Arnavkumar01's picture
change in main.py to explicitely tell Open AI what the format for audio is
b46da1b
raw
history blame
10.5 kB
import os
import io
import json
import re
import tempfile
import asyncio
from typing import Optional
import logging
from contextlib import asynccontextmanager
from fastapi import FastAPI, Request, status, Depends, Header, HTTPException
from fastapi.concurrency import run_in_threadpool
from pydantic import BaseModel
from dotenv import load_dotenv
from openai import OpenAI
from elevenlabs.client import ElevenLabs
from langchain_huggingface import HuggingFaceEmbeddings
from langchain_postgres.vectorstores import PGVector
from sqlalchemy import create_engine
# --- GRADIO ---
import gradio as gr
# --- SETUP ---
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
logging.getLogger('tensorflow').setLevel(logging.ERROR)
logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s')
load_dotenv()
NEON_DATABASE_URL = os.getenv("NEON_DATABASE_URL")
OPENAI_API_KEY = os.getenv("OPENAI_API_KEY")
ELEVENLABS_API_KEY = os.getenv("ELEVENLABS_API_KEY")
SHARED_SECRET = os.getenv("SHARED_SECRET")
# --- CONFIG ---
COLLECTION_NAME = "real_estate_embeddings"
EMBEDDING_MODEL = "hkunlp/instructor-large"
ELEVENLABS_VOICE_NAME = "Leo"
PLANNER_MODEL = "gpt-4o-mini"
ANSWERER_MODEL = "gpt-4o"
TABLE_DESCRIPTIONS = """
- "ongoing_projects_source": Details about projects currently under construction.
- "upcoming_projects_source": Information on future planned projects.
- "completed_projects_source": Facts about projects that are already finished.
- "historical_sales_source": Specific sales records, including price, date, and property ID.
- "past_customers_source": Information about previous customers.
- "feedback_source": Customer feedback and ratings for projects.
"""
# --- CLIENTS ---
embeddings = None
vector_store = None
client_openai = OpenAI(api_key=OPENAI_API_KEY)
client_elevenlabs = ElevenLabs(api_key=ELEVENLABS_API_KEY)
# --- LIFESPAN ---
@asynccontextmanager
async def lifespan(app: FastAPI):
global embeddings, vector_store
logging.info(f"Loading embedding model: {EMBEDDING_MODEL}")
embeddings = HuggingFaceEmbeddings(model_name=EMBEDDING_MODEL)
logging.info(f"Connecting to vector store: {COLLECTION_NAME}")
engine = create_engine(NEON_DATABASE_URL, pool_pre_ping=True)
vector_store = PGVector(
connection=engine,
collection_name=COLLECTION_NAME,
embeddings=embeddings,
)
logging.info("Vector store ready.")
yield
logging.info("Shutting down.")
app = FastAPI(lifespan=lifespan)
# --- PROMPTS ---
QUERY_FORMULATION_PROMPT = f"""
You are a query analysis agent. Transform the user's query into a precise search query and determine the correct table to filter by.
**Available Tables:**
{TABLE_DESCRIPTIONS}
**User's Query:** "{{user_query}}"
**Task:**
1. Rephrase into a clear, keyword-focused English search query.
2. If status keywords (ongoing, completed, upcoming, etc.) are present, pick the matching table.
3. If no status keyword, set filter_table to null.
4. Return JSON: {{"search_query": "...", "filter_table": "table_name or null"}}
"""
ANSWER_SYSTEM_PROMPT = """
You are an expert AI assistant for a premier real estate developer.
## CORE KNOWLEDGE
- Cities: Pune, Mumbai, Bengaluru, Delhi, Chennai, Hyderabad, Goa, Gurgaon, Kolkata.
- Properties: Luxury apartments, villas, commercial.
- Budget: 45 lakhs to 5 crores.
