main.py

import fitz
from fastapi import FastAPI, File, UploadFile, Form
from fastapi.responses import JSONResponse
from transformers import pipeline
from PIL import Image
from io import BytesIO
from starlette.middleware import Middleware
from starlette.middleware.cors import CORSMiddleware
from pdf2image import convert_from_bytes

app = FastAPI()

# Set up CORS middleware
origins = ["*"]  # or specify your list of allowed origins
app.add_middleware(
    CORSMiddleware,
    allow_origins=origins,
    allow_credentials=True,
    allow_methods=["*"],
    allow_headers=["*"],
)

nlp_qa = pipeline("document-question-answering", model="jinhybr/OCR-DocVQA-Donut")
nlp_qa_v2 = pipeline("document-question-answering", model="faisalraza/layoutlm-invoices")
nlp_qa_v3 = pipeline("question-answering", model="deepset/roberta-base-squad2")
nlp_classification = pipeline("text-classification", model="distilbert/distilbert-base-uncased-finetuned-sst-2-english")
nlp_classification_v2 = pipeline("text-classification", model="cardiffnlp/twitter-roberta-base-sentiment-latest")
nlp_speech_to_text = pipeline("automatic-speech-recognition", model="openai/whisper-base")

description = """
## Image-based Document QA
This API performs document question answering using a LayoutLMv2-based model.

### Endpoints:
- **POST /uploadfile/:** Upload an image file to extract text and answer provided questions.
- **POST /pdfQA/:** Provide a PDF file to extract text and answer provided questions.
"""

app = FastAPI(docs_url="/", description=description)

@app.post("/uploadfile/", description="Upload an image file to extract text and answer provided questions.")
async def perform_document_qa(
    file: UploadFile = File(...),
    questions: str = Form(...),
):
    try:
        # Read the uploaded file as bytes
        contents = await file.read()

        # Open the image using PIL
        image = Image.open(BytesIO(contents))

        # Perform document question answering for each question using LayoutLMv2-based model
        answers_dict = {}
        for question in questions.split(','):
            result = nlp_qa(
                image,
                question.strip()
            )

            # Access the 'answer' key from the first item in the result list
            answer = result[0]['answer']

            # Format the question as a string without extra characters
            formatted_question = question.strip("[]")

            answers_dict[formatted_question] = answer

        return answers_dict
    except Exception as e:
        return JSONResponse(content=f"Error processing file: {str(e)}", status_code=500)

@app.post("/uploadfilev2/", description="Upload an image file to extract text and answer provided questions.")
async def perform_document_qa(
    file: UploadFile = File(...),
    questions: str = Form(...),
):
    try:
        # Read the uploaded file as bytes
        contents = await file.read()

        # Open the image using PIL
        image = Image.open(BytesIO(contents))

        # Perform document question answering for each question using LayoutLMv2-based model
        answers_dict = {}
        for question in questions.split(','):
            result = nlp_qa_v2(
                image,
                question.strip()
            )

            # Access the 'answer' key from the first item in the result list
            answer = result[0]['answer']

            # Format the question as a string without extra characters
            formatted_question = question.strip("[]")

            answers_dict[formatted_question] = answer

        return answers_dict
    except Exception as e:
        return JSONResponse(content=f"Error processing file: {str(e)}", status_code=500)

@app.post("/uploadfilev3/", description="Upload an image file to extract text and answer provided questions.")
async def perform_document_qa(
    context: str = Form(...),
    question: str = Form(...),
):
    try:
        QA_input = {
            'question': question,
            'context': context
        }
        
        res = nlp_qa_v3(QA_input)

        return res['answer']
    except Exception as e:
        return JSONResponse(content=f"Error processing file: {str(e)}", status_code=500)

@app.post("/classify/", description="Classify the provided text.")
async def classify_text(text: str = Form(...)):
    try:
        # Perform text classification using the pipeline
        result = nlp_classification(text)

        # Return the classification result
        return result
    except Exception as e:
        return JSONResponse(content=f"Error classifying text: {str(e)}", status_code=500)

@app.post("/test_classify/", description="Classify the provided text with positive, neutral, or negative sentiment.")
async def test_classify_text(text: str = Form(...)):
    try:
        # Perform text classification using the updated model that returns positive, neutral, or negative
        result = nlp_classification_v2(text)

        # Print the raw label for debugging purposes (can be removed later)
        raw_label = result[0]['label']
        print(f"Raw label from model: {raw_label}")

        # Map the model labels to human-readable format
        label_map = {
            "negative": "Negative",  
            "neutral": "Neutral",   
            "positive": "Positive"   
        }

        # Get the readable label from the map
        formatted_label = label_map.get(raw_label, "Unknown")

        return {"label": formatted_label, "score": result[0]['score']}
    except Exception as e:
        return JSONResponse(content=f"Error classifying text: {str(e)}", status_code=500)


@app.post("/transcribe_and_match/", description="Transcribe audio and match responses to form fields.")
async def transcribe_and_match(
    file: UploadFile = File(...),
    field_data: str = Form(...)
):
    """
    Transcribe audio and match it to form fields.
    :param file: The uploaded audio file.
    :param field_data: A JSON string that contains form field information (field names and IDs).
    """
    try:
        # Step 1: Read and transcribe the audio file
        contents = await file.read()
        transcription_result = nlp_speech_to_text(contents)
        transcription_text = transcription_result['text']

        # Step 2: Parse the field_data (which contains field names/IDs)
        # Example: [{"field_id": "name_field", "field_label": "Name"}, {"field_id": "email_field", "field_label": "Email"}]
        import json
        fields = json.loads(field_data)
        
        # Step 3: Find the matching field for the transcription
        field_matches = {}

        for field in fields:
            field_label = field.get("field_label", "").lower()
            field_id = field.get("field_id", "")
            
            # Simple matching: if the transcribed text contains the field label (or something close)
            if field_label in transcription_text.lower():
                field_matches[field_id] = transcription_text

        # Step 4: Return transcription + matched fields
        return {
            "transcription": transcription_text,
            "matched_fields": field_matches
        }

    except Exception as e:
        return JSONResponse(content=f"Error processing audio or matching fields: {str(e)}", status_code=500)

# Set up CORS middleware
origins = ["*"]  # or specify your list of allowed origins
app.add_middleware(
    CORSMiddleware,
    allow_origins=origins,
    allow_credentials=True,
    allow_methods=["*"],
    allow_headers=["*"],
)