app.py

import os
import gradio as gr
import fitz  # PyMuPDF
import faiss
import numpy as np
from io import BytesIO
from sentence_transformers import SentenceTransformer
from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline
from langchain.text_splitter import RecursiveCharacterTextSplitter
from huggingface_hub import login

# Authenticate with Hugging Face
hf_token = os.environ.get("HUGGINGFACE_TOKEN")
if not hf_token:
    raise ValueError("⚠️ Please set the HUGGINGFACE_TOKEN environment variable.")
login(token=hf_token)

# Load embedding model
embed_model = SentenceTransformer("BAAI/bge-base-en-v1.5")

# Load Mistral without 4bit quantization (CPU-friendly)
model_id = "mistralai/Mistral-7B-Instruct-v0.1"
tokenizer = AutoTokenizer.from_pretrained(model_id, token=hf_token)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map={"": "cpu"},  # force CPU
    torch_dtype="auto",      # safe for CPU
    token=hf_token
)
llm = pipeline("text-generation", model=model, tokenizer=tokenizer)

# Globals
index = None
doc_texts = []

# Extract text from PDF or TXT
def extract_text(file):
    text = ""
    file_bytes = file.read()
    if file.name.endswith(".pdf"):
        pdf_stream = BytesIO(file_bytes)
        doc = fitz.open(stream=pdf_stream, filetype="pdf")
        for page in doc:
            text += page.get_text()
    elif file.name.endswith(".txt"):
        text = file_bytes.decode("utf-8")
    else:
        return "❌ Unsupported file type."
    return text

# Process the file, build FAISS index
def process_file(file):
    global index, doc_texts
    text = extract_text(file)
    if text.startswith("❌"):
        return text

    splitter = RecursiveCharacterTextSplitter(chunk_size=300, chunk_overlap=50)
    doc_texts = splitter.split_text(text)

    embeddings = embed_model.encode(doc_texts, convert_to_numpy=True)
    dim = embeddings.shape[1]
    index = faiss.IndexFlatL2(dim)
    index.add(embeddings)

    return "✅ File processed! You can now ask questions."

# Generate answer using context + LLM
def generate_answer(question):
    global index, doc_texts
    if index is None or not doc_texts:
        return "⚠️ Please upload and process a file first."

    question_emb = embed_model.encode([question], convert_to_numpy=True)
    _, I = index.search(question_emb, k=3)
    context = "\n".join([doc_texts[i] for i in I[0]])

    prompt = f"""[System: You are a helpful assistant. Answer based on the context.]

Context:
{context}

Question: {question}
Answer:"""

    response = llm(prompt, max_new_tokens=300, do_sample=True, temperature=0.7)
    return response[0]["generated_text"].split("Answer:")[-1].strip()

# Gradio UI
with gr.Blocks(title="RAG Chatbot (CPU Compatible)") as demo:
    gr.Markdown("## 📚 Upload PDF/TXT and Ask Questions using Mistral-7B")

    with gr.Row():
        file_input = gr.File(label="📁 Upload File (.pdf or .txt)", file_types=[".pdf", ".txt"])
        upload_status = gr.Textbox(label="Upload Status", interactive=False)

    with gr.Row():
        question_box = gr.Textbox(label="❓ Ask a Question", placeholder="What would you like to know?")
        answer_box = gr.Textbox(label="💬 Answer", interactive=False)

    file_input.change(fn=process_file, inputs=file_input, outputs=upload_status)
    question_box.submit(fn=generate_answer, inputs=question_box, outputs=answer_box)

demo.launch()