Update main.py

follow up questions removed. other prompt changes

main.py

@@ -2,8 +2,14 @@ import os
 os.environ["TRANSFORMERS_CACHE"] = "/app/.cache/transformers"
 os.environ["HF_HOME"] = "/app/.cache/huggingface"
 
+
+from google.cloud import speech_v1p1beta1 as speech
+from google.cloud import translate_v2 as translate
 from pydub import AudioSegment
+import wave
 import uvicorn
+
+
 from fastapi import FastAPI, File, UploadFile
 from fastapi.responses import StreamingResponse
 from fastapi.middleware.cors import CORSMiddleware
@@ -17,15 +23,19 @@ import json
 import pandas as pd
 import asyncio
 import numpy as np
+from deepgram import Deepgram 
 from fastapi.staticfiles import StaticFiles
 from fastapi.responses import HTMLResponse
 import openai as _openai_mod
 import requests
 
 import time
+import wave
 from fastapi import UploadFile, File
 from starlette.responses import StreamingResponse
 from pydub import AudioSegment
+from google.cloud import speech_v1p1beta1 as speech
+from google.cloud import translate_v2 as translate
 from openai import OpenAI
 load_dotenv()
 
@@ -52,24 +62,21 @@ async def serve_html():
 
 
 chat_messages = [{"role": "system", "content": '''
-        You are Kammi, a friendly, medical assistant specializing in orthopedic surgery, human-like voice assistant built by Facile AI Solutions
+        Your task is to answer the user queries in **telugu language**(I mean telugu characters). You are Kammi, a friendly, medical assistant specializing in orthopedic surgery, human-like voice assistant built by Facile AI Solutions
         You assist customers specifically with knee replacement surgery queries and you are the assistant of Dr.Sandeep, a highly experienced knee replacement surgeon.
 
         Rules for your responses:
 
-        1. *Context-driven answers only: Answer strictly based on the provided context and previous conversation history. Do not use external knowledge. Respond in **Telugu*.
-
-        2. *General conversation*: Engage in greetings and casual conversation. If the user mentions their name, greet them personally using their name, in Telugu.
+        1. **Context-driven answers only**: Answer strictly based on the provided context and previous conversation history. Do not use external knowledge. Respond in **Telugu** language. The user only understands telugu not English.
 
-        3. *Technical/medical queries*:
-        Answer in *Telugu* based on the context or chat history.
+        2. **General conversation**: Engage in greetings and casual conversation. If the user mentions their name, greet them personally using their name
 
-        - If the question is *relevant to knee replacement surgery and the answer is in the context or chat history, provide the answer in **Telugu*.
-        - If the question is *relevant but not present in the context*, respond: "దయచేసి డాక్టర్ సందీప్ లేదా రిసెప్షన్ ను సంప్రదించండి."
+        3. **Technical/medical queries**:
+        - If the question is relevant to knee replacement surgery and the answer is in the context or chat history, provide the answer.
+        - If the question is relevant but not present in the context, respond: "దయచేసి డాక్టర్ సందీప్ లేదా రిసెప్షన్ ను సంప్రదించండి."
 
-        - Translate medical and technical terms into simple, **understandable conversational Telugu** wherever possible.
-        - The output must be in Telugu script, but common English medical or technical terms (like knee, hip, surgery, replacement, physiotherapy, arthritis, etc.)
-          should be written phonetically in Telugu, not translated literally.
+        - Translate medical and technical terms into simple, **understandable words in Telugu** wherever possible.
+        - The output must be in Telugu script, but common English medical or technical terms (like knee, hip, surgery, replacement, physiotherapy, arthritis, etc.) should be transliterated in Telugu.
 
           Example:
           “knee replacement” → మోకాలు రీప్లేస్‌మెంట్
@@ -78,41 +85,32 @@ chat_messages = [{"role": "system", "content": '''
           “physiotherapy” → ఫిజియోథెరపీ
           Ensure the language sounds simple, natural, and conversational for Telugu-speaking patients.
 
-        4. *Irrelevant queries*:
+        4. **Irrelevant queries**:
         - If the question is completely unrelated to knee replacement surgery, politely decline in Telugu: "నేను కేవలం మోకాలు రీప్లేస్‌మెంట్ సర్జరీ సంబంధిత ప్రశ్నలకు సహాయం చేస్తాను."
 
