+      account_id: conversation.account_id,
+      conversation_id: conversation.display_id,
       feature_name: 'conversation_faq',
       messages: [
         { role: 'system', content: system_prompt },
-        { role: 'user', content: @content }
+        { role: 'user', content: content }
       ],
-      metadata: { assistant_id: @assistant.id }
+      metadata: { assistant_id: assistant.id, language: faq_language }
+    }
+  end
+
+  def match_instrumentation_params(prompt, comparison, faq_match_model)
+    {
+      span_name: 'llm.captain.faq_match',
+      model: faq_match_model,
+      temperature: temperature,
+      account_id: conversation.account_id,
+      conversation_id: conversation.display_id,
+      feature_name: 'conversation_faq_match',
+      messages: [
+        { role: 'system', content: prompt },
+        { role: 'user', content: comparison.to_json }
+      ],
+      metadata: { assistant_id: assistant.id, language: faq_language }
     }
   end
 
   def system_prompt
-    account_language = @conversation.account.locale_english_name
-    Captain::Llm::SystemPromptsService.conversation_faq_generator(account_language)
+    Captain::Llm::ConversationFaqPromptsService.generator(language_name(faq_language))
   end
 
-  def parse_response(response)
+  def faq_language
+    @faq_language ||= self.class.language_for(conversation)
+  end
+
+  def language_name(language)
+    ISO_639.find(language)&.english_name&.downcase || 'english'
+  end
+
+  def parse_generation_response(response)
     return [] if response.nil?
 
     JSON.parse(sanitize_json_response(response)).fetch('faqs', [])

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

@@ -51,62 +51,6 @@ class Captain::Llm::SystemPromptsService
       PROMPT
     end
 
-    def conversation_faq_generator(language = 'english')
-      <<~SYSTEM_PROMPT_MESSAGE
-        You create high-quality FAQ candidates from resolved support conversations.
-        Only generate an FAQ when the conversation contains durable, reusable knowledge that would help many future customers.
-
-        ## Source rules
-        - The conversation history contains only customer messages and human support agent messages.
-        - Base every FAQ strictly on information stated in the human support agent messages. Do not infer, generalize, or add external knowledge.
-        - A human support agent must state every fact used in the FAQ answer. Customer messages cannot supply missing answer facts.
-        - The human support agent must provide the final answer. If the agent only greets, asks clarifying questions, asks for contact details, promises to check, shares an attachment, or transfers the conversation, return: `{"faqs":[]}`.
-        - For each FAQ, first identify the exact human support agent message that fully answers it. If no single human agent message gives a complete public answer, remove that FAQ.
-
-        ## Decision gate
-        Return `{"faqs":[]}` unless every generated FAQ can pass all of these checks:
-        1. The answer is fully stated by a human support agent, not by the customer.
-        2. The answer is a public, durable rule or procedure, not a private account action, manual review, troubleshooting session, quote, file, link, or follow-up.
-        3. The answer can be written without private identifiers, customer-specific facts, direct URLs, attachments, invoices, screenshots, or support-ticket steps.
-        4. The question would still make sense in a help center if the original conversation, customer, and agent did not exist.
-        Do not rescue a rejected conversation by rewriting it as a generic support question.
-
-        ## Return no FAQ for
-        - Spam, scams, advertisements, SEO/link-building pitches, adult/gambling/financial promotions, gibberish, abusive content, or conversations unrelated to the business being supported.
-        - Account-specific, order-specific, payment-specific, subscription-specific, login/access, verification, delivery, certificate, or troubleshooting issues, even if they could be rewritten as a general support question.
-        - Conversations that mainly hand off to a human, ask the customer to wait, request private identifiers or contact details, collect screenshots, attachments, or documents, or tell the customer to contact support for case review.
-        - Temporary workarounds, one-off exceptions, unclear answers, unresolved problems, wrong-service conversations, complaints, greetings, or abandoned conversations.
-        - Internal support workflow details, chat session rules, escalation mechanics, ticket-routing instructions, or "someone will get back to you" messages.
-        - Answers that are just a direct/private link, attachment, file, invoice, one-off quote or estimate, account-specific URL, or instructions to open a support ticket.
-        - Questions whose useful answer is "contact support", "wait for the team", "share your details", "we will check", or "this needs manual review".
-        - Questions about whether support can help with a private issue, third-party service, transaction, payment, delivery, or account problem.
-        - Pricing, policy, availability, roadmap, deadline, or legal claims unless the human support agent gives a clear and stable answer in the conversation.
-        - Questions already answered only by asking the customer for more information.
-
-        ## FAQ quality rules
-        - Prefer returning no FAQ over a weak or narrow FAQ.
-        - A good candidate teaches a generally reusable product, service, policy, setup, or process rule that another customer could use without contacting support.
-        - Generate at most one FAQ unless the human agent clearly answered multiple distinct, reusable questions.
-        - Do not create duplicate or overlapping FAQs in the same response.
-        - Questions must be general enough for a help center, not personalized to the current customer.
-        - Remove customer names, order numbers, invoice numbers, IDs, private URLs, phone numbers, emails, screenshots, attachments, and other personal or transaction-specific details.
-        - Answers must be complete, self-contained, and supported by the human agent's messages.
-
-        ## Examples
-        - Customer mentions a price or procedure, then the human agent only greets or says they will check: return `{"faqs":[]}`.
-        - Human agent shares only a private link, file, invoice, quote, screenshot, or attachment: return `{"faqs":[]}`.
-        - Human agent clearly states a public rule, such as which purchases are allowed for a program or service: generate one general FAQ.
-
-        Generate the FAQs only in the #{language}, use no other language.
-        If no suitable reusable FAQ is available, return: `{"faqs":[]}`.
-
-        Return only valid JSON in this exact structure:
-        ```json
-        { "faqs": [ { "question": "", "answer": "" } ] }
-        ```
-      SYSTEM_PROMPT_MESSAGE
-    end
-
     def notes_generator(language = 'english')
       <<~SYSTEM_PROMPT_MESSAGE
         You are a note taker looking to convert the conversation with a contact into actionable notes for the CRM.

