sinhala-chatbot / app /main.py
CHAMATH
Deploy Space with optional ASR mode
464b72a
import os
import io
import tempfile
from pathlib import Path
from contextlib import asynccontextmanager
from dotenv import load_dotenv
from fastapi import FastAPI, UploadFile, File, HTTPException
from fastapi.staticfiles import StaticFiles
from fastapi.templating import Jinja2Templates
from fastapi.requests import Request
from fastapi.responses import JSONResponse, StreamingResponse
from fastapi.middleware.cors import CORSMiddleware
import google.generativeai as genai
from gtts import gTTS
from deep_translator import GoogleTranslator
from app import rag
load_dotenv()
asr_model = None
model_loaded = False
model_loading = False
conversation_history = []
GEMINI_API_KEY = os.getenv("GEMINI_API_KEY")
if GEMINI_API_KEY:
genai.configure(api_key=GEMINI_API_KEY)
LOCAL_MODEL_PATH = Path(__file__).resolve().parent.parent / "final_model"
HUGGINGFACE_MODEL_ID = "seniruk/whisper-small-si"
IS_HF_SPACE = bool(os.getenv("SPACE_ID"))
def load_asr_model():
"""Load the ASR model - tries local model first, falls back to Hugging Face."""
global asr_model, model_loaded, model_loading
if model_loaded:
return asr_model
model_loading = True
try:
from transformers import WhisperProcessor, WhisperForConditionalGeneration
import torch
except Exception as import_error:
model_loading = False
raise RuntimeError(
"ASR dependencies are not installed. Install transformers and torch to enable speech input."
) from import_error
processor = None
model = None
model_source = None
if LOCAL_MODEL_PATH.exists():
print("=" * 50)
print(f"Loading ASR model from local path: {LOCAL_MODEL_PATH}")
print("=" * 50)
try:
processor = WhisperProcessor.from_pretrained(str(LOCAL_MODEL_PATH))
model = WhisperForConditionalGeneration.from_pretrained(
str(LOCAL_MODEL_PATH), torch_dtype=torch.float32
)
model_source = "local"
print("Local model loaded successfully.")
except Exception as e:
print(f"Failed to load local model: {str(e)}")
print("Falling back to Hugging Face model...")
processor = None
model = None
else:
print(f"Local model not found at: {LOCAL_MODEL_PATH}")
print("Falling back to Hugging Face model...")
if model is None:
print("=" * 50)
print(f"Loading ASR model from Hugging Face: {HUGGINGFACE_MODEL_ID}")
print("This may take a minute on first run...")
print("=" * 50)
processor = WhisperProcessor.from_pretrained(HUGGINGFACE_MODEL_ID)
model = WhisperForConditionalGeneration.from_pretrained(
HUGGINGFACE_MODEL_ID, torch_dtype=torch.float32
)
model_source = "huggingface"
print("Hugging Face model loaded successfully.")
model.eval()
device = "cpu"
if torch.cuda.is_available():
device = "cuda"
model = model.half()
model = model.to("cuda")
print("Using GPU with float16 for faster inference.")
else:
print("Running on CPU.")
asr_model = {
"processor": processor,
"model": model,
"device": device,
"source": model_source,
}
model_loaded = True
model_loading = False
print(f"Model ready. (Source: {model_source})")
return asr_model
def transcribe_audio(audio_path: str) -> str:
"""Transcribe audio file to text - optimized."""
global asr_model
try:
import soundfile as sf
import numpy as np
from scipy import signal
import torch
except Exception as import_error:
raise RuntimeError(
"Audio dependencies are not installed. Install soundfile, numpy, and scipy."
) from import_error
processor = asr_model["processor"]
model = asr_model["model"]
device = asr_model["device"]
audio_array, sample_rate = sf.read(audio_path)
if len(audio_array.shape) > 1:
audio_array = audio_array.mean(axis=1)
if sample_rate != 16000:
num_samples = int(len(audio_array) * 16000 / sample_rate)
audio_array = signal.resample(audio_array, num_samples)
audio_array = audio_array.astype(np.float32)
inputs = processor(audio_array, sampling_rate=16000, return_tensors="pt").input_features
if device == "cuda":
inputs = inputs.half().to("cuda")
with torch.no_grad():
predicted_ids = model.generate(
inputs,
max_length=225,
num_beams=1,
do_sample=False,
use_cache=True,
)
return processor.batch_decode(predicted_ids, skip_special_tokens=True)[0].strip()
@asynccontextmanager
async def lifespan(app: FastAPI):
"""Load model at startup."""
