Merge pull request #229 from alphagov/fix-length-bug

Fix JSON parsing error for long question routing responses

Author: kevindew

lib/answer_composition/pipeline/openai/question_router.rb

@@ -18,14 +18,14 @@ def call
       if genuine_rag?
         answer.assign_attributes(
           question_routing_label:,
-          question_routing_confidence_score: llm_classification_data["confidence"],
+          question_routing_confidence_score: confidence_score,
         )
       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"],
+          question_routing_confidence_score: confidence_score,
         )
 
         context.abort_pipeline unless use_llm_answer?
@@ -58,6 +58,12 @@ def openai_token_limit_reached?
       openai_response_choice["finish_reason"] == "length"
     end
 
+    def confidence_score
+      return if openai_token_limit_reached?
+
+      llm_classification_data["confidence"]
+    end
+
     def llm_answer
       llm_classification_data["answer"]
     end

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

@@ -38,7 +38,7 @@
   end
 
   let(:classification_response) do
-    { answer: "Hello!", confidence: 0.85 }
+    { answer: "Hello!", confidence: 0.85 }.to_json
   end
 
   let(:expected_message_history) do
@@ -113,7 +113,7 @@
           expected_message_history,
           tools:,
           function_name: "genuine_rag",
-          function_arguments: { answer: "Generic answer", confidence: 0.9 },
+          function_arguments: { answer: "Generic answer", confidence: 0.9 }.to_json,
         )
       end
 
@@ -144,10 +144,7 @@
           expected_message_history,
           tools:,
           function_name: "vague_acronym_grammar",
-          function_arguments: {
-            answer: "A long answer that is terminated mid senten",
-            confidence: 0.99,
-          },
+          function_arguments: '{"answer": "A long answer that is terminated mid senten',
           finish_reason: "length",
         )
       end
@@ -184,7 +181,7 @@
           expected_message_history,
           tools:,
           function_name: "multi_questions",
-          function_arguments: { answer: answer_message, confidence: 0.9 },
+          function_arguments: { answer: answer_message, confidence: 0.9 }.to_json,
         )
       end
 
@@ -211,7 +208,7 @@
           expected_message_history,
           tools:,
           function_name: "harmful_vulgar_controversy",
-          function_arguments: { answer: "Ignored", confidence: 0.9 },
+          function_arguments: { answer: "Ignored", confidence: 0.9 }.to_json,
         )
       end