app.py

import os
import gradio as gr
from huggingface_hub import InferenceClient
from langchain_community.vectorstores import Chroma
from langchain_community.embeddings import HuggingFaceBgeEmbeddings
from langchain_community.document_loaders import PyPDFLoader, UnstructuredFileLoader, CSVLoader
from langchain.text_splitter import RecursiveCharacterTextSplitter

# Load Hugging Face API token
HF_API_TOKEN = os.getenv("HF_API_TOKEN")
if not HF_API_TOKEN:
    raise ValueError("Hugging Face API token is not set in environment variables.")

# Initialize Zephyr client
client = InferenceClient("HuggingFaceH4/zephyr-7b-beta", token=HF_API_TOKEN)

# Load documents based on file type
def load_documents(file_path):
    if file_path.endswith(".pdf"):
        loader = PyPDFLoader(file_path)
    elif file_path.endswith(".txt") or file_path.endswith(".docx"):
        loader = UnstructuredFileLoader(file_path)
    elif file_path.endswith(".csv"):
        loader = CSVLoader(file_path)
    else:
        raise ValueError("Unsupported file format")
    documents = loader.load()
    text_splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=100)
    return text_splitter.split_documents(documents)

# Create vector store
def create_vector_store(documents, persist_dir="vector_db"):
    embeddings = HuggingFaceBgeEmbeddings(
        model_name="BAAI/bge-large-en",
        model_kwargs={"device": "cpu"},
    )
    vector_store = Chroma.from_documents(documents, embeddings, persist_directory=persist_dir)
    return vector_store

# Initialize retriever and vector store
persist_dir = "vector_db"
retriever = None  # Will be dynamically updated

# Handle queries and uploads
def handle_query(message, history, system_message, max_tokens, temperature, top_p, file=None):
    global retriever
    if file:  # If a file is uploaded, process it
        documents = load_documents(file.name)
        vector_store = create_vector_store(documents, persist_dir)
        retriever = vector_store.as_retriever()
    if not retriever:
        return "No documents uploaded yet. Please upload a file first."
    
    # Retrieve relevant context
    relevant_docs = retriever.get_relevant_documents(message)
    context = "\n".join([doc.page_content for doc in relevant_docs])
    
    # Build the prompt
    prompt = f"""
    Use the following context to answer the user's question.
    Context:
    {context}
    
    Question:
    {message}
    Answer:"""
    
    response = ""
    for msg in client.chat_completion(
        messages=[{"role": "system", "content": system_message}, {"role": "user", "content": prompt}],
        max_tokens=max_tokens,
        stream=True,
        temperature=temperature,
        top_p=top_p,
    ):
        token = msg.choices[0].delta.content
        response += token
        yield response

# Gradio app setup
demo = gr.Interface(
    fn=handle_query,
    inputs=[
        gr.File(label="Upload Document"),
        gr.Textbox(value="You are a knowledgeable assistant.", label="System Message"),
        gr.Textbox(label="Enter Your Query", placeholder="Ask a question..."),
        gr.Slider(1, 2048, step=1, value=512, label="Max Tokens"),
        gr.Slider(0.1, 4.0, step=0.1, value=0.7, label="Temperature"),
        gr.Slider(0.1, 1.0, step=0.05, value=0.95, label="Top-p"),
    ],
    outputs="text",
    title="RAG with Zephyr-7B",
    description="Upload documents and ask questions using RAG.",
)

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