app.py

import torch
import transformers
from transformers import RagRetriever, RagSequenceForGeneration, AutoTokenizer, AutoModelForCausalLM
import gradio as gr

device = 'cuda' if torch.cuda.is_available() else 'cpu'


dataset_path = "./5k_index_data/my_knowledge_dataset"
index_path = "./5k_index_data/my_knowledge_dataset_hnsw_index.faiss"

tokenizer = AutoTokenizer.from_pretrained("facebook/rag-sequence-nq")
retriever = RagRetriever.from_pretrained("facebook/rag-sequence-nq", index_name="custom",
                                            passages_path = dataset_path,
                                            index_path = index_path,
                                            n_docs = 5)
rag_model = RagSequenceForGeneration.from_pretrained('facebook/rag-sequence-nq', retriever=retriever)
rag_model.retriever.init_retrieval()
rag_model.to(device)
model = AutoModelForCausalLM.from_pretrained('HuggingFaceH4/zephyr-7b-beta',
                              device_map = 'auto',
                              torch_dtype = torch.bfloat16,
                             )



def strip_title(title):
    if title.startswith('"'):
        title = title[1:]
    if title.endswith('"'):
        title = title[:-1]
    
    return title

# getting the correct format to input in gemma model
def input_format(query, context):
    sys_instruction = f'Context:\n {context} \n Given the following information, generate answer to the question. Provide links in the answer from the information to increase credebility.'
    message = f'Question: {query}'

    return f'<bos><start_of_turn>\n{sys_instruction}' + f' {message}<end_of_turn>\n'

# retrieving and generating  answer in one call
def retrieved_info(query, rag_model = rag_model, generating_model = model):
    # Tokenize Query
    retriever_input_ids = rag_model.retriever.question_encoder_tokenizer.batch_encode_plus(
        [query],
        return_tensors = 'pt',
        padding = True,
        truncation = True,
    )['input_ids'].to(device)

    # Retrieve Documents
    question_encoder_output = rag_model.rag.question_encoder(retriever_input_ids)
    question_encoder_pool_output = question_encoder_output[0]

    result = rag_model.retriever(
        retriever_input_ids,
        question_encoder_pool_output.cpu().detach().to(torch.float32).numpy(),
        prefix = rag_model.rag.generator.config.prefix,
        n_docs = rag_model.config.n_docs,
        return_tensors = 'pt',
    )

    # Preparing query and retrieved docs for model
    all_docs = rag_model.retriever.index.get_doc_dicts(result.doc_ids)
    retrieved_context = []
    for docs in all_docs:
        titles = [strip_title(title) for title in docs['title']]
        texts = docs['text']
        for title, text in zip(titles, texts):
            retrieved_context.append(f'{title}: {text}')

    generation_model_input = input_format(query, retrieved_context)

    # Generating answer using gemma model
    tokenizer = AutoTokenizer.from_pretrained("HuggingFaceH4/zephyr-7b-beta")
    input_ids = tokenizer(generation_model_input, return_tensors='pt').to(device)
    output = generating_model.generate(input_ids, max_new_tokens = 256)
    
    return tokenizer.decode(output[0])






def respond(
    message,
    history: list[tuple[str, str]],
    system_message,
    max_tokens ,
    temperature,
    top_p,
):
    if message:  # If there's a user query
        response = retrieved_info(message)  # Get the answer from your local FAISS and Q&A model
        return response

    # In case no message, return an empty string
    return ""



"""
For information on how to customize the ChatInterface, peruse the gradio docs: https://www.gradio.app/docs/chatinterface
"""
# Custom title and description
title = "🧠 Welcome to Your AI Knowledge Assistant"
description = """
Hi!! I am your loyal assistant. My functionality is based on the RAG model. I retrieve relevant information and provide answers based on that. Ask me any questions, and let me assist you.
My capabilities are limited because I am still in the development phase. I will do my best to assist you. SOOO LET'S BEGGINNNN......
"""

demo = gr.ChatInterface(
    respond,
    type = 'messages',
    additional_inputs=[
        gr.Textbox(value="You are a helpful and friendly assistant.", label="System message"),
        gr.Slider(minimum=1, maximum=2048, value=256, step=1, label="Max new tokens"),
        gr.Slider(minimum=0.1, maximum=4.0, value=0.7, step=0.1, label="Temperature"),
        gr.Slider(
            minimum=0.1,
            maximum=1.0,
            value=0.95,
            step=0.05,
            label="Top-p (nucleus sampling)",
        ),
    ],
    title=title,
    description=description,
    textbox=gr.Textbox(placeholder=["'What is the future of AI?' or 'App Development'"]),
    examples=[["✨Future of AI"], ["📱App Development"]],
    example_icons=["🤖", "📱"],
    theme="compact",
    submit_btn = True,
)


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