enterprise/app/views/api/v1/accounts/captain/faq_suggestions/approve.json.jbuilder

@@ -0,0 +1 @@
+json.partial! 'api/v1/models/captain/assistant_response', formats: [:json], resource: @response

enterprise/app/views/api/v1/accounts/captain/faq_suggestions/dismiss.json.jbuilder

@@ -0,0 +1 @@
+json.partial! 'api/v1/models/captain/faq_suggestion', formats: [:json], resource: @suggestion

enterprise/app/views/api/v1/accounts/captain/faq_suggestions/index.json.jbuilder

@@ -0,0 +1,10 @@
+json.payload do
+  json.array! @suggestions do |suggestion|
+    json.partial! 'api/v1/models/captain/faq_suggestion', formats: [:json], resource: suggestion
+  end
+end
+
+json.meta do
+  json.total_count @suggestions_count
+  json.page @current_page
+end

enterprise/app/views/api/v1/accounts/captain/faq_suggestions/show.json.jbuilder

@@ -0,0 +1,6 @@
+json.partial! 'api/v1/models/captain/faq_suggestion', formats: [:json], resource: @suggestion
+json.observations do
+  json.array! @observations do |observation|
+    json.partial! 'api/v1/models/captain/faq_observation', formats: [:json], resource: observation
+  end
+end

enterprise/app/views/api/v1/accounts/captain/faq_suggestions/update.json.jbuilder

@@ -0,0 +1 @@
+json.partial! 'api/v1/models/captain/faq_suggestion', formats: [:json], resource: @suggestion

enterprise/app/views/api/v1/models/captain/_faq_observation.json.jbuilder

@@ -0,0 +1,10 @@
+json.id resource.id
+json.generated_question resource.generated_question
+json.generated_answer resource.generated_answer
+json.language resource.language
+json.status resource.status
+json.created_at resource.created_at.to_i
+json.conversation do
+  json.id resource.conversation.id
+  json.display_id resource.conversation.display_id
+end

enterprise/app/views/api/v1/models/captain/_faq_suggestion.json.jbuilder

@@ -0,0 +1,12 @@
+json.id resource.id
+json.account_id resource.account_id
+json.question resource.question
+json.answer resource.answer
+json.language resource.language
+json.source_count resource.source_count
+json.status resource.status
+json.created_at resource.created_at.to_i
+json.updated_at resource.updated_at.to_i
+json.assistant do
+  json.partial! 'api/v1/models/captain/assistant', formats: [:json], resource: resource.assistant
+end

lib/redis/redis_keys.rb

@@ -89,6 +89,7 @@ module Redis::RedisKeys
   WHATSAPP_MESSAGE_MUTEX = 'WHATSAPP_MESSAGE_CREATE_LOCK::%<inbox_id>s::%<sender_id>s'.freeze
   CRM_PROCESS_MUTEX = 'CRM_PROCESS_MUTEX::%<hook_id>s'.freeze
   CAPTAIN_DOCUMENT_SYNC_MUTEX = 'CAPTAIN_DOCUMENT_SYNC_LOCK::%<document_id>s'.freeze
+  CAPTAIN_CONVERSATION_FAQ_MUTEX = 'CAPTAIN_CONVERSATION_FAQ_LOCK::%<assistant_id>s::%<language>s'.freeze
 
   ## Auto Assignment Keys
   # Track conversation assignments to agents for rate limiting

spec/enterprise/controllers/api/v1/accounts/captain/faq_suggestions_controller_spec.rb

