Fix OpenAI transcription zero token usage

lib/ruby_llm/providers/openai/transcription.rb

@@ -53,15 +53,17 @@ def parse_transcription_response(response, model:)
           return RubyLLM::Transcription.new(text: data, model: model) if data.is_a?(String)
 
           usage = data['usage'] || {}
+          input_tokens = usage['input_tokens'].nil? ? usage['prompt_tokens'] : usage['input_tokens']
+          output_tokens = usage['output_tokens'].nil? ? usage['completion_tokens'] : usage['output_tokens']
 
           RubyLLM::Transcription.new(
             text: data['text'],
             model: model,
             language: data['language'],
             duration: data['duration'],
             segments: data['segments'],
-            input_tokens: usage['input_tokens'] || usage['prompt_tokens'],
-            output_tokens: usage['output_tokens'] || usage['completion_tokens']
+            input_tokens: input_tokens,
+            output_tokens: output_tokens
           )
         end
       end

spec/ruby_llm/generators/chat_ui_generator_spec.rb

@@ -134,13 +134,13 @@
         expect(message_content).to include('acts_as_message')
 
         # Check broadcasting setup
-        expect(message_content).to include(%q(broadcasts_to ->(message) { "chat_#{message.chat_id}" }))
+        expect(message_content).to include("broadcasts_to ->(message) { \"chat_\#{message.chat_id}\" }")
         expect(message_content).to include('inserts_by: :append')
 
         # Check broadcast_append_chunk method
         expect(message_content).to include('def broadcast_append_chunk(content)')
-        expect(message_content).to include(%q(broadcast_append_to "chat_#{chat_id}"))
-        expect(message_content).to include(%q(target: "message_#{id}_content"))
+        expect(message_content).to include("broadcast_append_to \"chat_\#{chat_id}\"")
+        expect(message_content).to include("target: \"message_\#{id}_content\"")
         expect(message_content).to include('content: ERB::Util.html_escape(content.to_s)')
       end
     end
@@ -318,13 +318,13 @@
         expect(message_content).to include("model: :llm_model, model_class: 'Llm::Model'")
 
         # Check broadcasting setup
-        expect(message_content).to include(%q(broadcasts_to ->(llm_message) { "llm_chat_#{llm_message.llm_chat_id}" }))
+        expect(message_content).to include("broadcasts_to ->(llm_message) { \"llm_chat_\#{llm_message.llm_chat_id}\" }")
         expect(message_content).to include('inserts_by: :append')
 
         # Check broadcast_append_chunk method
         expect(message_content).to include('def broadcast_append_chunk(content)')
-        expect(message_content).to include(%q(broadcast_append_to "llm_chat_#{llm_chat_id}"))
-        expect(message_content).to include(%q(target: "llm_message_#{id}_content"))
+        expect(message_content).to include("broadcast_append_to \"llm_chat_\#{llm_chat_id}\"")
+        expect(message_content).to include("target: \"llm_message_\#{id}_content\"")
         expect(message_content).to include('content: ERB::Util.html_escape(content.to_s)')
       end
     end

spec/ruby_llm/providers/open_ai/transcription_spec.rb

@@ -0,0 +1,55 @@
+# frozen_string_literal: true
+
+require 'spec_helper'
+
+RSpec.describe RubyLLM::Providers::OpenAI::Transcription do
+  describe '.parse_transcription_response' do
+    let(:response) { instance_double(Faraday::Response, body: body) }
+
+    context 'when explicit zero-valued usage fields are present' do
+      let(:body) do
+        {
+          'text' => 'Transcript',
+          'language' => 'en',
+          'duration' => 1.23,
+          'segments' => [],
+          'usage' => {
+            'input_tokens' => 0,
+            'prompt_tokens' => 12,
+            'output_tokens' => 0,
+            'completion_tokens' => 8
+          }
+        }
+      end
+
+      it 'preserves zero-valued token usage fields' do
+        transcription = described_class.parse_transcription_response(response, model: 'gpt-4o-transcribe')
+
+        expect(transcription.input_tokens).to eq(0)
+        expect(transcription.output_tokens).to eq(0)
+      end
+    end
+
+    context 'when only fallback token usage fields are present' do
+      let(:body) do
+        {
+          'text' => 'Transcript',
+          'language' => 'en',
+          'duration' => 1.23,
+          'segments' => [],
+          'usage' => {
+            'prompt_tokens' => 12,
+            'completion_tokens' => 8
+          }
+        }
+      end
+
+      it 'falls back to prompt/completion token usage fields' do
+        transcription = described_class.parse_transcription_response(response, model: 'gpt-4o-transcribe')
+
+        expect(transcription.input_tokens).to eq(12)
+        expect(transcription.output_tokens).to eq(8)
+      end
+    end
+  end
+end