spike always sending question to rephraser

Author: davidgisbey

lib/answer_composition/pipeline/question_rephraser.rb

@@ -18,8 +18,6 @@ def initialize(context)
       end
 
       def call
-        return if message_records.blank?
-
         start_time = Clock.monotonic_time
         response = anthropic_bedrock_client.messages.create(
           system: [{ type: "text", text: config[:system_prompt] }],

spec/lib/answer_composition/pipeline/question_rephraser_spec.rb

@@ -30,95 +30,82 @@
     end
   end
 
-  context "when the question is the beginning of the conversation" do
-    let(:context) { build(:answer_pipeline_context) }
-
-    it "returns nil" do
-      expect(described_class.call(context)).to be_nil
-    end
+  it "includes the current question in the user prompt" do
+    described_class.call(context)
+    expect(stub).to have_been_requested
   end
 
-  context "when all other recent answers have statuses in Answer::STATUSES_EXCLUDED_FROM_REPHRASING" do
-    it "returns nil" do
-      conversation = create(:conversation)
-      create(:question, conversation:)
-      Answer::STATUSES_EXCLUDED_FROM_REPHRASING.sample(4) do |status|
-        question = create(:question, conversation:)
-        create(:answer, question:, status:)
-      end
-      latest_question = create(:question, conversation:)
-      context = build(:answer_pipeline_context, question: latest_question)
+  it "includes the message_history in the user prompt" do
+    message_history = <<~HISTORY.strip
+      user:
+      """
+      How do I pay my tax
+      """
+      assistant:
+      """
+      What type of tax
+      """
+      user:
+      """
+      What types are there
+      """
+      assistant:
+      """
+      Self-assessment, PAYE, Corporation tax
+      """
+    HISTORY
+
+    anthropic_request = stub_claude_question_rephrasing(
+      Regexp.new(message_history),
+      rephrased,
+    )
+
+    described_class.call(context)
+
+    expect(anthropic_request).to have_been_made
+  end
 
-      expect(described_class.call(context)).to be_nil
-    end
+  it "updates the context's question_message with the rephrased question" do
+    described_class.call(context)
+    expect(context.question_message).to eq(rephrased)
   end
 
-  context "when the question is part of an ongoing chat" do
-    it "includes the current question in the user prompt" do
-      described_class.call(context)
-      expect(stub).to have_been_requested
-    end
+  it "assigns metrics to the answer" do
+    allow(Clock).to receive(:monotonic_time).and_return(100.0, 101.5)
 
-    it "includes the message_history in the user prompt" do
-      message_history = <<~HISTORY.strip
-        user:
-        """
-        How do I pay my tax
-        """
-        assistant:
-        """
-        What type of tax
-        """
-        user:
-        """
-        What types are there
-        """
-        assistant:
-        """
-        Self-assessment, PAYE, Corporation tax
-        """
-      HISTORY
-
-      anthropic_request = stub_claude_question_rephrasing(
-        Regexp.new(message_history),
-        rephrased,
-      )
+    described_class.call(context)
 
-      described_class.call(context)
+    expect(context.answer.metrics["question_rephrasing"])
+      .to eq({
+        duration: 1.5,
+        llm_prompt_tokens: 10,
+        llm_completion_tokens: 20,
+        llm_cached_tokens: nil,
+        model: BedrockModels.model_id(:claude_sonnet_4_0),
+      })
+  end
 
-      expect(anthropic_request).to have_been_made
-    end
+  it "assigns the llm response to the answer" do
+    described_class.call(context)
 
-    it "updates the context's question_message with the rephrased question" do
-      described_class.call(context)
-      expect(context.question_message).to eq(rephrased)
-    end
+    expected_llm_response = claude_messages_response(
+      content: [claude_messages_text_block(rephrased)],
+      usage: claude_messages_usage_block(input_tokens: 10, output_tokens: 20),
+    ).to_h
 
-    it "assigns metrics to the answer" do
-      allow(Clock).to receive(:monotonic_time).and_return(100.0, 101.5)
+    expect(context.answer.llm_responses["question_rephrasing"])
+      .to eq(expected_llm_response)
+  end
 
-      described_class.call(context)
+  context "when the question is the first in the conversation" do
+    let(:question) { create(:question) }
+    let(:context) { build(:answer_pipeline_context, question:) }
+    let!(:stub) { stub_claude_question_rephrasing(question.message, rephrased) }
 
-      expect(context.answer.metrics["question_rephrasing"])
-        .to eq({
-          duration: 1.5,
-          llm_prompt_tokens: 10,
-          llm_completion_tokens: 20,
-          llm_cached_tokens: nil,
-          model: BedrockModels.model_id(:claude_sonnet_4_0),
-        })
-    end
 
