app.py

import gradio as gr
from gpt4all import GPT4All
from huggingface_hub import hf_hub_download
import faiss
#from langchain_community.embeddings import HuggingFaceEmbeddings
from langchain_huggingface import HuggingFaceEmbeddings
import numpy as np
from pypdf import PdfReader
from gradio_pdf import PDF
from pdf2image import convert_from_path
from transformers import pipeline
from pathlib import Path

title = "Mistral-7B-Instruct-GGUF Run On CPU-Basic Free Hardware"

description = """
🔎 [Mistral AI's Mistral 7B Instruct v0.1](https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.1) [GGUF format model](https://huggingface.co/TheBloke/Mistral-7B-Instruct-v0.1-GGUF) , 4-bit quantization balanced quality gguf version, running on CPU. English Only (Also support other languages but the quality's not good). Using [GitHub - llama.cpp](https://github.com/ggerganov/llama.cpp) [GitHub - gpt4all](https://github.com/nomic-ai/gpt4all). 
🔨 Running on CPU-Basic free hardware. Suggest duplicating this space to run without a queue. 
Mistral does not support system prompt symbol (such as ```<<SYS>>```) now, input your system prompt in the first message if you need. Learn more: [Guardrailing Mistral 7B](https://docs.mistral.ai/usage/guardrailing). 
"""

"""
[Model From TheBloke/Mistral-7B-Instruct-v0.1-GGUF](https://huggingface.co/TheBloke/Mistral-7B-Instruct-v0.1-GGUF)
[Mistral-instruct-v0.1 System prompt](https://docs.mistral.ai/usage/guardrailing)
"""

model_path = "models"
model_name = "mistral-7b-instruct-v0.1.Q4_K_M.gguf"

hf_hub_download(repo_id="TheBloke/Mistral-7B-Instruct-v0.1-GGUF", filename=model_name, local_dir=model_path, local_dir_use_symlinks=False)

print("Start the model init process")
model = model = GPT4All(model_name, model_path, allow_download = False, device="cpu")


model.config["promptTemplate"] = "[INST] {0} [/INST]"
model.config["systemPrompt"] = "Tu es un assitant et tu dois répondre en français"
model._is_chat_session_activated = False

max_new_tokens = 2048

model_kwargs = {'device': 'cpu'}
encode_kwargs = {'normalize_embeddings': False}
embeddings = HuggingFaceEmbeddings(
    
    model_kwargs=model_kwargs,
    encode_kwargs=encode_kwargs
)

chunk_size = 2048

# creating a pdf reader object

reader = PdfReader("./resource/NGAP 01042024.pdf")

text = []
for p in np.arange(0, len(reader.pages), 1):
    page = reader.pages[int(p)]
    
    # extracting text from page
    text.append(page.extract_text())
    
text = ' '.join(text)

chunks = [text[i:i + chunk_size] for i in range(0, len(text), chunk_size)]


def get_text_embedding(text):

    return embeddings.embed_query(text)

text_embeddings = np.array([get_text_embedding(chunk) for chunk in chunks])

d = text_embeddings.shape[1]
index = faiss.IndexFlatL2(d)
index.add(text_embeddings)

print("Finish the model init process")

def qa(question: str) -> str:
    
    

    question_embeddings = np.array([get_text_embedding(question)])

    D, I = index.search(question_embeddings, k=2) # distance, index
    retrieved_chunk = [chunks[i] for i in I.tolist()[0]]

    prompt = f"""<s>[INST]
            Les informations contextuelles sont ci-dessous.
            ---------------------
            {retrieved_chunk}
            ---------------------
            [/INST]
            Compte tenu des informations contextuelles et non des connaissances préalables, répondez à la requête.  </s>
            [INST] Requête: {question} [/INST]
            Réponse: 
                """
    outputs = model.generate(prompt=prompt, temp=0.5, top_k = 40, top_p = 1, max_tokens = max_new_tokens)
    return "".join(outputs)


demo = gr.Interface(
    qa,
    [gr.Textbox(label="Question")#, PDF(label="Document")
    ],
    gr.Textbox()
)
if __name__ == "__main__":
    demo.queue(max_size=3).launch()