## RULES
1. Match user language (Hinglish β†’ Hinglish, English β†’ English).
2. Use CONTEXT if available, else use core knowledge.
3. Only answer real estate questions.
"""
# --- FIXED: transcribe_audio accepts path + bytes ---
def transcribe_audio(audio_path: str, audio_bytes: bytes) -> str:
for attempt in range(3):
try:
audio_file = io.BytesIO(audio_bytes)
filename = os.path.basename(audio_path) # e.g., "audio.wav"
logging.info(f"Transcribing audio: {filename} ({len(audio_bytes)} bytes)")
transcript = client_openai.audio.transcriptions.create(
model="whisper-1",
file=(filename, audio_file) # ← Critical: gives format hint
)
text = transcript.text.strip()
# Hinglish transliteration
if re.search(r'[\u0900-\u097F]', text):
response = client_openai.chat.completions.create(
model="gpt-4o-mini",
messages=[{"role": "user", "content": f"Transliterate to Roman (Hinglish): {text}"}],
temperature=0.0
)
text = response.choices[0].message.content.strip()
logging.info(f"Transcribed: {text}")
return text
except Exception as e:
logging.error(f"Transcription error (attempt {attempt+1}): {e}")
if attempt == 2:
return ""
return ""
def generate_elevenlabs_sync(text: str, voice: str) -> bytes:
for attempt in range(3):
try:
return client_elevenlabs.generate(
text=text,
voice=voice,
model="eleven_multilingual_v2",
output_format="mp3_44100_128"
)
except Exception as e:
logging.error(f"ElevenLabs error (attempt {attempt+1}): {e}")
if attempt == 2:
return b''
return b''
async def formulate_search_plan(user_query: str) -> dict:
for attempt in range(3):
try:
response = await run_in_threadpool(
client_openai.chat.completions.create,
model=PLANNER_MODEL,
messages=[{"role": "user", "content": QUERY_FORMULATION_PROMPT.format(user_query=user_query)}],
response_format={"type": "json_object"},
temperature=0.0
)
return json.loads(response.choices[0].message.content)
except Exception as e:
logging.error(f"Planner error (attempt {attempt+1}): {e}")
if attempt == 2:
return {"search_query": user_query, "filter_table": None}
return {"search_query": user_query, "filter_table": None}
async def get_agent_response(user_text: str) -> str:
for attempt in range(3):
try:
plan = await formulate_search_plan(user_text)
search_query = plan.get("search_query", user_text)
filter_table = plan.get("filter_table")
search_filter = {"source_table": filter_table} if filter_table else {}
docs = await run_in_threadpool(
vector_store.similarity_search,
search_query, k=3, filter=search_filter
)
if not docs:
docs = await run_in_threadpool(vector_store.similarity_search, search_query, k=3)
context = "\n\n".join([d.page_content for d in docs])
response = await run_in_threadpool(
client_openai.chat.completions.create,
model=ANSWERER_MODEL,
messages=[
{"role": "system", "content": ANSWER_SYSTEM_PROMPT},
{"role": "system", "content": f"CONTEXT:\n{context}"},
{"role": "user", "content": f"Question: {user_text}"}
]
)
return response.choices[0].message.content.strip()
except Exception as e:
logging.error(f"RAG error (attempt {attempt+1}): {e}")
if attempt == 2:
return "Sorry, I couldn't respond. Please try again."
return "Sorry, I couldn't respond."
# --- AUTH ENDPOINT ---
class TextQuery(BaseModel):
query: str
async def verify_token(x_auth_token: str = Header(...)):
if not SHARED_SECRET or x_auth_token != SHARED_SECRET:
logging.warning("Auth failed for /test-text-query")
raise HTTPException(status_code=401, detail="Invalid token")
logging.info("Auth passed")
@app.post("/test-text-query", dependencies=[Depends(verify_token)])
async def test_text_query_endpoint(query: TextQuery):
logging.info(f"Text query: {query.query}")
response = await get_agent_response(query.query)
return {"response": response}
# --- FIXED: process_audio passes path + bytes ---
async def process_audio(audio_path):
if not audio_path or not os.path.exists(audio_path):
return None, "No valid audio file received."
try:
# Read raw bytes
with open(audio_path, "rb") as f:
audio_bytes = f.read()
if len(audio_bytes) == 0:
return None, "Empty audio file."
# 1. Transcribe β€” pass path + bytes
user_text = await run_in_threadpool(transcribe_audio, audio_path, audio_bytes)
if not user_text:
return None, "Couldn't understand audio. Try again."
logging.info(f"User: {user_text}")
# 2. AI Response
agent_response = await get_agent_response(user_text)
if not agent_response:
return None, "No response generated."
logging.info(f"AI: {agent_response[:100]}...")
# 3. Generate Speech
ai_audio_bytes = await run_in_threadpool(
generate_elevenlabs_sync, agent_response, ELEVENLABS_VOICE_NAME
)
if not ai_audio_bytes:
return None, "Failed to generate voice."
# Save to temp file
with tempfile.NamedTemporaryFile(delete=False, suffix=".mp3") as f:
f.write(ai_audio_bytes)
out_path = f.name
return out_path, f"**You:** {user_text}\n\n**AI:** {agent_response}"
except Exception as e:
logging.error(f"Audio processing error: {e}", exc_info=True)
return None, f"Error: {str(e)}"
# --- GRADIO UI ---
with gr.Blocks(title="Real Estate AI") as demo:
gr.Markdown("# Real Estate Voice Assistant")
gr.Markdown("Ask about projects in Pune, Mumbai, Bengaluru, etc.")
with gr.Row():
inp = gr.Audio(sources=["microphone"], type="filepath", label="Speak")
out_audio = gr.Audio(label="AI Response", type="filepath")
out_text = gr.Textbox(label="Conversation", lines=8)
# Only trigger on real file (not example text)
inp.change(process_audio, inp, [out_audio, out_text])
# --- FIXED: Examples now use real audio files (optional) ---
# Remove text examples to avoid FileNotFoundError
# Or: Record real .wav files and upload to repo
# For now: disable examples
# gr.Examples(examples=[], inputs=inp)
# --- MOUNT GRADIO ---
app = gr.mount_gradio_app(app, demo, path="/")