-    it "assigns the llm response to the answer" do
+    it "calls the llm and rephrases the question" do
       described_class.call(context)
-
-      expected_llm_response = claude_messages_response(
-        content: [claude_messages_text_block(rephrased)],
-        usage: claude_messages_usage_block(input_tokens: 10, output_tokens: 20),
-      ).to_h
-
-      expect(context.answer.llm_responses["question_rephrasing"])
-        .to eq(expected_llm_response)
+      expect(stub).to have_been_requested
     end
   end
 

spec/system/conversation_js_features_spec.rb

@@ -182,7 +182,6 @@ def when_the_second_answer_is_generated
 
     stubs_for_mock_answer(@second_question,
                           @second_answer,
-                          rephrase_question: true,
                           sources_used: %w[link_1],
                           create_content_chunk: false)
 
@@ -236,18 +235,15 @@ def then_i_see_a_character_count_error
 
   def stubs_for_mock_answer(question,
                             answer,
-                            rephrase_question: false,
                             sources_used: [],
                             create_content_chunk: true)
     stub_claude_jailbreak_guardrails(question)
 
-    if rephrase_question
-      rephrased_question = "Rephrased #{question}"
+    rephrased_question = "Rephrased #{question}"
 
-      stub_claude_question_rephrasing(question, rephrased_question)
+    stub_claude_question_rephrasing(question, rephrased_question)
 
-      question = rephrased_question
-    end
+    question = rephrased_question
 
     stub_bedrock_titan_embedding(question)
 

spec/system/conversation_with_claude_structured_answer_spec.rb

@@ -35,7 +35,8 @@ def then_i_see_the_answer_is_pending
   end
 
   def when_the_first_answer_is_generated
-    titan_embedding = mock_titan_embedding(@first_question)
+    rephrased_question = "Rephrased #{@first_question}"
+    titan_embedding = mock_titan_embedding(rephrased_question)
     allow(Search::TextToEmbedding)
       .to receive(:call)
       .and_return(titan_embedding)
@@ -47,24 +48,25 @@ def when_the_first_answer_is_generated
     @first_answer = "Lots of tax."
 
     stub_claude_jailbreak_guardrails(@first_question)
-    stub_claude_question_routing(@first_question)
-    stub_claude_structured_answer(@first_question, @first_answer)
+    stub_claude_question_rephrasing(@first_question, rephrased_question)
+    stub_claude_question_routing(rephrased_question)
+    stub_claude_structured_answer(rephrased_question, @first_answer)
     stub_claude_output_guardrails(@first_answer)
-    stub_bedrock_invoke_model_openai_oss_topic_tagger(@first_question)
+    stub_bedrock_invoke_model_openai_oss_topic_tagger(rephrased_question)
     stub_bedrock_invoke_model_openai_oss_answer_relevancy(
-      question_message: @first_question,
+      question_message: rephrased_question,
       answer_message: @first_answer,
     )
     stub_bedrock_invoke_model_openai_oss_faithfulness(
       retrieval_context: "Some content",
       answer_message: @first_answer,
     )
     stub_bedrock_invoke_model_openai_oss_coherence(
-      question_message: @first_question,
+      question_message: rephrased_question,
       answer_message: @first_answer,
     )
     stub_bedrock_invoke_model_openai_oss_context_relevancy(
-      question_message: @first_question,
+      question_message: rephrased_question,
     )
 
     execute_queued_sidekiq_jobs

spec/system/user_conversation_activity_is_shown_in_admin_spec.rb

@@ -34,7 +34,8 @@ def and_i_enter_a_question
   end
 
   def and_the_answer_is_generated
-    titan_embedding = mock_titan_embedding(@question)
+    rephrased_question = "Rephrased #{@question}"
+    titan_embedding = mock_titan_embedding(rephrased_question)
     allow(Search::TextToEmbedding)
       .to receive(:call)
       .and_return(titan_embedding)
@@ -46,24 +47,25 @@ def and_the_answer_is_generated
     @answer = "Maybe. You could get some benefits."
 
     stub_claude_jailbreak_guardrails(@question)
-    stub_claude_question_routing(@question)
-    stub_claude_structured_answer(@question, @answer)
+    stub_claude_question_rephrasing(@question, rephrased_question)
+    stub_claude_question_routing(rephrased_question)
+    stub_claude_structured_answer(rephrased_question, @answer)
     stub_claude_output_guardrails(@answer)
-    stub_bedrock_invoke_model_openai_oss_topic_tagger(@question)
+    stub_bedrock_invoke_model_openai_oss_topic_tagger(rephrased_question)
     stub_bedrock_invoke_model_openai_oss_answer_relevancy(
-      question_message: @question,
+      question_message: rephrased_question,
       answer_message: @answer,
     )
     stub_bedrock_invoke_model_openai_oss_faithfulness(
       retrieval_context: "Some content",
       answer_message: @answer,
     )
     stub_bedrock_invoke_model_openai_oss_coherence(
-      question_message: @question,
+      question_message: rephrased_question,
       answer_message: @answer,
     )
     stub_bedrock_invoke_model_openai_oss_context_relevancy(
-      question_message: @question,
+      question_message: rephrased_question,
     )
 
     execute_queued_sidekiq_jobs