app.py

import gradio as gr
from huggingface_hub import InferenceClient
import pandas as pd
import transformers
import torch
from sentence_transformers import SentenceTransformer, util
"""
For more information on `huggingface_hub` Inference API support, please check the docs: https://huggingface.co/docs/huggingface_hub/v0.22.2/en/guides/inference
"""
client = InferenceClient("HuggingFaceH4/zephyr-7b-beta")

# Load the SBERT model
sbert_model = SentenceTransformer('sentence-transformers/all-MiniLM-L6-v2')

# Function to initialize the Llama 3 8B model pipeline
def initiate_pipeline():
    model = "meta-llama/Meta-Llama-3-8B-Instruct"
    device = "cuda:0" if torch.cuda.is_available() else "cpu"
    return transformers.pipeline(
        "text-generation",
        model=model,
        model_kwargs={"torch_dtype": torch.bfloat16},
        device=device,
    )

# Initialize the model
llama_model = initiate_pipeline()

# Load the Q&A pairs from the CSV
qa_data = pd.read_csv("rag_juri_cv.csv")

# Function to retrieve the top 5 relevant Q&A pairs using Sentence-BERT
def retrieve_top_k(query, k=5):
    # Combine the questions from the CSV into a list
    questions = qa_data['QUESTION'].tolist()

    # Encode the questions and the query using Sentence-BERT
    question_embeddings = sbert_model.encode(questions, convert_to_tensor=True)
    query_embedding = sbert_model.encode(query, convert_to_tensor=True)

    # Compute cosine similarities between the query and all questions
    cosine_scores = util.pytorch_cos_sim(query_embedding, question_embeddings).flatten()

    # Get the indices of the top k most similar questions
    top_k_indices = torch.topk(cosine_scores, k=k).indices.cpu()

    # Retrieve the corresponding Q&A pairs
    top_k_qa = qa_data.iloc[top_k_indices]

    return top_k_qa

def chatbot(query):
    # Retrieve the top 5 relevant Q&A pairs
    top_k_qa = retrieve_top_k(query)

    # Generate the prefix, body, and suffix for the prompt
    prefix = """
    <|begin_of_text|><|start_header_id|>user<|end_header_id|>You are a chatbot specialized in answering questions about Juri Grosjean's CV.
    Please only use the information provided in the context to answer the question.
    Here is the question to answer:
    """ + query + "\n\n"

    context = "This is the context information to answer the question:\n"
    for index, row in top_k_qa.iterrows():
        context += f"Information {index}: {row['ANSWER']}\n\n"

    suffix = "<|eot_id|><|start_header_id|>assistant<|end_header_id|>"

    prompt = prefix + context + suffix

    # Generate a response
    outputs = llama_model(
        prompt,
        max_new_tokens=500,
        do_sample=True,
        temperature=0.6,
        top_p=0.9,
    )

    # Extract and return the chatbot's answer
    output = outputs[0]["generated_text"]
    return output.split("assistant")[-1].strip()


# Set up the Gradio interface
demo = gr.Interface(
        fn=chatbot,
        inputs=gr.Textbox(lines=5, placeholder="Ask a question about Juri Grosjean's CV"),
        outputs="text"
    )

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