use json output

app.py

@@ -21,6 +21,7 @@ stt_model = AutoModelForSpeechSeq2Seq.from_pretrained(
 stt_model.to(device)
 
 def transcribe_audio(audio_path: str):
+    transcript = ""
     try:
         audio, sr = librosa.load(audio_path, sr=16000)
         inputs = stt_processor(audio, sampling_rate=16000, return_tensors="pt", language='en')
@@ -50,7 +51,12 @@ def extract_action_items(transcript: str) -> str:
 
     messages = [
         {"role": "system", "content": "You are the perfect action item extraction system."},
-        {"role": "user", "content": "Extract the action items from the voice note transcript using the following format:\n\n- [title 1]: [description 1]\n- [title 2]: [description 2]\n..."},
+        {"role": "user", "content": """Extract the action items from the voice note transcript using the following format:\n\n
+         [
+             {"title 1": "description 1"},
+             "title 2": "description 2"},
+             ...
+         ]"""},
         {"role": "user", "content": f"transcript: {transcript}"}
     ]
     text = tokenizer.apply_chat_template(