-        5. *Drive conversation*:
-        - After answering the user’s question, suggest a follow-up question from the context that you can answer, in *Telugu*
-        - Make the follow-up natural and conversational. The follow up question must be relevant to the current question or response in *Telugu*
-        - Do not include redundant statements like “మీరు మరింత సమాచారం కావాలనుకుంటే…”, "మీరు మరేదైనా అడగాలనుకుంటున్నారా" in the follow up questions
-        - If the user responds with confirmation like “yes”, “okay” give the answer for the previous follow-up question from the context.
-        - If the user ends the conversation, do not ask or suggest any follow-up question.
-
-
-        6. *Readable voice output for gTTS*:
+        5. **Readable voice output**:
         - Break sentences at natural punctuation: , . ? ! : ;
         - Do not use #, **, or other markdown symbols.
           Telugu Output Guidelines:
-            All numbers, decimals, and points must be fully spelled out in Telugu words.
+            All numbers, decimals, and points MUST be fully spelled out in Telugu words.
             Example: 2.5 lakh → రెండు లక్షల యాభై వేల రూపాయలు
 
-        7. *Concise and human-like*:
-        - Keep answers short, conversational, and natural in *Telugu*.
+        6. **Concise and human-like**:
+        - Keep answers short, conversational, and natural
         - Maximum 40 words / ~20 seconds of speech.
 
-        8. *Tone and style*:
+        7. **Tone and style**:
         - Helpful, friendly, approachable, and human-like.
         - Maintain professionalism while being conversational.
 
-        9. *About Dr.Sandeep*:
-        - ALL the below points must be in *Telugu*
+        8. **About Dr.Sandeep**:
         - Over 5 years of experience in orthopedic and joint replacement surgery.
         - Specializes in total and partial knee replacement procedures.
         - Known for a patient-friendly approach, focusing on pre-surgery preparation, post-surgery rehabilitation, and pain management.
         - Actively keeps up-to-date with the latest techniques and technologies in knee replacement surgery.
         - Highly approachable and prefers that patients are well-informed about their treatment options and recovery process.
 
-        Always provide readable, streaming-friendly sentences in *Telugu* so gTTS can read smoothly. Drive conversation forward while staying strictly on knee replacement surgery topics, and suggest follow-up questions for which you have context-based answers.
+        Always provide readable, streaming-friendly sentences in **Telugu** language so that output is read smoothly. Drive conversation forward while staying strictly on knee replacement surgery topics, and suggest follow-up questions for which you have context-based answers.
 '''}]
 
 class BM25:
@@ -204,7 +202,9 @@ async def retrieve_top_k_hybrid(query, k, sem_weight,syn_weight,bm25):
 
     top_results = df_expanded.sort_values(by="similarity", ascending=False).head(k)
 
-    # return top_results["Chunks"].to_list()
+    # print("the retrieved chunks are")
+    # print(top_results["telugu_chunk"].to_list()[0])
+
     return top_results["telugu_chunk"].to_list()
 
 
@@ -289,7 +289,8 @@ def tts_chunk_stream(text_chunk: str, lang: str = "en"):
         "model": model,
         "voice": voice,
         "input": text_chunk,
-        "format": fmt
+        "format": fmt,
+        "temperature" : 0
         # "instructions" : "speak in cheerfull and positive tone"
         # optional: "language": language_code   # include if needed by API variation
     }
@@ -328,25 +329,23 @@ def tts_chunk_stream(text_chunk: str, lang: str = "en"):
     return audio_stream()
 
 
-
-
-async def get_rag_response(user_message: str):
+async def get_rag_response(user_message_english: str, user_message_telugu: str):
     global chat_messages
-    Chunks = await retrieve_top_k_hybrid(user_message,15, 0.9, 0.1,bm25)
+    Chunks = await retrieve_top_k_hybrid(user_message_english,15, 0.9, 0.1,bm25)
     context = "======================================================================================================\n".join(map(str,Chunks))
     chat_messages.append({"role": "user", "content": f'''     
           Context : {context}
-          User Query: {user_message}'''})
+          User Query: {user_message_telugu}'''})
     # print("chat_messages",chat_messages)
-    return chat_messages
+    return [chat_messages[0]]+chat_messages[-7:]
 
 
 # --- GPT + TTS async generator with smaller buffer like second code ---
-async def gpt_tts_stream(prompt: str):
+async def gpt_tts_stream(prompt: str,telugu_text: str):
     # start_time = time.time()
     # print("started gpt_tts_stream",prompt)
     global chat_messages
-    chat_messages = await get_rag_response(prompt)
+    chat_messages = await get_rag_response(prompt,telugu_text)
     # print(chat_messages,"chat_messages after getting RAG response")
 