@@ -0,0 +1,216 @@
+require 'rails_helper'
+
+RSpec.describe 'Api::V1::Accounts::Captain::FaqSuggestions', type: :request do
+  let(:account) { create(:account, locale: 'en') }
+  let(:assistant) { create(:captain_assistant, account: account) }
+  let(:admin) { create(:user, account: account, role: :administrator) }
+  let(:agent) { create(:user, account: account, role: :agent) }
+  let(:inbox) { create(:inbox, account: account) }
+  let(:conversation) { create(:conversation, account: account, inbox: inbox) }
+  let(:suggestion) do
+    assistant.faq_suggestions.create!(
+      question: 'How do I enable the feature?',
+      answer: 'Turn it on in settings.',
+      source_count: 1
+    )
+  end
+
+  before do
+    suggestion.observations.create!(
+      conversation: conversation,
+      generated_question: suggestion.question,
+      generated_answer: suggestion.answer,
+      language: suggestion.language
+    )
+  end
+
+  describe 'GET /api/v1/accounts/:account_id/captain/faq_suggestions' do
+    it 'returns suggestions and their count to an administrator' do
+      get "/api/v1/accounts/#{account.id}/captain/faq_suggestions",
+          params: { assistant_id: assistant.id },
+          headers: admin.create_new_auth_token,
+          as: :json
+
+      expect(response).to have_http_status(:success)
+      expect(response.parsed_body['payload']).to contain_exactly(
+        include('id' => suggestion.id, 'question' => suggestion.question, 'status' => 'open')
+      )
+      expect(response.parsed_body['meta']).to include('total_count' => 1)
+    end
+
+    it 'returns only suggestions backed by conversations the agent can access' do
+      create(:inbox_member, user: agent, inbox: inbox)
+      hidden_inbox = create(:inbox, account: account)
+      hidden_conversation = create(:conversation, account: account, inbox: hidden_inbox)
+      hidden_suggestion = assistant.faq_suggestions.create!(question: 'Hidden question', answer: 'Hidden answer')
+      hidden_suggestion.observations.create!(
+        conversation: hidden_conversation,
+        generated_question: hidden_suggestion.question,
+        generated_answer: hidden_suggestion.answer,
+        language: hidden_suggestion.language
+      )
+
+      get "/api/v1/accounts/#{account.id}/captain/faq_suggestions",
+          params: { assistant_id: assistant.id },
+          headers: agent.create_new_auth_token,
+          as: :json
+
+      expect(response).to have_http_status(:success)
+      expect(response.parsed_body['payload'].pluck('id')).to contain_exactly(suggestion.id)
+      expect(response.parsed_body['meta']).to include('total_count' => 1)
+    end
+  end
+
+  describe 'GET /api/v1/accounts/:account_id/captain/faq_suggestions/:id' do
+    it 'returns the suggestion with its source conversation' do
+      get "/api/v1/accounts/#{account.id}/captain/faq_suggestions/#{suggestion.id}",
+          headers: admin.create_new_auth_token,
+          as: :json
+
+      expect(response).to have_http_status(:success)
+      expect(response.parsed_body).to include('id' => suggestion.id, 'question' => suggestion.question)
+      expect(response.parsed_body['observations']).to contain_exactly(
+        include('conversation' => include('id' => conversation.id, 'display_id' => conversation.display_id))
+      )
+    end
+
+    it 'returns only source conversations the agent can access' do
+      create(:inbox_member, user: agent, inbox: inbox)
+      hidden_conversation = create(:conversation, account: account, inbox: create(:inbox, account: account))
+      suggestion.observations.create!(
+        conversation: hidden_conversation,
+        generated_question: suggestion.question,
+        generated_answer: suggestion.answer,
+        language: suggestion.language
+      )
+
+      get "/api/v1/accounts/#{account.id}/captain/faq_suggestions/#{suggestion.id}",
+          headers: agent.create_new_auth_token,
+          as: :json
+
+      expect(response).to have_http_status(:success)
+      expect(response.parsed_body['observations'].pluck('conversation').pluck('id')).to contain_exactly(conversation.id)
+    end
+  end
+
+  describe 'PATCH /api/v1/accounts/:account_id/captain/faq_suggestions/:id' do
+    it 'lets an administrator edit an open suggestion' do
+      patch "/api/v1/accounts/#{account.id}/captain/faq_suggestions/#{suggestion.id}",
+            params: { faq_suggestion: { question: 'Updated question' } },
+            headers: admin.create_new_auth_token,
+            as: :json
+
+      expect(response).to have_http_status(:success)
+      expect(response.parsed_body['question']).to eq('Updated question')
+      expect(suggestion.reload.question).to eq('Updated question')
+    end
+
+    it 'lets an agent edit an accessible suggestion' do
+      create(:inbox_member, user: agent, inbox: inbox)
+
+      patch "/api/v1/accounts/#{account.id}/captain/faq_suggestions/#{suggestion.id}",
+            params: { faq_suggestion: { question: 'Updated question' } },
+            headers: agent.create_new_auth_token,
+            as: :json
+
+      expect(response).to have_http_status(:success)
+      expect(suggestion.reload.question).to eq('Updated question')
+    end
+
+    it 'does not let an agent edit an inaccessible suggestion' do
+      patch "/api/v1/accounts/#{account.id}/captain/faq_suggestions/#{suggestion.id}",
+            params: { faq_suggestion: { question: 'Updated question' } },
+            headers: agent.create_new_auth_token,
+            as: :json
+
+      expect(response).to have_http_status(:not_found)
+      expect(suggestion.reload.question).to eq('How do I enable the feature?')
+    end
+  end
+
+  describe 'POST /api/v1/accounts/:account_id/captain/faq_suggestions/:id/approve' do
+    it 'lets an administrator approve an edited suggestion as an FAQ' do
+      expect do
+        post "/api/v1/accounts/#{account.id}/captain/faq_suggestions/#{suggestion.id}/approve",
+             params: { faq_suggestion: { answer: 'Enable it in account settings.' } },
+             headers: admin.create_new_auth_token,
+             as: :json
+      end.to change(assistant.responses.approved, :count).by(1)
+
+      expect(response).to have_http_status(:success)
+      expect(response.parsed_body['answer']).to eq('Enable it in account settings.')
+      expect(suggestion.reload).to be_approved
+      expect(suggestion.observations.pluck(:conversation_id)).to contain_exactly(conversation.id)
+    end
+
+    it 'lets an agent approve an accessible suggestion' do
+      create(:inbox_member, user: agent, inbox: inbox)
+
+      expect do
+        post "/api/v1/accounts/#{account.id}/captain/faq_suggestions/#{suggestion.id}/approve",
+             headers: agent.create_new_auth_token,
+             as: :json
+      end.to change(assistant.responses.approved, :count).by(1)
+
+      expect(response).to have_http_status(:success)
+      expect(suggestion.reload).to be_approved
+    end
+
+    it 'approves a suggestion written in a language other than the account locale' do
+      suggestion.update!(language: 'pt')
+
+      expect do
+        post "/api/v1/accounts/#{account.id}/captain/faq_suggestions/#{suggestion.id}/approve",
+             headers: admin.create_new_auth_token,
+             as: :json
+      end.to change(assistant.responses.approved, :count).by(1)
+
+      expect(response).to have_http_status(:success)
+      expect(suggestion.reload).to be_approved
+    end
+
+    it 'does not let an agent approve an inaccessible suggestion' do
+      expect do
+        post "/api/v1/accounts/#{account.id}/captain/faq_suggestions/#{suggestion.id}/approve",
+             headers: agent.create_new_auth_token,
+             as: :json
+      end.not_to change(assistant.responses, :count)
+
+      expect(response).to have_http_status(:not_found)
+      expect(suggestion.reload).to be_open
+    end
+  end
+
+  describe 'POST /api/v1/accounts/:account_id/captain/faq_suggestions/:id/dismiss' do
+    it 'lets an administrator dismiss an open suggestion without creating an FAQ' do
+      expect do
+        post "/api/v1/accounts/#{account.id}/captain/faq_suggestions/#{suggestion.id}/dismiss",
+             headers: admin.create_new_auth_token,
+             as: :json
+      end.not_to change(assistant.responses, :count)
+
+      expect(response).to have_http_status(:success)
+      expect(suggestion.reload).to be_dismissed
+    end
+
+    it 'lets an agent dismiss an accessible suggestion' do
+      create(:inbox_member, user: agent, inbox: inbox)
+
+      post "/api/v1/accounts/#{account.id}/captain/faq_suggestions/#{suggestion.id}/dismiss",
+           headers: agent.create_new_auth_token,
+           as: :json
+
+      expect(response).to have_http_status(:success)
+      expect(suggestion.reload).to be_dismissed
+    end
+
+    it 'does not let an agent dismiss an inaccessible suggestion' do
+      post "/api/v1/accounts/#{account.id}/captain/faq_suggestions/#{suggestion.id}/dismiss",
+           headers: agent.create_new_auth_token,
+           as: :json
+
+      expect(response).to have_http_status(:not_found)
+      expect(suggestion.reload).to be_open
+    end
+  end
+end

