Upload 5 files

app.py

@@ -0,0 +1,400 @@
+from typing import List, Optional
+
+import gradio as gr
+from langchain_core.vectorstores import VectorStore
+
+from config import (
+    LLM_MODEL_REPOS,
+    EMBED_MODEL_REPOS,
+    SUBTITLES_LANGUAGES,
+    GENERATE_KWARGS,
+)
+
+from utils import (
+    load_llm_model,
+    load_embed_model,
+    load_documents_and_create_db,
+    user_message_to_chatbot,
+    update_user_message_with_context,
+    get_llm_response,
+    get_gguf_model_names,
+    add_new_model_repo,
+    clear_llm_folder,
+    clear_embed_folder,
+    get_memory_usage,
+)
+
+
+# ============ INTERFACE COMPONENT INITIALIZATION FUNCS ============
+
+def get_rag_settings(rag_mode: bool, render: bool = True):
+    k = gr.Radio(
+        choices=[1, 2, 3, 4, 5, 'all'],
+        value=2,
+        label='Number of relevant documents for search',
+        visible=rag_mode,
+        render=render,
+        )
+    score_threshold = gr.Slider(
+        minimum=0,
+        maximum=1,
+        value=0.5,
+        step=0.05,
+        label='relevance_scores_threshold',
+        visible=rag_mode,
+        render=render,
+        )
+    return k, score_threshold
+
+
+def get_user_message_with_context(text: str, rag_mode: bool) -> gr.component:
+    num_lines = len(text.split('\n'))
+    max_lines = 10
+    num_lines = max_lines if num_lines > max_lines else num_lines
+    return gr.Textbox(
+        text,
+        visible=rag_mode,
+        interactive=False,
+        label='User Message With Context',
+        lines=num_lines,
+        )
+
+
+def get_system_prompt_component(interactive: bool) -> gr.Textbox:
+    value = '' if interactive else 'System prompt is not supported by this model'
+    return gr.Textbox(value=value, label='System prompt', interactive=interactive)
+
+
+def get_generate_args(do_sample: bool) -> List[gr.component]:
+    generate_args = [
+        gr.Slider(minimum=0.1, maximum=3, value=GENERATE_KWARGS['temperature'], step=0.1, label='temperature', visible=do_sample),
+        gr.Slider(minimum=0.1, maximum=1, value=GENERATE_KWARGS['top_p'], step=0.01, label='top_p', visible=do_sample),
+        gr.Slider(minimum=1, maximum=50, value=GENERATE_KWARGS['top_k'], step=1, label='top_k', visible=do_sample),
+        gr.Slider(minimum=1, maximum=5, value=GENERATE_KWARGS['repeat_penalty'], step=0.1, label='repeat_penalty', visible=do_sample),
+    ]
+    return generate_args
+
+
+def get_rag_mode_component(db: Optional[VectorStore]) -> gr.Checkbox:
+    value = visible = db is not None
+    return gr.Checkbox(value=value, label='RAG Mode', scale=1, visible=visible)
+
+
+# ================ LOADING AND INITIALIZING MODELS ========================
+
+start_llm_model, start_support_system_role, load_log = load_llm_model(LLM_MODEL_REPOS[0], 'gemma-2-2b-it-Q8_0.gguf')
+start_embed_model, load_log = load_embed_model(EMBED_MODEL_REPOS[0])
+
+
+
+# ================== APPLICATION WEB INTERFACE ============================
+
+theme = gr.themes.Base(primary_hue='green', secondary_hue='yellow', neutral_hue='zinc').set(
+    loader_color='rgb(0, 255, 0)',
+    slider_color='rgb(0, 200, 0)',
+    body_text_color_dark='rgb(0, 200, 0)',
+    button_secondary_background_fill_dark='green',
+)
+css = '''.gradio-container {width: 60% !important}'''
+
+with gr.Blocks(theme=theme, css=css) as interface:
+
+    # ==================== GRADIO STATES ===============================
+
+    documents = gr.State([])
+    db = gr.State(None)
+    user_message_with_context = gr.State('')
+    support_system_role = gr.State(start_support_system_role)
+    llm_model_repos = gr.State(LLM_MODEL_REPOS)
+    embed_model_repos = gr.State(EMBED_MODEL_REPOS)
+    llm_model = gr.State(start_llm_model)
+    embed_model = gr.State(start_embed_model)
+
+
+
+    # ==================== BOT PAGE =================================
+
+    with gr.Tab(label='Chatbot'):
+        with gr.Row():
+            with gr.Column(scale=3):
+                chatbot = gr.Chatbot(
+                    show_copy_button=True,
+                    bubble_full_width=False,
+                    height=480,
+                )
+                user_message = gr.Textbox(label='User')