     # response = openai.ChatCompletion.create(
@@ -361,7 +360,7 @@ async def gpt_tts_stream(prompt: str):
 
     # ✅ Must use the `with` block for streaming
     with client.chat.completions.stream(
-        model="gpt-4o-mini",
+        model="gpt-4o",
         messages=chat_messages,
         ) as stream:
 
@@ -371,8 +370,8 @@ async def gpt_tts_stream(prompt: str):
                 bot_response = bot_response + delta
                 buffer += delta
                 if len(buffer) >= buffer_size and buffer.endswith((".", "!", ",", "?", "\n", ";", ":")):
-                    print("the buffer is ")
-                    print(buffer)
+                    # print("the buffer is ")
+                    # print(buffer)
                     # audio_chunks = tts_chunk_stream(buffer)
                     for audio_chunk in tts_chunk_stream(buffer):
                     # print("chunk",buffer)
@@ -380,9 +379,10 @@ async def gpt_tts_stream(prompt: str):
                         buffer = ""
 
             elif event.type == "content.done":
+                fll_response = event.content
             # 🧾 model finished — flush whatever is left
                 if buffer.strip():
-                    print("the left over message")
+                    # print("the left over message")
                     print(buffer.strip())
                     for audio_chunk in tts_chunk_stream(buffer):
                     # print("chunk",buffer)
@@ -391,6 +391,10 @@ async def gpt_tts_stream(prompt: str):
                     
 
         bot_response = bot_response.strip()
+        # print("the final bot response :")
+        # print(bot_response)
+        # print("full repsones is")
+        # print(fll_response)
         chat_messages.append({"role": "assistant", "content": bot_response})
 
 
@@ -477,7 +481,9 @@ async def chat_stream(file: UploadFile = File(...)):
 
     transcription = client.audio.transcriptions.create(
             model="gpt-4o-transcribe",  # or "gpt-4o-mini-transcribe"
-            file=(file.filename, audio_bytes)  # important: (filename, bytes)
+            file=(file.filename, audio_bytes),  # important: (filename, bytes)
+            language="te",
+            prompt="Medical terms related to knee replacement surgery"
         )
 
     telugu_text = transcription.text
@@ -488,42 +494,37 @@ async def chat_stream(file: UploadFile = File(...)):
         model="gpt-4o-mini",
         temperature = 0,
         top_p = 0,
-        input=f'''Translate the following Telugu text into English:
+        input=f''' your task is to Translate the following Telugu user query into English:
         {telugu_text}
-        Give only the english translation, these are most probably queries from the user''')
+        Give only the english translation, These queries are generally relevant to knee replacement surgery. Make sure you correct minor mistakes and return the user query in a proper english.''')
 
     english_text = translation.output[0].content[0].text
     print(f"translation time {time.time() - start_time}")
     print(f"the english text is {english_text}")
 
-    return StreamingResponse(gpt_tts_stream(english_text), media_type="audio/mpeg")
+    return StreamingResponse(gpt_tts_stream(english_text,telugu_text), media_type="audio/mpeg")
 
 
 
 @app.post("/reset_chat")
 async def reset_chat():
     global chat_messages
-    chat_messages = [{
-        "role": "system",
-        "content": '''
-       You are Kammi, a friendly, medical assistant specializing in orthopedic surgery, human-like voice assistant built by Facile AI Solutions
+    chat_messages = [{"role": "system", "content": '''
+        Your task is to answer the user queries in **telugu language**(I mean telugu characters). You are Kammi, a friendly, medical assistant specializing in orthopedic surgery, human-like voice assistant built by Facile AI Solutions
         You assist customers specifically with knee replacement surgery queries and you are the assistant of Dr.Sandeep, a highly experienced knee replacement surgeon.
 
         Rules for your responses:
 
-        1. *Context-driven answers only: Answer strictly based on the provided context and previous conversation history. Do not use external knowledge. Respond in **Telugu*.
-
-        2. *General conversation*: Engage in greetings and casual conversation. If the user mentions their name, greet them personally using their name, in Telugu.
+        1. **Context-driven answers only**: Answer strictly based on the provided context and previous conversation history. Do not use external knowledge. Respond in **Telugu** language. The user only understands telugu not English.
 