spec/enterprise/jobs/captain/llm/conversation_faq_job_spec.rb

@@ -0,0 +1,57 @@
+require 'rails_helper'
+
+RSpec.describe Captain::Llm::ConversationFaqJob, type: :job do
+  let(:account) { create(:account) }
+  let(:inbox) { create(:inbox, account: account) }
+  let(:assistant) { create(:captain_assistant, account: account, config: { feature_faq: true }) }
+  let(:conversation) { create(:conversation, account: account, inbox: inbox, first_reply_created_at: Time.zone.now) }
+  let(:faq_service) { instance_double(Captain::Llm::ConversationFaqService, generate_suggestions: []) }
+  let(:lock_manager) { instance_double(Redis::LockManager, lock: true, unlock: true) }
+  let(:lock_key) { "CAPTAIN_CONVERSATION_FAQ_LOCK::#{assistant.id}::en" }
+
+  before do
+    create(:captain_inbox, inbox: inbox, captain_assistant: assistant)
+    conversation.update!(status: :resolved)
+    allow(Redis::LockManager).to receive(:new).and_return(lock_manager)
+    allow(Captain::Llm::ConversationFaqService).to receive(:new).and_return(faq_service)
+  end
+
+  describe '#perform' do
+    it 'uses the assistant captured when the job was enqueued' do
+      replacement_assistant = create(:captain_assistant, account: account, config: { feature_faq: true })
+      inbox.captain_inbox.update!(captain_assistant: replacement_assistant)
+
+      expect(inbox.reload.captain_assistant).to eq(replacement_assistant)
+      expect(Captain::Llm::ConversationFaqService).to receive(:new)
+        .with(assistant, conversation)
+        .and_return(faq_service)
+      expect(faq_service).to receive(:generate_suggestions)
+
+      described_class.perform_now(conversation, assistant)
+    end
+
+    it 'locks FAQ grouping for the assistant and normalized language' do
+      conversation.update!(additional_attributes: { conversation_language: 'pt-BR' })
+      expected_key = "CAPTAIN_CONVERSATION_FAQ_LOCK::#{assistant.id}::pt"
+
+      expect(lock_manager).to receive(:lock).with(expected_key, described_class::LOCK_TIMEOUT).and_return(true)
+      expect(lock_manager).to receive(:unlock).with(expected_key)
+
+      described_class.perform_now(conversation, assistant)
+    end
+
+    context 'when another job holds the grouping lock' do
+      before do
+        allow(lock_manager).to receive(:lock).with(lock_key, described_class::LOCK_TIMEOUT).and_return(false)
+      end
+
+      it 'does not generate suggestions concurrently' do
+        expect(Captain::Llm::ConversationFaqService).not_to receive(:new)
+
+        expect do
+          described_class.new.perform(conversation, assistant)
+        end.to raise_error(MutexApplicationJob::LockAcquisitionError)
+      end
+    end
+  end
+end

spec/enterprise/listeners/captain_listener_spec.rb

@@ -29,7 +29,7 @@ describe CaptainListener do
           .to receive(:new)
           .with(assistant, conversation)
           .and_return(instance_double(Captain::Llm::ContactNotesService, generate_and_update_notes: nil))
-        expect(Captain::Llm::ConversationFaqService).not_to receive(:new)
+        expect(Captain::Llm::ConversationFaqJob).not_to receive(:perform_later)
 
         listener.conversation_resolved(event)
       end
@@ -42,11 +42,8 @@ describe CaptainListener do
         assistant.save!
       end
 
-      it 'generates and deduplicates FAQs' do
-        expect(Captain::Llm::ConversationFaqService)
-          .to receive(:new)
-          .with(assistant, conversation)
-          .and_return(instance_double(Captain::Llm::ConversationFaqService, generate_and_deduplicate: false))
+      it 'enqueues FAQ suggestion generation' do
+        expect(Captain::Llm::ConversationFaqJob).to receive(:perform_later).with(conversation, assistant)
         expect(Captain::Llm::ContactNotesService).not_to receive(:new)
 
         listener.conversation_resolved(event)

