app.py

import os
os.environ["TRANSFORMERS_NO_TORCHVISION"] = "1"

import gradio as gr
import torch
import faiss
import numpy as np
import re
from pypdf import PdfReader
from sentence_transformers import SentenceTransformer
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM


embedder = SentenceTransformer("sentence-transformers/all-MiniLM-L6-v2")

model_name = "google/flan-t5-base"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSeq2SeqLM.from_pretrained(model_name)

index = None
chunks = []


def split_text(text, chunk_size=600, overlap=100):
    pieces = []
    start = 0
    while start < len(text):
        end = start + chunk_size
        chunk = text[start:end]

        # Remove numeric-only chunks
        if len(chunk.strip()) > 50 and not chunk.strip().isdigit():
            pieces.append(chunk)

        start = end - overlap
    return pieces


def clean_text(text):
    # Remove References section completely
    text = re.split(r'References|REFERENCES', text)[0]

    # Remove arXiv IDs
    text = re.sub(r'arXiv:\d+\.\d+', '', text)

    # Remove citation numbers like [12]
    text = re.sub(r'\[\d+\]', '', text)

    # Remove lines that are mostly numbers
    text = re.sub(r'^\d+\s*$', '', text, flags=re.MULTILINE)

    return text


def process_pdf(file):
    global index, chunks

    if file is None:
        return "Please upload a PDF."

    reader = PdfReader(file)
    full_text = ""

    for page in reader.pages:
        text = page.extract_text()
        if text:
            full_text += text

    if full_text.strip() == "":
        return "PDF has no extractable text."

    full_text = clean_text(full_text)
    chunks = split_text(full_text)

    embeddings = embedder.encode(chunks)
    embeddings = np.array(embeddings).astype("float32")

    dimension = embeddings.shape[1]
    index = faiss.IndexFlatL2(dimension)
    index.add(embeddings)

    return "PDF processed successfully! Ask your question."


def ask_question(question):
    global index, chunks

    if index is None:
        return "Please process a PDF first."

    question_embedding = embedder.encode([question]).astype("float32")
    D, I = index.search(question_embedding, k=3)

    retrieved_chunks = [chunks[i] for i in I[0]]
    context = "\n\n".join(retrieved_chunks)

    prompt = f"""
You are a research assistant.

Explain clearly what the paper is about.
Answer in 3-5 complete sentences.
Do not include citations or reference numbers.
If unclear, say the document does not clearly specify.

Context:
{context}

Question:
{question}

Answer:
"""

    inputs = tokenizer(prompt, return_tensors="pt", truncation=True)
    outputs = model.generate(
        **inputs,
        max_new_tokens=250,
        temperature=0.5
    )

    answer = tokenizer.decode(outputs[0], skip_special_tokens=True)
    return answer.strip()


with gr.Blocks() as demo:
    gr.Markdown("# 📚 Clean RAG Paper QA")

    file_input = gr.File(label="Upload Research PDF", file_types=[".pdf"])
    process_btn = gr.Button("Process PDF")
    status_output = gr.Textbox(label="Status")

    question_input = gr.Textbox(label="Ask a Question")
    ask_btn = gr.Button("Get Answer")
    answer_output = gr.Textbox(label="Answer")

    process_btn.click(process_pdf, inputs=file_input, outputs=status_output)
    ask_btn.click(ask_question, inputs=question_input, outputs=answer_output)

demo.launch()