app.py

import gradio as gr
from datasets import load_dataset

import os
import spaces
from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer, BitsAndBytesConfig
import torch
from threading import Thread
from sentence_transformers import SentenceTransformer
import numpy as np

token = os.environ["HF_TOKEN"]
ST = SentenceTransformer("mixedbread-ai/mxbai-embed-large-v1")

dataset = load_dataset("Yoxas/statistical_literacyv2")

data = dataset["train"]

# Convert the string embeddings to numerical arrays and ensure they are 2D
def convert_and_ensure_2d_embeddings(example):
    # Clean the embedding string
    embedding_str = example['embedding']
    embedding_str = embedding_str.replace('\n', ' ').replace('...', '')
    embedding_list = list(map(float, embedding_str.strip("[]").split()))
    embeddings = np.array(embedding_list, dtype=np.float32)
    # Ensure the embeddings are 2-dimensional
    if embeddings.ndim == 1:
        embeddings = embeddings.reshape(1, -1)
    return {'embedding': embeddings}

# Apply the function to ensure embeddings are 2-dimensional and of type float32
data = data.map(convert_and_ensure_2d_embeddings)

# Flatten embeddings if they are nested 2D arrays
def flatten_embeddings(example):
    embedding = example['embedding']
    if embedding.ndim == 2 and embedding.shape[0] == 1:
        embedding = embedding.flatten()
    return {'embedding': embedding}

data = data.map(flatten_embeddings)

# Ensure embeddings are in the correct shape for FAISS
embeddings = np.array(data['embedding'].tolist(), dtype=np.float32)

# Add FAISS index
data.add_faiss_index_from_external_arrays(embeddings)

model_id = "meta-llama/Meta-Llama-3-8B-Instruct"

# use quantization to lower GPU usage
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True, bnb_4bit_use_double_quant=True, bnb_4bit_quant_type="nf4", bnb_4bit_compute_dtype=torch.bfloat16
)

tokenizer = AutoTokenizer.from_pretrained(model_id, token=token)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    quantization_config=bnb_config,
    token=token
)
terminators = [
    tokenizer.eos_token_id,
    tokenizer.convert_tokens_to_ids("")
]

SYS_PROMPT = """You are an assistant for answering questions.
You are given the extracted parts of a long document and a question. Provide a conversational answer.
If you don't know the answer, just say "I do not know." Don't make up an answer."""

def search(query: str, k: int = 3):
    """a function that embeds a new query and returns the most probable results"""
    embedded_query = ST.encode(query)  # embed new query
    scores, retrieved_examples = data.get_nearest_examples(  # retrieve results
        "embedding", embedded_query,  # compare our new embedded query with the dataset embeddings
        k=k  # get only top k results
    )
    return scores, retrieved_examples

def format_prompt(prompt, retrieved_documents, k):
    """using the retrieved documents we will prompt the model to generate our responses"""
    PROMPT = f"Question:{prompt}\nContext:"
    for idx in range(k):
        PROMPT += f"{retrieved_documents['text'][idx]}\n"
    return PROMPT

@spaces.GPU(duration=150)
def talk(prompt, history):
    k = 1  # number of retrieved documents
    scores, retrieved_documents = search(prompt, k)
    formatted_prompt = format_prompt(prompt, retrieved_documents, k)
    formatted_prompt = formatted_prompt[:2000]  # to avoid GPU OOM
    messages = [{"role": "system", "content": SYS_PROMPT}, {"role": "user", "content": formatted_prompt}]
    # tell the model to generate
    input_ids = tokenizer.apply_chat_template(
        messages,
        add_generation_prompt=True,
        return_tensors="pt"
    ).to(model.device)
    outputs = model.generate(
        input_ids,
        max_new_tokens=1024,
        eos_token_id=terminators,
        do_sample=True,
        temperature=0.6,
        top_p=0.9,
    )
    streamer = TextIteratorStreamer(
        tokenizer, timeout=10.0, skip_prompt=True, skip_special_tokens=True
    )
    generate_kwargs = dict(
        input_ids=input_ids,
        streamer=streamer,
        max_new_tokens=1024,
        do_sample=True,
        top_p=0.95,
        temperature=0.75,
        eos_token_id=terminators,
    )
    t = Thread(target=model.generate, kwargs=generate_kwargs)
    t.start()

    outputs = []
    for text in streamer:
        outputs.append(text)
        print(outputs)
        yield "".join(outputs)

TITLE = "# RAG"

DESCRIPTION = """
A rag pipeline with a chatbot feature
Resources used to build this project :
* embedding model : https://huggingface.co/mixedbread-ai/mxbai-embed-large-v1
* dataset : https://huggingface.co/datasets/not-lain/wikipedia
* faiss docs : https://huggingface.co/docs/datasets/v2.18.0/en/package_reference/main_classes#datasets.Dataset.add_faiss_index 
* chatbot : https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct
"""

demo = gr.ChatInterface(
    fn=talk,
    chatbot=gr.Chatbot(
        show_label=True,
        show_share_button=True,
        show_copy_button=True,
        likeable=True,
        layout="bubble",
        bubble_full_width=False,
    ),
    theme="Soft",
    examples=[["what's anarchy ? "]],
    title=TITLE,
    description=DESCRIPTION,
)
demo.launch(debug=True)