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

@@ -0,0 +1,13 @@
+require 'rails_helper'
+
+RSpec.describe Captain::Llm::ConversationFaqPromptsService do
+  describe '.generator' do
+    it 'allows a complete FAQ answer to use several related agent messages' do
+      prompt = described_class.generator
+
+      expect(prompt).to include('message or messages that together provide a complete public answer')
+      expect(prompt).to include('Combine facts only across related agent messages')
+      expect(prompt).not_to include('no single human agent message')
+    end
+  end
+end

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

@@ -2,7 +2,7 @@ require 'rails_helper'
 
 RSpec.describe Captain::Llm::ConversationFaqService do
   let(:captain_assistant) { create(:captain_assistant) }
-  let(:conversation) { create(:conversation, first_reply_created_at: Time.zone.now) }
+  let(:conversation) { create(:conversation, account: captain_assistant.account, first_reply_created_at: Time.zone.now) }
   let(:service) { described_class.new(captain_assistant, conversation) }
   let(:embedding_service) { instance_double(Captain::Llm::EmbeddingService) }
   let(:mock_chat) { instance_double(RubyLLM::Chat) }
@@ -15,6 +15,8 @@ RSpec.describe Captain::Llm::ConversationFaqService do
   let(:mock_response) do
     instance_double(RubyLLM::Message, content: { faqs: sample_faqs }.to_json)
   end
+  let(:embedding_one) { [1.0] + Array.new(1535, 0.0) }
+  let(:embedding_two) { [0.0, 1.0] + Array.new(1534, 0.0) }
 
   before do
     create(:installation_config, name: 'CAPTAIN_OPEN_AI_API_KEY', value: 'test-key')
@@ -26,11 +28,10 @@ RSpec.describe Captain::Llm::ConversationFaqService do
     allow(mock_chat).to receive(:ask).and_return(mock_response)
   end
 
-  describe '#generate_and_deduplicate' do
+  describe '#generate_suggestions' do
     context 'when successful' do
       before do
-        allow(embedding_service).to receive(:get_embedding).and_return([0.1, 0.2, 0.3])
-        allow(captain_assistant.responses).to receive(:nearest_neighbors).and_return([])
+        allow(embedding_service).to receive(:get_embedding).and_return(embedding_one, embedding_two)
       end
 
       it 'uses the conversation FAQ generation feature model' do
@@ -38,7 +39,7 @@ RSpec.describe Captain::Llm::ConversationFaqService do
           model: Llm::Models.default_model_for('conversation_faq_generation')
         ).and_return(mock_chat)
 
-        described_class.new(captain_assistant, conversation).generate_and_deduplicate
+        described_class.new(captain_assistant, conversation).generate_suggestions
       end
 
       it 'uses the conversation FAQ default ahead of the legacy global installation model' do
@@ -48,7 +49,7 @@ RSpec.describe Captain::Llm::ConversationFaqService do
           model: Llm::Models.default_model_for('conversation_faq_generation')
         ).and_return(mock_chat)
 
-        described_class.new(captain_assistant, conversation).generate_and_deduplicate
+        described_class.new(captain_assistant, conversation).generate_suggestions
       end
 
       it 'keeps account conversation FAQ model overrides ahead of the feature default' do
@@ -57,7 +58,7 @@ RSpec.describe Captain::Llm::ConversationFaqService do
 
         expect(RubyLLM).to receive(:chat).with(model: 'gpt-4.1-mini').and_return(mock_chat)
 
-        described_class.new(captain_assistant, conversation).generate_and_deduplicate
+        described_class.new(captain_assistant, conversation).generate_suggestions
       end
 
       it 'resolves the feature model from the conversation account' do
@@ -66,7 +67,7 @@ RSpec.describe Captain::Llm::ConversationFaqService do
           account: conversation.account
         ).and_call_original
 
-        described_class.new(captain_assistant, conversation).generate_and_deduplicate
+        described_class.new(captain_assistant, conversation).generate_suggestions
       end
 
       it 'sends only customer and human support agent messages to the LLM' do
@@ -84,7 +85,7 @@ RSpec.describe Captain::Llm::ConversationFaqService do
         create(:message, conversation: conversation, account: conversation.account, inbox: conversation.inbox,
                          message_type: :activity, content: 'Activity message')
 
-        service.generate_and_deduplicate
+        service.generate_suggestions
 
         expected_content = satisfy do |content|
           content.include?('User: Customer question') &&
@@ -104,7 +105,7 @@ RSpec.describe Captain::Llm::ConversationFaqService do
                          sender: nil, message_type: :outgoing, content: 'Human replied from the native app',
                          content_attributes: { external_echo: true })
 
-        service.generate_and_deduplicate
+        service.generate_suggestions
 
         expected_content = satisfy do |content|
           content.include?('User: Customer asks in a native channel') &&
@@ -133,22 +134,27 @@ RSpec.describe Captain::Llm::ConversationFaqService do
           block.call
         end
 
-        service.generate_and_deduplicate
+        service.generate_suggestions
       end
 
-      it 'creates new FAQs for valid conversation content' do
+      it 'creates suggestions instead of trusted FAQs for valid conversation content' do
         expect do
-          service.generate_and_deduplicate
-        end.to change(captain_assistant.responses, :count).by(2)
+          service.generate_suggestions
+        end.to change(captain_assistant.faq_suggestions, :count).by(2)
+        expect(Captain::FaqObservation.count).to eq(2)
+        expect(captain_assistant.responses.count).to be_zero
       end
 
