Merge pull request #64 from alphagov/duplicate-openai-directories

Tidy up pipeline namespaces

lib/answer_composition/composer.rb

@@ -46,7 +46,7 @@ def compose_answer
         PipelineRunner.call(question:, pipeline: [
           Pipeline::JailbreakGuardrails,
           Pipeline::QuestionRephraser.new(llm_provider: :openai),
-          Pipeline::QuestionRouter,
+          Pipeline::OpenAI::QuestionRouter,
           Pipeline::QuestionRoutingGuardrails,
           Pipeline::SearchResultFetcher,
           Pipeline::OpenAI::StructuredAnswerComposer,

lib/answer_composition/pipeline/openai/question_router.rb

@@ -0,0 +1,179 @@
+module AnswerComposition::Pipeline::OpenAI
+  class QuestionRouter
+    OPENAI_MODEL = "gpt-4o-mini".freeze
+    MAX_COMPLETION_TOKENS = 160
+
+    def self.call(...) = new(...).call
+
+    def initialize(context)
+      @context = context
+    end
+
+    def call
+      start_time = Clock.monotonic_time
+
+      answer = context.answer
+      answer.assign_llm_response("question_routing", openai_response_choice)
+
+      if genuine_rag?
+        answer.assign_attributes(
+          question_routing_label:,
+          question_routing_confidence_score: llm_classification_data["confidence"],
+        )
+      else
+        answer.assign_attributes(
+          message: use_llm_answer? ? llm_answer : Answer::CannedResponses.response_for_question_routing_label(question_routing_label),
+          status: answer_status,
+          question_routing_label:,
+          question_routing_confidence_score: llm_classification_data["confidence"],
+        )
+
+        context.abort_pipeline unless use_llm_answer?
+      end
+
+      answer.assign_metrics(
+        "question_routing", build_metrics(start_time)
+      )
+    end
+
+  private
+
+    attr_reader :context
+
+    def label_config
+      Rails.configuration.question_routing_labels[question_routing_label]
+    end
+
+    def use_llm_answer?
+      return false if openai_token_limit_reached?
+
+      label_config[:use_answer]
+    end
+
+    def answer_status
+      label_config[:answer_status]
+    end
+
+    def openai_token_limit_reached?
+      openai_response_choice["finish_reason"] == "length"
+    end
+
+    def llm_answer
+      llm_classification_data["answer"]
+    end
+
+    def genuine_rag?
+      question_routing_label == "genuine_rag"
+    end
+
+    def question_routing_label
+      llm_classification_function["name"]
+    end
+
+    def llm_classification_function
+      @llm_classification_function ||= openai_response_choice.dig(
+        "message", "tool_calls", 0, "function"
+      )
+    end
+
+    def openai_response_choice
+      @openai_response_choice ||= openai_response.dig("choices", 0)
+    end
+
+    def raw_llm_classification_data
+      llm_classification_function["arguments"]
+    end
+
+    def llm_classification_data
+      JSON.parse(raw_llm_classification_data)
+    end
+
+    def openai_response
+      @openai_response ||= openai_client.chat(
+        parameters: {
+          model: OPENAI_MODEL,
+          messages:,
+          temperature: 0.0,
+          tools:,
+          tool_choice: "required",
+          parallel_tool_calls: false,
+          max_completion_tokens: MAX_COMPLETION_TOKENS,
+        },
+      )
+    end
+
+    def messages
+      [
+        { role: "system", content: config[:system_prompt] },
+        { role: "user", content: context.question_message },
+      ]
+    end
+
+    def config
+      Rails.configuration.govuk_chat_private.llm_prompts.openai.question_routing
+    end
+
+    def openai_client
+      @openai_client ||= OpenAIClient.build
+    end
+
+    def tools
+      config[:classifications].map do |classification|
+        parameters = {
+          type: "object",
+          additionalProperties: false,
+          properties: {
+            confidence: config[:confidence_property],
+          },
+          required: %w[confidence],
+        }
+
+        if classification[:properties].present?
+          parameters[:required].concat(classification[:required])
+          parameters[:properties].merge!(classification[:properties])
+        end
+
+        {
+          type: "function",
+          function: {
+            name: classification[:name],
+            description: build_description(classification),
+            strict: true,
+            parameters:,
+          },
+        }
+      end
+    end
+
+    def build_description(classification)
+      description = [classification[:description].strip]
+
+      examples = {
+        positive_examples: Array(classification[:examples]),
+        negative_examples: Array(classification[:negative_examples]),
+      }
+
+      examples.each do |key, value|
+        next unless value.any?
+
+        example_string = value.map { |str| "'#{str}'" }.join(", ")
+        description << config["description_#{key}_template"].sub("{examples}", example_string).strip
+      end
+
+      description.compact.join(" ")
+    end
+
+    def error_message(error)
+      "class: #{error.class} message: #{error.message}"
+    end
+
+    def build_metrics(start_time)
+      {
+        duration: Clock.monotonic_time - start_time,
+        llm_prompt_tokens: openai_response.dig("usage", "prompt_tokens"),
+        llm_completion_tokens: openai_response.dig("usage", "completion_tokens"),
+        llm_cached_tokens: openai_response.dig("usage", "prompt_tokens_details", "cached_tokens"),
+      }
+    end
+  end
+end

spec/lib/answer_composition/composer_spec.rb

@@ -43,7 +43,7 @@
         expected_pipeline = [
           AnswerComposition::Pipeline::JailbreakGuardrails,
           AnswerComposition::Pipeline::QuestionRephraser.new(llm_provider: :openai),
-          AnswerComposition::Pipeline::QuestionRouter,
+          AnswerComposition::Pipeline::OpenAI::QuestionRouter,
           AnswerComposition::Pipeline::QuestionRoutingGuardrails,
           AnswerComposition::Pipeline::SearchResultFetcher,
           AnswerComposition::Pipeline::OpenAI::StructuredAnswerComposer,

spec/lib/answer_composition/pipeline/open_ai/question_rephraser_spec.rb → spec/lib/answer_composition/pipeline/openai/question_rephraser_spec.rb

@@ -1,4 +1,4 @@
-RSpec.describe AnswerComposition::Pipeline::OpenAI::QuestionRephraser do
+RSpec.describe AnswerComposition::Pipeline::OpenAI::QuestionRephraser do # rubocop:disable RSpec/SpecFilePathFormat
   let(:conversation) { create :conversation, :with_history }
   let(:question) { conversation.questions.strict_loading(false).last }
   let(:question_records) { conversation.questions.joins(:answer) }

spec/lib/answer_composition/pipeline/question_router_spec.rb → spec/lib/answer_composition/pipeline/openai/question_router_spec.rb

@@ -1,4 +1,4 @@
-RSpec.describe AnswerComposition::Pipeline::QuestionRouter do
+RSpec.describe AnswerComposition::Pipeline::OpenAI::QuestionRouter do # rubocop:disable RSpec/SpecFilePathFormat
   let(:classification_attributes) do
     {
       name: "greetings",