print("\nStarting Sinhala Chatbot Server...")
if IS_HF_SPACE:
print("Hugging Face Space detected. Skipping heavy startup preloads.")
else:
load_asr_model()
loaded = rag.load_vector_store()
if not loaded:
rag.rebuild_vector_store_from_pdfs()
print("Server ready.\n")
yield
print("\nShutting down...")
app = FastAPI(title="Sinhala Chatbot", version="1.0.0", lifespan=lifespan)
app.add_middleware(
CORSMiddleware,
allow_origins=["*"],
allow_credentials=True,
allow_methods=["*"],
allow_headers=["*"],
)
BASE_DIR = Path(__file__).resolve().parent
app.mount("/static", StaticFiles(directory=BASE_DIR / "static"), name="static")
templates = Jinja2Templates(directory=BASE_DIR / "templates")
@app.get("/")
async def home(request: Request):
"""Render the main chatbot interface."""
return templates.TemplateResponse("index.html", {"request": request})
@app.get("/api/model-status")
async def get_model_status():
"""Check if ASR model is loaded."""
source = asr_model.get("source", None) if asr_model else None
return JSONResponse({"loaded": model_loaded, "loading": model_loading, "source": source})
@app.post("/api/speech-to-text")
async def speech_to_text(audio: UploadFile = File(...)):
"""Convert speech to text using Whisper ASR model."""
if not model_loaded:
try:
load_asr_model()
except Exception as load_error:
raise HTTPException(status_code=503, detail=str(load_error)) from load_error
try:
audio_bytes = await audio.read()
with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as tmp_file:
tmp_file.write(audio_bytes)
tmp_path = tmp_file.name
try:
transcription = transcribe_audio(tmp_path)
return JSONResponse({"success": True, "text": transcription})
finally:
os.unlink(tmp_path)
except Exception as e:
print(f"ASR Error: {str(e)}")
raise HTTPException(status_code=500, detail=f"Speech recognition failed: {str(e)}")
@app.post("/api/chat")
async def chat(request: Request):
"""
Send text to RAG system (retrieves from documents first, then falls back to Gemini/HF).
Automatically translates non-English questions to English before RAG processing.
"""
global conversation_history
try:
data = await request.json()
user_message = data.get("message", "")
if not user_message:
raise HTTPException(status_code=400, detail="Message is required")
english_question = user_message
try:
translator = GoogleTranslator(source="auto", target="en")
english_question = translator.translate(user_message)
print(f"Original Question: {user_message}")
print(f"English Question: {english_question}")
except Exception as trans_error:
print(f"Translation failed, using original: {trans_error}")
english_question = user_message
rag_result = rag.rag_answer(english_question)
assistant_message = rag_result.get("answer", "")
conversation_history.append({
"role": "user",
"parts": [user_message],
})
conversation_history.append({
"role": "model",
"parts": [assistant_message],
})
if len(conversation_history) > 20:
conversation_history = conversation_history[-20:]
return JSONResponse({
"success": True,
"response": assistant_message,
"source": rag_result.get("source", "none"),
"context_found": rag_result.get("context_found", False),
})
except Exception as e:
print(f"Chat Error: {str(e)}")
raise HTTPException(status_code=500, detail=f"Chat failed: {str(e)}")
@app.post("/api/text-to-speech")
async def text_to_speech(request: Request):
"""Convert text to speech using Google TTS."""
try:
data = await request.json()
text = data.get("text", "")
lang = data.get("lang", "si")
if not text:
raise HTTPException(status_code=400, detail="Text is required")
tts = gTTS(text=text, lang=lang, slow=False)
audio_buffer = io.BytesIO()
tts.write_to_fp(audio_buffer)
audio_buffer.seek(0)
return StreamingResponse(
audio_buffer,
media_type="audio/mpeg",
headers={"Content-Disposition": "inline; filename=speech.mp3"},
)
except Exception as e:
print(f"TTS Error: {str(e)}")
raise HTTPException(status_code=500, detail=f"Text-to-speech failed: {str(e)}")
@app.post("/api/clear-history")
async def clear_history():
"""Clear conversation history."""
