downgrade outlines to 0.0.36 to avoid issue 820

app.py

@@ -11,8 +11,8 @@ from pathlib import Path
 from faster_whisper import WhisperModel
 from huggingface_hub import hf_hub_download
 
-llm_repo_id = "Qwen/Qwen2.5-3B-Instruct-GGUF"
-llm_filename="qwen2.5-3b-instruct-q2_k.gguf"
+llm_repo_id = "Qwen/Qwen2.5-1.5B-Instruct-GGUF"
+llm_filename="qwen2.5-1.5b-instruct-q8_0.gguf"
 asr_repo_id = "Luigi/whisper-small-zh_tw-ct2"
 
 llm_model_path = hf_hub_download(repo_id=llm_repo_id, filename=llm_filename)
@@ -55,14 +55,18 @@ def predict_with_llm(text):
     model = load_outlines_model()
 
     prompt = f"""
-        You are an agent specialized in classifying reservation information.
+        You are an expert in classification of restautant customers' message.
 
-        I'm going to provide you with a consumer restaurant reservation message.
-        You need to categorize it under one of the two labels.
-        Please reply with *only* the name of the labels in a JSON object like:
+        I'm going to provide you with a message from a restautant customer.
+        You have to classify it in one of the follwing two intents:
+
+        Reservation: Inquiries and requests highly related to table reservations and seating
+        Others: All other messages that do not involve table booking or reservations
+
+        Please reply with *only* the name of the intent labels in a JSON object like:
         {{\"result\": \"Reservation\"}} or {{\"result\": \"Others\"}}
 
-        Here is the reservation information: {text}
+        Here is the message to classify: {text}
     """.strip()
 
     classifier = choice(model, ["Reservation", "Others"])

requirements.txt

@@ -5,6 +5,6 @@ torch
 torchaudio
 faster-whisper
 soundfile
-outlines[llamacpp]==0.0.36 # https://github.com/dottxt-ai/outlines/issues/820
+outlines[llamacpp]==0.0.36 # issue beyond 0.0.36 https://github.com/dottxt-ai/outlines/issues/820
 numpy>=1.24,<2.0
 llama-cpp-python