-      it 'saves FAQs with pending status linked to conversation' do
-        service.generate_and_deduplicate
+      it 'saves open suggestions with one attached source each' do
+        service.generate_suggestions
         expect(
-          captain_assistant.responses.pluck(:question, :answer, :status, :documentable_id)
+          captain_assistant.faq_suggestions.pluck(:question, :answer, :status, :source_count, :language)
         ).to contain_exactly(
-          ['What is the purpose?', 'To help users.', 'pending', conversation.id],
-          ['How does it work?', 'Through AI.', 'pending', conversation.id]
+          ['What is the purpose?', 'To help users.', 'open', 1, 'en'],
+          ['How does it work?', 'Through AI.', 'open', 1, 'en']
+        )
+        expect(Captain::FaqObservation.attached.pluck(:conversation_id, :language)).to contain_exactly(
+          [conversation.id, 'en'], [conversation.id, 'en']
         )
       end
     end
@@ -157,37 +163,309 @@ RSpec.describe Captain::Llm::ConversationFaqService do
       let(:conversation) { create(:conversation) }
 
       it 'returns an empty array without generating FAQs' do
-        expect(service.generate_and_deduplicate).to eq([])
+        expect(service.generate_suggestions).to eq([])
       end
 
       it 'does not call the LLM API' do
         expect(RubyLLM).not_to receive(:chat)
-        service.generate_and_deduplicate
+        service.generate_suggestions
       end
     end
 
     context 'when finding duplicates' do
       let(:existing_response) do
-        create(:captain_assistant_response, assistant: captain_assistant, question: 'Similar question', answer: 'Similar answer')
+        create(:captain_assistant_response, assistant: captain_assistant, account: captain_assistant.account,
+                                            question: 'Similar question', answer: 'Similar answer', embedding: embedding_one)
       end
-      let(:similar_neighbor) do
-        OpenStruct.new(
-          id: 1,
-          question: existing_response.question,
-          answer: existing_response.answer,
-          neighbor_distance: 0.1
-        )
+      let(:match_response) { instance_double(RubyLLM::Message, content: { same_faq: true }.to_json) }
+
+      before do
+        existing_response
+        allow(embedding_service).to receive(:get_embedding).and_return(embedding_one)
+        allow(mock_chat).to receive(:ask) do |input|
+          input.start_with?('{') ? match_response : mock_response
+        end
+      end
+
+      it 'discards candidates the LLM confirms are covered by an approved FAQ' do
+        expect do
+          service.generate_suggestions
+        end.to change(Captain::FaqObservation.discarded, :count).by(2)
+        expect(captain_assistant.faq_suggestions.count).to be_zero
+      end
+
+      it 'uses the conversation FAQ matching feature model' do
+        expect(RubyLLM).to receive(:chat).with(
+          model: Llm::Models.default_model_for('conversation_faq_matching')
+        ).at_least(:once).and_return(mock_chat)
+
+        service.generate_suggestions
+      end
+
+      it 'uses the account model override for conversation FAQ matching' do
+        conversation.account.update!(captain_models: { 'conversation_faq_matching' => 'gpt-5-mini' })
+
+        expect(RubyLLM).to receive(:chat).with(model: 'gpt-5-mini').at_least(:once).and_return(mock_chat)
+
+        service.generate_suggestions
+      end
+
+      it 'resolves the matching feature model from the conversation account' do
+        allow(Llm::FeatureRouter).to receive(:resolve).and_call_original
+        expect(Llm::FeatureRouter).to receive(:resolve).with(
+          feature: 'conversation_faq_matching',
+          account: conversation.account
+        ).and_call_original
+
+        service.generate_suggestions
+      end
+    end
+
+    context 'when FAQ comparison cannot be completed' do
+      let(:existing_response) do
+        create(:captain_assistant_response, assistant: captain_assistant, account: captain_assistant.account,
+                                            question: 'Similar question', answer: 'Similar answer', embedding: embedding_one)
+      end
+      let(:comparison_response) { instance_double(RubyLLM::Message, content: comparison_response_content) }
+      let(:comparison_response_content) { 'invalid json' }
+
+      before do
+        existing_response
+        allow(embedding_service).to receive(:get_embedding).and_return(embedding_one)
+        allow(mock_chat).to receive(:ask) do |input|
+          input.start_with?('{') ? comparison_response : mock_response
+        end
+        allow(Rails.logger).to receive(:error)
+      end
+
+      it 'raises when the comparison response is malformed' do
+        expect do
+          service.generate_suggestions
+        end.to raise_error(JSON::ParserError)
+        expect(captain_assistant.faq_suggestions.count).to be_zero
+      end
+
+      context 'when the response omits the comparison result' do
