app.py

import os
import gradio as gr
from langchain.chains import RetrievalQA
from langchain_pinecone import Pinecone
from langchain_openai import ChatOpenAI
from langchain_community.llms import HuggingFacePipeline
from langchain_community.embeddings import HuggingFaceEmbeddings
#from dotenv import load_dotenv
import torch
from transformers import AutoModelForCausalLM, AutoModelForSequenceClassification, pipeline, AutoTokenizer
from huggingface_hub import login


login(token=os.getenv("HF_TOKEN"))

# Initialize Embedding Model
embedding_model = HuggingFaceEmbeddings(model_name="sentence-transformers/all-MiniLM-L6-v2")

# Pinecone Retriever
api_key = os.getenv("PINCE_CONE_LIGHT")
if api_key is None:
    raise ValueError("Pinecone API key missing.")
else:
    pc = Pinecone(pinecone_api_key=api_key, embedding=embedding_model, index_name='rag-rubic', namespace='vectors_lightmodel')
    retriever = pc.as_retriever()

# LLM Options
llm_options = {
    "OpenAI": "gpt-4o-mini",
    "Microsoft-Phi": "microsoft/Phi-3.5-mini-instruct",
    "DeepSeek-R1": "deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B",
    "Intel-tinybert": "Intel/dynamic_tinybert"
}

def load_llm(name, model_name):
    """Loads the selected LLM model only when needed."""
    if name == "OpenAI":
        openai_api_key = os.getenv("OPEN_AI_KEY")
        return ChatOpenAI(model='gpt-4o-mini', openai_api_key=openai_api_key)
    
    if "Phi" in name or "DeepSeek" in name:
        model = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto", torch_dtype=torch.float16)
        tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
        pipe = pipeline("text-generation", model=model, tokenizer=tokenizer, max_length=4096, eos_token_id=tokenizer.eos_token_id, return_full_text=False,
                        do_sample=False, num_return_sequences=1, max_new_tokens=50, temperature=0.1)
    elif "tinybert" in name:
        model = AutoModelForSequenceClassification.from_pretrained(model_name)
        tokenizer = AutoTokenizer.from_pretrained(model_name)
        pipe = pipeline("feature-extraction", model=model, tokenizer=tokenizer, truncation=True, padding=True, max_length=512)
    else:
        return None
    
    return HuggingFacePipeline(pipeline=pipe)

# Initialize default LLM
selected_llm = list(llm_options.keys())[0]
llm = load_llm(selected_llm, llm_options[selected_llm])

# Create QA Retrieval Chain
qa = RetrievalQA.from_llm(llm=llm, retriever=retriever)

# Chatbot function
def chatbot(selected_llm, user_input, chat_history):
    global llm
    if hasattr(llm, "pipeline"):  # Ensure llm has a pipeline
        current_model = llm.pipeline.model.name_or_path  # Get the model name
    else:
        current_model = None  # Handle cases where llm is not initialized
    
    if selected_llm != current_model:
        llm = load_llm(selected_llm, llm_options[selected_llm])

    response = qa.invoke({"query": user_input})
    answer = response.get("result", "No response received.")
    chat_history.append(("🧑‍💻 You", user_input))
    chat_history.append(("🤖 Bot", answer))
    return chat_history, ""


# Gradio Interface
with gr.Blocks() as demo:
    gr.Markdown("# 🤖 RAG-Powered Chatbot")
    llm_selector = gr.Dropdown(choices=list(llm_options.keys()), value=selected_llm, label="Choose an LLM")
    chat_history = gr.State([])
    chatbot_ui = gr.Chatbot()
    user_input = gr.Textbox(label="💬 Type your message and press Enter:")
    send_button = gr.Button("Send")
    
    send_button.click(chatbot, inputs=[llm_selector, user_input, chat_history], outputs=[chatbot_ui, user_input])
    user_input.submit(chatbot, inputs=[llm_selector, user_input, chat_history], outputs=[chatbot_ui, user_input])

demo.launch()