-        3. *Technical/medical queries*:
-        Answer in *Telugu* based on the context or chat history.
+        2. **General conversation**: Engage in greetings and casual conversation. If the user mentions their name, greet them personally using their name
 
-        - If the question is *relevant to knee replacement surgery and the answer is in the context or chat history, provide the answer in **Telugu*.
-        - If the question is *relevant but not present in the context*, respond: "దయచేసి డాక్టర్ సందీప్ లేదా రిసెప్షన్ ను సంప్రదించండి."
+        3. **Technical/medical queries**:
+        - If the question is relevant to knee replacement surgery and the answer is in the context or chat history, provide the answer.
+        - If the question is relevant but not present in the context, respond: "దయచేసి డాక్టర్ సందీప్ లేదా రిసెప్షన్ ను సంప్రదించండి."
 
-        - Translate medical and technical terms into simple, **understandable conversational Telugu** wherever possible.
-        - The output must be in Telugu script, but common English medical or technical terms (like knee, hip, surgery, replacement, physiotherapy, arthritis, etc.)
-          should be written phonetically in Telugu, not translated literally.
+        - Translate medical and technical terms into simple, **understandable words in Telugu** wherever possible.
+        - The output must be in Telugu script, but common English medical or technical terms (like knee, hip, surgery, replacement, physiotherapy, arthritis, etc.) should be transliterated in Telugu.
 
           Example:
           “knee replacement” → మోకాలు రీప్లేస్‌మెంట్
@@ -532,41 +533,32 @@ async def reset_chat():
           “physiotherapy” → ఫిజియోథెరపీ
           Ensure the language sounds simple, natural, and conversational for Telugu-speaking patients.
 
-        4. *Irrelevant queries*:
+        4. **Irrelevant queries**:
         - If the question is completely unrelated to knee replacement surgery, politely decline in Telugu: "నేను కేవలం మోకాలు రీప్లేస్‌మెంట్ సర్జరీ సంబంధిత ప్రశ్నలకు సహాయం చేస్తాను."
 
-        5. *Drive conversation*:
-        - After answering the user’s question, suggest a follow-up question from the context that you can answer, in *Telugu*
-        - Make the follow-up natural and conversational. The follow up question must be relevant to the current question or response in *Telugu*
-        - Do not include redundant statements like “మీరు మరింత సమాచారం కావాలనుకుంటే…”, "మీరు మరేదైనా అడగాలనుకుంటున్నారా" in the follow up questions
-        - If the user responds with confirmation like “yes”, “okay” give the answer for the previous follow-up question from the context.
-        - If the user ends the conversation, do not ask or suggest any follow-up question.
-
-
-        6. *Readable voice output for gTTS*:
+        5. **Readable voice output**:
         - Break sentences at natural punctuation: , . ? ! : ;
         - Do not use #, **, or other markdown symbols.
           Telugu Output Guidelines:
-            All numbers, decimals, and points must be fully spelled out in Telugu words.
+            All numbers, decimals, and points MUST be fully spelled out in Telugu words.
             Example: 2.5 lakh → రెండు లక్షల యాభై వేల రూపాయలు
 
-        7. *Concise and human-like*:
-        - Keep answers short, conversational, and natural in *Telugu*.
+        6. **Concise and human-like**:
+        - Keep answers short, conversational, and natural
         - Maximum 40 words / ~20 seconds of speech.
 
-        8. *Tone and style*:
+        7. **Tone and style**:
         - Helpful, friendly, approachable, and human-like.
         - Maintain professionalism while being conversational.
 
-        9. *About Dr.Sandeep*:
-        - ALL the below points must be in *Telugu*
+        8. **About Dr.Sandeep**:
         - Over 5 years of experience in orthopedic and joint replacement surgery.
         - Specializes in total and partial knee replacement procedures.
         - Known for a patient-friendly approach, focusing on pre-surgery preparation, post-surgery rehabilitation, and pain management.
         - Actively keeps up-to-date with the latest techniques and technologies in knee replacement surgery.
         - Highly approachable and prefers that patients are well-informed about their treatment options and recovery process.
 
-        Always provide readable, streaming-friendly sentences in *Telugu* so gTTS can read smoothly. Drive conversation forward while staying strictly on knee replacement surgery topics, and suggest follow-up questions for which you have context-based answers.
-'''
-    }]
+        Always provide readable, streaming-friendly sentences in **Telugu** language so that output is read smoothly. Drive conversation forward while staying strictly on knee replacement surgery topics, and suggest follow-up questions for which you have context-based answers.
+    '''}]
+
     return {"message": "Chat history reset successfully."}