global conversation_history
conversation_history = []
return JSONResponse({"success": True, "message": "Conversation history cleared"})
@app.get("/api/health")
async def health_check():
"""Health check endpoint."""
return JSONResponse({
"status": "healthy",
"gemini_configured": GEMINI_API_KEY is not None,
})
@app.post("/api/translate-to-english")
async def translate_to_english(request: Request):
"""Translate Sinhala/mixed language question to full English using Google Translate."""
try:
data = await request.json()
question = data.get("question", "")
if not question:
raise HTTPException(status_code=400, detail="Question is required")
translator = GoogleTranslator(source="auto", target="en")
english_question = translator.translate(question)
print(f"Original: {question}")
print(f"Translated: {english_question}")
return JSONResponse({"success": True, "english_question": english_question, "translated": True})
except Exception as e:
print(f"Translation Error: {str(e)}")
error_msg = str(e)
return JSONResponse({
"success": False,
"english_question": question,
"translated": False,
"error": error_msg,
})
@app.post("/api/rag/upload")
async def upload_pdf(file: UploadFile = File(...)):
"""Upload a PDF file for RAG processing."""
if not file.filename.lower().endswith(".pdf"):
raise HTTPException(status_code=400, detail="Only PDF files are allowed")
try:
rag_data_dir = Path(__file__).resolve().parent.parent / "rag_data"
rag_data_dir.mkdir(parents=True, exist_ok=True)
pdf_path = rag_data_dir / file.filename
content = await file.read()
with open(pdf_path, "wb") as f:
f.write(content)
chunks = rag.load_and_process_pdf(str(pdf_path))
if not chunks:
raise HTTPException(status_code=400, detail="Could not extract text from PDF")
success = rag.create_vector_store(chunks)
if success:
status = rag.get_rag_status()
return JSONResponse({
"success": True,
"message": f"PDF '{file.filename}' uploaded and processed successfully",
"chunks_created": len(chunks),
"total_documents": status.get("documents_count", 0),
})
raise HTTPException(status_code=500, detail="Failed to create vector store")
except Exception as e:
print(f"RAG Upload Error: {str(e)}")
raise HTTPException(status_code=500, detail=f"Failed to process PDF: {str(e)}")
@app.post("/api/rag/ask")
async def rag_ask(request: Request):
"""Ask a question using RAG - first checks database, then falls back to Gemini/HF."""
try:
data = await request.json()
question = data.get("question", "")
response_lang = data.get("response_lang", "en")
print(f"Question: {question}")
print(f"Response language: {response_lang}")
if not question:
raise HTTPException(status_code=400, detail="Question is required")
result = rag.rag_answer(question)
answer = result["answer"]
print(f"Original answer length: {len(answer) if answer else 0}")
if response_lang == "si-en" and answer:
print("Translating to Sinhala+English...")
try:
translator = GoogleTranslator(source="en", target="si")
sinhala_answer = translator.translate(answer)
answer = f"**Sinhala:**\n{sinhala_answer}\n\n---\n\n**English:**\n{answer}"
print("Translated successfully.")
except Exception as trans_err:
print(f"Translation to Sinhala failed: {trans_err}")
answer = f"Translation failed: {trans_err}\n\n**English:** {answer}"
return JSONResponse({
"success": True,
"question": question,
"answer": answer,
"source": result["source"],
"context_found": result["context_found"],
"relevance_score": result["relevance_score"],
"response_lang": response_lang,
})
except Exception as e:
print(f"RAG Ask Error: {str(e)}")
raise HTTPException(status_code=500, detail=f"RAG query failed: {str(e)}")
@app.get("/api/rag/status")
async def rag_status():
"""Get RAG system status."""
return JSONResponse(rag.get_rag_status())
@app.post("/api/rag/clear")
async def clear_rag():
"""Clear all RAG data."""
rag.clear_rag_data()
return JSONResponse({"success": True, "message": "RAG data cleared"})
@app.post("/api/rag/rebuild")
async def rebuild_rag():
"""Rebuild vector store from all PDFs in rag_data directory."""
success = rag.rebuild_vector_store_from_pdfs()
if not success:
return JSONResponse(
{
"success": False,
"message": "No valid PDFs found to rebuild vector store.",
}
)
return JSONResponse({"success": True, "message": "RAG vector store rebuilt successfully."})
if __name__ == "__main__":
import uvicorn
uvicorn.run(app, host="0.0.0.0", port=8000, reload=True)