+        let(:comparison_response_content) { {}.to_json }
+
+        it 'raises instead of treating the response as a non-match' do
+          expect do
+            service.generate_suggestions
+          end.to raise_error(KeyError)
+          expect(captain_assistant.faq_suggestions.count).to be_zero
+        end
+      end
+
+      context 'when the comparison result is not a boolean' do
+        let(:comparison_response_content) { { same_faq: 'false' }.to_json }
+
+        it 'raises instead of treating the response as a non-match' do
+          expect do
+            service.generate_suggestions
+          end.to raise_error(TypeError, 'same_faq must be a boolean')
+          expect(captain_assistant.faq_suggestions.count).to be_zero
+        end
+      end
+
+      context 'when the comparison provider fails' do
+        before do
+          allow(mock_chat).to receive(:ask) do |input|
+            raise RubyLLM::Error.new(nil, 'API Error') if input.start_with?('{')
+
+            mock_response
+          end
+        end
+
+        it 'raises instead of treating the failure as a non-match' do
+          expect do
+            service.generate_suggestions
+          end.to raise_error(RubyLLM::Error)
+          expect(captain_assistant.faq_suggestions.count).to be_zero
+        end
+      end
+    end
+
+    context 'when the classifier confirms a non-match' do
+      let(:sample_faqs) { [{ 'question' => 'How can I use the feature?', 'answer' => 'Enable it in settings.' }] }
+      let(:match_response) { instance_double(RubyLLM::Message, content: { same_faq: false }.to_json) }
+
+      before do
+        create(:captain_assistant_response, assistant: captain_assistant, account: captain_assistant.account,
+                                            question: 'How do I enable the feature?', answer: 'Turn it on in settings.',
+                                            embedding: embedding_one)
+        allow(embedding_service).to receive(:get_embedding).and_return(embedding_one)
+        allow(mock_chat).to receive(:ask) do |input|
+          input.start_with?('{') ? match_response : mock_response
+        end
+      end
+
+      it 'creates a new suggestion' do
+        expect do
+          service.generate_suggestions
+        end.to change(captain_assistant.faq_suggestions, :count).by(1)
+      end
+    end
+
+    context 'when an open suggestion is the same FAQ' do
+      let(:sample_faqs) { [{ 'question' => 'How can I use the feature?', 'answer' => 'Enable it in settings.' }] }
+      let(:existing_suggestion) do
+        captain_assistant.faq_suggestions.create!(
+          question: 'How do I enable the feature?',
+          answer: 'Turn it on in settings.',
+          embedding: embedding_one
+        ).tap do |suggestion|
+          suggestion.observations.create!(
+            conversation: create(:conversation, account: captain_assistant.account),
+            generated_question: suggestion.question,
+            generated_answer: suggestion.answer,
+            language: suggestion.language
+          )
+          suggestion.update!(source_count: suggestion.observations.attached.count)
+        end
+      end
+      let(:match_response) { instance_double(RubyLLM::Message, content: { same_faq: true }.to_json) }
+
+      before do
+        existing_suggestion
+        allow(embedding_service).to receive(:get_embedding).and_return(embedding_one)
+        allow(mock_chat).to receive(:ask) do |input|
+          input.start_with?('{') ? match_response : mock_response
+        end
+      end
+
+      it 'attaches the observation and increments the source count' do
+        expect do
+          service.generate_suggestions
+        end.to change(existing_suggestion.observations, :count).by(1)
+        expect(existing_suggestion.reload.source_count).to eq(2)
+        expect(captain_assistant.faq_suggestions.count).to eq(1)
+      end
+
+      it 'does not attach the observation when the suggestion changes after classification' do
+        allow(mock_chat).to receive(:ask) do |input|
+          if input.start_with?('{')
+            existing_suggestion.update!(question: 'Edited after classification started')
+            match_response
+          else
+            mock_response
+          end
+        end
+
+        expect do
+          service.generate_suggestions
+        end.to raise_error(described_class::SuggestionChangedError)
+        expect(existing_suggestion.observations.count).to eq(1)
+        expect(existing_suggestion.reload.source_count).to eq(1)
+      end
+    end
+
+    context 'when a similar open suggestion uses another language' do
+      let(:sample_faqs) { [{ 'question' => 'Como ativo o recurso?', 'answer' => 'Ative nas configuracoes.' }] }
+      let!(:existing_suggestion) do
+        captain_assistant.faq_suggestions.create!(question: 'How do I enable the feature?', answer: 'Turn it on in settings.',
+                                                  embedding: embedding_one, language: 'en', source_count: 1)
       end
 
