app.py

import re
from transformers import AutoModelForCausalLM, AutoTokenizer, AutoConfig, AutoModel, StoppingCriteria, StoppingCriteriaList, TextIteratorStreamer
from vllm import LLM, SamplingParams
import torch
import gradio as gr
import json
import os
import shutil
import requests
import numpy as np
import pandas as pd
from threading import Thread
from FlagEmbedding import BGEM3FlagModel
from sklearn.metrics.pairwise import cosine_similarity

from transformers import AutoModelForSequenceClassification

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

#Importing the embedding model
embedding_model = BGEM3FlagModel('BAAI/bge-m3',  
                       use_fp16=True) # Setting use_fp16 to True speeds up computation with a slight performance degradation

embeddings = np.load("embeddings_albert_tchap.npy")
embeddings_data = pd.read_json("embeddings_albert_tchap.json")
embeddings_text = embeddings_data["text_with_context"].tolist()

#Importing the classifier/router (deberta)
classifier_model = AutoModelForSequenceClassification.from_pretrained("AgentPublic/chatrag-deberta")
tokenizer = AutoTokenizer.from_pretrained("AgentPublic/chatrag-deberta")

#Importing the actual generative LLM (llama-based)
model_name = "Pclanglais/Tchap"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.float16)
model = model.to('cuda:0')

system_prompt = "<|begin_of_text|><|start_header_id|>system<|end_header_id|>\n\nTu es Albert, l'agent conversationnel des services publics qui peut décrire des documents de référence ou aider à des tâches de rédaction<|eot_id|>"

#Function to guess whether we use the RAG or not.
def classification_chatrag(query):
  encoding = tokenizer(query, return_tensors="pt")
  encoding = {k: v.to(model_classifier.device) for k,v in encoding.items()}

  outputs = model_classifier(**encoding)

  logits = outputs.logits
  logits.shape

  # apply sigmoid + threshold
  sigmoid = torch.nn.Sigmoid()
  probs = sigmoid(logits.squeeze().cpu())
  predictions = np.zeros(probs.shape)

  # Extract the float value from the tensor
  float_value = round(probs.item()*100)

  if float_value > 50:
    status = True
    print("We activate RAG")
  else:
    status = False
    print("We remove RAG")
  return status

#Vector search over the database
def vector_search(sentence_query):

    query_embedding = embedding_model.encode(sentence_query, 
                            batch_size=12, 
                            max_length=256, # If you don't need such a long length, you can set a smaller value to speed up the encoding process.
                            )['dense_vecs']

    # Reshape the query embedding to fit the cosine_similarity function requirements
    query_embedding_reshaped = query_embedding.reshape(1, -1)
    
    # Compute cosine similarities
    similarities = cosine_similarity(query_embedding_reshaped, embeddings)
    
    # Find the index of the closest document (highest similarity)
    closest_doc_index = np.argmax(similarities)
    
    # Closest document's embedding
    closest_doc_embedding = embeddings_text[closest_doc_index]
    
    return closest_doc_embedding


class StopOnTokens(StoppingCriteria):
    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
        stop_ids = [29, 0]
        for stop_id in stop_ids:
            if input_ids[0][-1] == stop_id:
                return True
        return False


def predict(message, history):

    global source_text
    global assess_rag
    #For now, we only query the vector database once, at the start.
    if len(history) == 0:
        assess_rag = classification_chatrag(message)
        if assess_rag:
            source_text = vector_search(message)
    
    history_transformer_format = history + [[message, ""]]

    print(history_transformer_format)
    stop = StopOnTokens()

    messages = []
    id_message = 1
    total_message = len(history_transformer_format)
    for item in history_transformer_format:

        #Once we target the ongoing post we add the source.
        if id_message == total_message:
            if assess_rag:
                question = "<|start_header_id|>user<|end_header_id|>\n\n"+ item[0] + "\n\n### Source ###\n" + source_text
            else:
                question = "<|start_header_id|>user<|end_header_id|>\n\n"+ item[0]
        else:
            question = "<|start_header_id|>user<|end_header_id|>\n\n"+ item[0]
        answer = "<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"+item[1]
        result = "".join([question, answer])
        messages.append(result)
        id_message = id_message + 1

    messages = "".join(messages)

    print(messages)

    messages = system_prompt + messages

    model_inputs = tokenizer([messages], return_tensors="pt").to("cuda")
    streamer = TextIteratorStreamer(tokenizer, timeout=10., skip_prompt=True, skip_special_tokens=True)
    generate_kwargs = dict(
        model_inputs,
        streamer=streamer,
        max_new_tokens=1024,
        do_sample=False,
        top_p=0.95,
        temperature=0.4,
        stopping_criteria=StoppingCriteriaList([stop])
        )
    t = Thread(target=model.generate, kwargs=generate_kwargs)
    t.start()

    partial_message = ""
    for new_token in streamer:
        if new_token != '<':
            partial_message += new_token
            yield partial_message
    return messages

# Define the Gradio interface
title = "Tchap"
description = "Le chatbot du service public"
examples = [
    [
        "Qui peut bénéficier de l'AIP?",  # user_message
        0.7  # temperature
    ]
]


gr.ChatInterface(predict).launch()