app.py

import warnings
warnings.filterwarnings("ignore")

import os
import glob
import textwrap
import time

import langchain

### loaders
from langchain.document_loaders import PyPDFLoader, DirectoryLoader

### splits
from langchain.text_splitter import RecursiveCharacterTextSplitter

### prompts
from langchain import PromptTemplate, LLMChain

### vector stores
from langchain.vectorstores import FAISS

### models
from langchain.llms import HuggingFacePipeline
from langchain.embeddings import HuggingFaceInstructEmbeddings

### retrievers
from langchain.chains import RetrievalQA

import torch
import transformers
from transformers import (
    AutoTokenizer, AutoModelForCausalLM,
    BitsAndBytesConfig,
    pipeline
)
import gradio as gr
import locale
import time
from langchain.embeddings.huggingface import HuggingFaceEmbeddings

class CFG:
    # LLMs
    model_name = 'llama2-13b-chat' # wizardlm, llama2-7b-chat, llama2-13b-chat, mistral-7B
    temperature = 0
    top_p = 0.95
    repetition_penalty = 1.15

    # splitting
    split_chunk_size = 800
    split_overlap = 0

    # embeddings
    embeddings_model_repo = 'sentence-transformers/all-MiniLM-L6-v2'

    # similar passages
    k = 6

    # paths
    PDFs_path = './'
    Embeddings_path =  './faiss-hp-sentence-transformers'
    Output_folder = './rag-vectordb'

def get_model(model = CFG.model_name):

    print('\nDownloading model: ', model, '\n\n')

    if model == 'wizardlm':
        model_repo = 'TheBloke/wizardLM-7B-HF'

        tokenizer = AutoTokenizer.from_pretrained(model_repo)

        bnb_config = BitsAndBytesConfig(
            load_in_4bit = True,
            bnb_4bit_quant_type = "nf4",
            bnb_4bit_compute_dtype = torch.float16,
            bnb_4bit_use_double_quant = True,
        )

        model = AutoModelForCausalLM.from_pretrained(
            model_repo,
            quantization_config = bnb_config,
            device_map = 'auto',
            low_cpu_mem_usage = True
        )

        max_len = 1024

    elif model == 'llama2-7b-chat':
        model_repo = 'daryl149/llama-2-7b-chat-hf'

        tokenizer = AutoTokenizer.from_pretrained(model_repo, use_fast=True)

        bnb_config = BitsAndBytesConfig(
            load_in_4bit = True,
            bnb_4bit_quant_type = "nf4",
            bnb_4bit_compute_dtype = torch.float16,
            bnb_4bit_use_double_quant = True,
        )

        model = AutoModelForCausalLM.from_pretrained(
            model_repo,
            quantization_config = bnb_config,
            device_map = 'auto',
            low_cpu_mem_usage = True,
            trust_remote_code = True
        )

        max_len = 2048

    elif model == 'llama2-13b-chat':
        model_repo = 'daryl149/llama-2-13b-chat-hf'

        tokenizer = AutoTokenizer.from_pretrained(model_repo, use_fast=True)

        bnb_config = BitsAndBytesConfig(
            load_in_4bit = True,
            bnb_4bit_quant_type = "nf4",
            bnb_4bit_compute_dtype = torch.float16,
            bnb_4bit_use_double_quant = True,
        )

        model = AutoModelForCausalLM.from_pretrained(
            model_repo,
            quantization_config = bnb_config,

            low_cpu_mem_usage = True,
            trust_remote_code = True
        )

        max_len = 2048  #8192
        truncation=True,  # Explicitly enable truncation
        padding="max_len"  # Optional: pad to max_length

    elif model == 'mistral-7B':
        model_repo = 'mistralai/Mistral-7B-v0.1'

        tokenizer = AutoTokenizer.from_pretrained(model_repo)

        bnb_config = BitsAndBytesConfig(
            load_in_4bit = True,
            bnb_4bit_quant_type = "nf4",
            bnb_4bit_compute_dtype = torch.float16,
            bnb_4bit_use_double_quant = True,
        )

        model = AutoModelForCausalLM.from_pretrained(
            model_repo,
            quantization_config = bnb_config,
            device_map = 'auto',
            low_cpu_mem_usage = True,
        )

        max_len = 1024

    else:
        print("Not implemented model (tokenizer and backbone)")

    return tokenizer, model, max_len

def wrap_text_preserve_newlines(text, width=700):
    # Split the input text into lines based on newline characters
    lines = text.split('\n')

    # Wrap each line individually
    wrapped_lines = [textwrap.fill(line, width=width) for line in lines]

    # Join the wrapped lines back together using newline characters
    wrapped_text = '\n'.join(wrapped_lines)

    return wrapped_text


def process_llm_response(llm_response):
    ans = wrap_text_preserve_newlines(llm_response['result'])

    sources_used = ' \n'.join(
        [
            source.metadata['source'].split('/')[-1][:-4]
            + ' - page: '
            + str(source.metadata['page'])
            for source in llm_response['source_documents']
        ]
    )

    ans = ans + '\n\nSources: \n' + sources_used
    return ans

def llm_ans(query):
    start = time.time()

    llm_response = qa_chain.invoke(query)
    ans = process_llm_response(llm_response)

    end = time.time()

    time_elapsed = int(round(end - start, 0))
    time_elapsed_str = f'\n\nTime elapsed: {time_elapsed} s'
    return ans + time_elapsed_str

def predict(message, history):
     output = str(llm_ans(message)).replace("\n", "<br/>")
     return output




tokenizer, model, max_len = get_model(model = CFG.model_name)

pipe = pipeline(
    task = "text-generation",
    model = model,
    tokenizer = tokenizer,
    pad_token_id = tokenizer.eos_token_id,
#     do_sample = True,
    max_length = max_len,
    temperature = CFG.temperature,
    top_p = CFG.top_p,
    repetition_penalty = CFG.repetition_penalty
)

### langchain pipeline
llm = HuggingFacePipeline(pipeline = pipe)

loader = DirectoryLoader(
    CFG.PDFs_path,
    glob="./*.pdf",
    loader_cls=PyPDFLoader,
    show_progress=True,
    use_multithreading=True
)

documents = loader.load()
text_splitter = RecursiveCharacterTextSplitter(
    chunk_size = CFG.split_chunk_size,
    chunk_overlap = CFG.split_overlap
)

texts = text_splitter.split_documents(documents)

vectordb = FAISS.from_documents(
    texts,
    HuggingFaceEmbeddings(model_name='sentence-transformers/all-mpnet-base-v2')
)

### persist vector database
vectordb.save_local(f"{CFG.Output_folder}/faiss_index_rag")

retriever = vectordb.as_retriever(search_kwargs = {"k": CFG.k, "search_type" : "similarity"})

qa_chain = RetrievalQA.from_chain_type(
    llm = llm,
    chain_type = "stuff", # map_reduce, map_rerank, stuff, refine
    retriever = retriever,
    chain_type_kwargs = {"prompt": PROMPT},
    return_source_documents = True,
    verbose = False
)

prompt_template = """
Don't try to make up an answer, if you don't know just say that you don't know.
Answer in the same language the question was asked.
Use only the following pieces of context to answer the question at the end.

{context}

Question: {question}
Answer:"""


PROMPT = PromptTemplate(
    template = prompt_template,
    input_variables = ["context", "question"]
)


locale.getpreferredencoding = lambda: "UTF-8"

demo = gr.ChatInterface(
     predict,
     title = f' Open-Source LLM ({CFG.model_name}) Question Answering'
 )

demo.queue()
demo.launch()