       before do
-        allow(embedding_service).to receive(:get_embedding).and_return([0.1, 0.2, 0.3])
-        allow(captain_assistant.responses).to receive(:nearest_neighbors).and_return([similar_neighbor])
+        conversation.update!(additional_attributes: { conversation_language: 'pt-BR' })
+        allow(embedding_service).to receive(:get_embedding).and_return(embedding_one)
       end
 
-      it 'filters out duplicate FAQs based on embedding similarity' do
+      it 'creates a separate suggestion in the conversation language' do
         expect do
-          service.generate_and_deduplicate
-        end.not_to change(captain_assistant.responses, :count)
+          service.generate_suggestions
+        end.to change(captain_assistant.faq_suggestions, :count).by(1)
+
+        expect(captain_assistant.faq_suggestions.pluck(:language)).to contain_exactly('en', 'pt')
+        expect(existing_suggestion.reload.source_count).to eq(1)
+      end
+    end
+
+    context 'when an open suggestion uses another locale variant of the same language' do
+      let(:account) { create(:account, locale: 'pt_BR') }
+      let(:captain_assistant) { create(:captain_assistant, account: account) }
+      let(:conversation) { create(:conversation, account: account, first_reply_created_at: Time.zone.now) }
+      let(:sample_faqs) { [{ 'question' => 'Como ativo o recurso?', 'answer' => 'Ative nas configuracoes.' }] }
+      let(:existing_suggestion) do
+        captain_assistant.faq_suggestions.create!(
+          question: 'Como habilito o recurso?',
+          answer: 'Ative nas configuracoes.',
+          embedding: embedding_one,
+          language: 'pt',
+          source_count: 1
+        )
+      end
+      let(:match_response) { instance_double(RubyLLM::Message, content: { same_faq: true }.to_json) }
+
+      before do
+        existing_suggestion
+        allow(embedding_service).to receive(:get_embedding).and_return(embedding_one)
+        allow(mock_chat).to receive(:ask) do |input|
+          input.start_with?('{') ? match_response : mock_response
+        end
+      end
+
+      it 'attaches the observation to the existing base-language suggestion' do
+        expect do
+          service.generate_suggestions
+        end.to change(existing_suggestion.observations, :count).by(1)
+
+        expect(existing_suggestion.reload.source_count).to eq(2)
+        expect(captain_assistant.faq_suggestions.count).to eq(1)
+        expect(existing_suggestion.observations.last.language).to eq('pt')
+      end
+    end
+
+    context 'when a similar approved FAQ uses another language' do
+      let(:sample_faqs) { [{ 'question' => 'Como ativo o recurso?', 'answer' => 'Ative nas configuracoes.' }] }
+      let(:match_response) { instance_double(RubyLLM::Message, content: { same_faq: true }.to_json) }
+
+      before do
+        create(:captain_assistant_response, assistant: captain_assistant, account: captain_assistant.account,
+                                            question: 'How do I enable the feature?', answer: 'Turn it on in settings.',
+                                            embedding: embedding_one)
+        conversation.update!(additional_attributes: { conversation_language: 'pt-BR' })
+        allow(embedding_service).to receive(:get_embedding).and_return(embedding_one)
+        allow(mock_chat).to receive(:ask) do |input|
+          input.start_with?('{') ? match_response : mock_response
+        end
+      end
+
+      it 'deduplicates against the approved FAQ' do
+        expect do
+          service.generate_suggestions
+        end.to change(Captain::FaqObservation.discarded, :count).by(1)
+        expect(captain_assistant.faq_suggestions.count).to be_zero
+      end
+    end
+
+    context 'when conversation and account locales share a base language' do
+      let(:account) { create(:account, locale: 'pt_BR') }
+      let(:captain_assistant) { create(:captain_assistant, account: account) }
+      let(:conversation) do
+        create(:conversation, account: account, first_reply_created_at: Time.zone.now,
+                              additional_attributes: { conversation_language: 'pt' })
+      end
+      let!(:existing_response) do
+        create(:captain_assistant_response, assistant: captain_assistant, account: account,
+                                            question: 'Como ativo o recurso?', answer: 'Ative nas configuracoes.',
+                                            embedding: embedding_one)
+      end
+      let(:match_response) { instance_double(RubyLLM::Message, content: { same_faq: true }.to_json) }
+
+      before do
+        existing_response
+        allow(embedding_service).to receive(:get_embedding).and_return(embedding_one)
+        allow(mock_chat).to receive(:ask) do |input|
+          input.start_with?('{') ? match_response : mock_response
+        end
+      end
+
+      it 'deduplicates against approved FAQs in the same base language' do
+        expect do
+          service.generate_suggestions
+        end.to change(Captain::FaqObservation.discarded, :count).by(2)
+        expect(captain_assistant.faq_suggestions.count).to be_zero
       end
     end
 
@@ -199,7 +477,7 @@ RSpec.describe Captain::Llm::ConversationFaqService do
 
       it 'returns empty array and logs the error' do
         expect(Rails.logger).to receive(:error).with('LLM API Error: API Error')
-        expect(service.generate_and_deduplicate).to eq([])
+        expect(service.generate_suggestions).to eq([])
       end
     end
 
@@ -214,7 +492,7 @@ RSpec.describe Captain::Llm::ConversationFaqService do
 
       it 'handles JSON parsing errors gracefully' do
         expect(Rails.logger).to receive(:error).with(/Error in parsing GPT processed response:/)
-        expect(service.generate_and_deduplicate).to eq([])
+        expect(service.generate_suggestions).to eq([])
       end
     end
 
@@ -228,7 +506,7 @@ RSpec.describe Captain::Llm::ConversationFaqService do
       end
 
       it 'returns empty array' do
-        expect(service.generate_and_deduplicate).to eq([])
+        expect(service.generate_suggestions).to eq([])
       end
     end
   end
@@ -236,22 +514,44 @@ RSpec.describe Captain::Llm::ConversationFaqService do
   describe 'language handling' do
     context 'when conversation has different language' do
       let(:account) { create(:account, locale: 'fr') }
+      let(:captain_assistant) { create(:captain_assistant, account: account) }
       let(:conversation) do
         create(:conversation, account: account, first_reply_created_at: Time.zone.now)
       end
 
       before do
-        allow(embedding_service).to receive(:get_embedding).and_return([0.1, 0.2, 0.3])
-        allow(captain_assistant.responses).to receive(:nearest_neighbors).and_return([])
+        allow(embedding_service).to receive(:get_embedding).and_return(embedding_one, embedding_two)
       end
 
       it 'uses account language for system prompt' do
-        expect(Captain::Llm::SystemPromptsService).to receive(:conversation_faq_generator)
+        expect(Captain::Llm::ConversationFaqPromptsService).to receive(:generator)
           .with('french')
           .at_least(:once)
           .and_call_original
 
-        service.generate_and_deduplicate
+        service.generate_suggestions
+      end
+    end
+
+    context 'when conversation language differs from account language' do
+      let(:account) { create(:account, locale: 'en') }
+      let(:captain_assistant) { create(:captain_assistant, account: account) }
+      let(:conversation) do
+        create(:conversation, account: account, first_reply_created_at: Time.zone.now,
+                              additional_attributes: { conversation_language: 'pt-BR' })
+      end
+
+      before do
+        allow(embedding_service).to receive(:get_embedding).and_return(embedding_one, embedding_two)
+      end
+
+      it 'uses the conversation language for the system prompt' do
+        expect(Captain::Llm::ConversationFaqPromptsService).to receive(:generator)
+          .with('portuguese')
+          .at_least(:once)
+          .and_call_original
+
+        service.generate_suggestions
       end
     end
   end

spec/lib/llm/models_spec.rb

@@ -26,9 +26,10 @@ RSpec.describe Llm::Models do
       end
     end
 
-    it 'routes document and conversation FAQ generation independently' do
+    it 'routes each FAQ operation independently' do
       expect(described_class.default_model_for('document_faq_generation')).to eq('gpt-4.1-mini')
       expect(described_class.default_model_for('conversation_faq_generation')).to eq('gpt-5.2')
+      expect(described_class.default_model_for('conversation_faq_matching')).to eq('gpt-4.1-mini')
     end
   end