app.py

import os
import tempfile
import gradio as gr
import spaces  # Required for GPU-enabled Spaces

from pinecone import Pinecone
from langchain_pinecone import PineconeVectorStore
from langchain_community.document_loaders import PyPDFLoader
from langchain_mistralai import ChatMistralAI
from langchain.chains import RetrievalQAWithSourcesChain
from langchain_text_splitters import RecursiveCharacterTextSplitter
from langchain.embeddings import HuggingFaceEmbeddings

# Initialize Pinecone (safe, does not use CUDA)
pc = Pinecone(api_key=os.getenv("PINECONE_API_KEY"))
INDEX_NAME = "ragreader"

# This function does all GPU work: embedding creation, document processing, and vector store population
@spaces.GPU
def process_documents(files):
    device = "cuda" if hasattr(__import__('torch'), 'cuda') and __import__('torch').cuda.is_available() else "cpu"
    embeddings = HuggingFaceEmbeddings(
        model_name="BAAI/bge-large-en-v1.5",
        model_kwargs={"device": device}
    )
    docs = []
    for file in files:
        with tempfile.NamedTemporaryFile(delete=False, suffix=".pdf") as tmp:
            file.seek(0)
            tmp.write(file.read())
            tmp.flush()
            loader = PyPDFLoader(tmp.name)
            docs.extend(loader.load())
        os.unlink(tmp.name)
    text_splitter = RecursiveCharacterTextSplitter(
        chunk_size=1000,
        chunk_overlap=200
    )
    split_docs = text_splitter.split_documents(docs)
    PineconeVectorStore.from_documents(
        documents=split_docs,
        embedding=embeddings,
        index_name=INDEX_NAME
    )
    return "Documents processed and stored."

# This function creates the QA chain (CPU only)
def init_qa_chain():
    # Embeddings must be created inside a GPU function, so we do not re-create here.
    # PineconeVectorStore uses the embeddings stored in Pinecone.
    llm = ChatMistralAI(
        model="mistral-tiny",
        temperature=0.3,
        mistral_api_key=os.getenv("MISTRAL_API_KEY")
    )
    # Pass None for embeddings since vectors are already in Pinecone
    vector_store = PineconeVectorStore(
        index_name=INDEX_NAME,
        embedding=None
    )
    return RetrievalQAWithSourcesChain.from_chain_type(
        llm=llm,
        chain_type="stuff",
        retriever=vector_store.as_retriever(search_kwargs={"k": 3}),
        return_source_documents=True
    )

# State: store the QA chain after processing
qa_chain = None

with gr.Blocks() as demo:
    gr.Markdown("## RAG Chatbot - PDF Reader")

    file_input = gr.File(file_types=[".pdf"], file_count="multiple", type="file", label="Upload PDFs")
    process_btn = gr.Button("Process Documents")
    process_output = gr.Textbox(label="Processing Status")

    chat_input = gr.Textbox(label="Ask a question about your documents")
    chat_btn = gr.Button("Submit Question")
    chat_output = gr.Textbox(label="Answer")
    source_output = gr.Textbox(label="Relevant Sources")

    def process_wrapper(files):
        global qa_chain
        if not files or len(files) == 0:
            return "Please upload at least one PDF."
        msg = process_documents(files)
        qa_chain = init_qa_chain()
        return msg

    def chat_with_docs(question):
        global qa_chain
        if not qa_chain:
            return "Please upload and process documents first.", ""
        if not question.strip():
            return "Please enter a question.", ""
        response = qa_chain.invoke({"question": question}, return_only_outputs=True)
        sources = "\n".join(
            f"{os.path.basename(doc.metadata.get('source', 'unknown'))} (Page {doc.metadata.get('page', 'N/A')})"
            for doc in response.get('source_documents', [])[:3]
        )
        return response.get('answer', "No answer found."), sources

    process_btn.click(fn=process_wrapper, inputs=[file_input], outputs=process_output)
    chat_btn.click(fn=chat_with_docs, inputs=chat_input, outputs=[chat_output, source_output])

if __name__ == "__main__":
    demo.launch()