+
+                with gr.Row():
+                    user_message_btn = gr.Button('Send')
+                    stop_btn = gr.Button('Stop')
+                    clear_btn = gr.Button('Clear')
+
+            # ------------- GENERATION PARAMETERS -------------------
+
+            with gr.Column(scale=1, min_width=80):
+                with gr.Group():
+                    gr.Markdown('History size')
+                    history_len = gr.Slider(
+                        minimum=0,
+                        maximum=5,
+                        value=0,
+                        step=1,
+                        info='Number of previous messages taken into account in history',
+                        label='history_len',
+                        show_label=False,
+                        )
+
+                    with gr.Group():
+                        gr.Markdown('Generation parameters')
+                        do_sample = gr.Checkbox(
+                            value=False,
+                            label='do_sample',
+                            info='Activate random sampling',
+                            )
+                        generate_args = get_generate_args(do_sample.value)
+                        do_sample.change(
+                            fn=get_generate_args,
+                            inputs=do_sample,
+                            outputs=generate_args,
+                            show_progress=False,
+                            )
+
+        rag_mode = get_rag_mode_component(db=db.value)
+        k, score_threshold = get_rag_settings(rag_mode=rag_mode.value, render=False)
+        rag_mode.change(
+            fn=get_rag_settings,
+            inputs=[rag_mode],
+            outputs=[k, score_threshold],
+            )
+        with gr.Row():
+            k.render()
+            score_threshold.render()
+
+        # ---------------- SYSTEM PROMPT AND USER MESSAGE -----------
+
+        with gr.Accordion('Prompt', open=True):
+            system_prompt = get_system_prompt_component(interactive=support_system_role.value)
+            user_message_with_context = get_user_message_with_context(text='', rag_mode=rag_mode.value)
+
+        # ---------------- SEND, CLEAR AND STOP BUTTONS ------------
+
+        generate_event = gr.on(
+            triggers=[user_message.submit, user_message_btn.click],
+            fn=user_message_to_chatbot,
+            inputs=[user_message, chatbot],
+            outputs=[user_message, chatbot],
+            queue=False,
+        ).then(
+            fn=update_user_message_with_context,
+            inputs=[chatbot, rag_mode, db, k, score_threshold],
+            outputs=[user_message_with_context],
+        ).then(
+            fn=get_user_message_with_context,
+            inputs=[user_message_with_context, rag_mode],
+            outputs=[user_message_with_context],
+        ).then(
+            fn=get_llm_response,
+            inputs=[chatbot, llm_model, user_message_with_context, rag_mode, system_prompt,
+                    support_system_role, history_len, do_sample, *generate_args],
+            outputs=[chatbot],
+        )
+
+        stop_btn.click(
+            fn=None,
+            inputs=None,
+            outputs=None,
+            cancels=generate_event,
+            queue=False,
+        )
+
+        clear_btn.click(
+            fn=lambda: (None, ''),
+            inputs=None,
+            outputs=[chatbot, user_message_with_context],
+            queue=False,
+            )
+
+
+
+    # ================= FILE DOWNLOAD PAGE =========================
+
+    with gr.Tab(label='Load documents'):
+        with gr.Row(variant='compact'):
+            upload_files = gr.File(file_count='multiple', label='Loading text files')
+            web_links = gr.Textbox(lines=6, label='Links to Web sites or YouTube')
+
+        with gr.Row(variant='compact'):
+            chunk_size = gr.Slider(50, 2000, value=500, step=50, label='Chunk size')
+            chunk_overlap = gr.Slider(0, 200, value=20, step=10, label='Chunk overlap')
+
+            subtitles_lang = gr.Radio(
+                SUBTITLES_LANGUAGES,
+                value=SUBTITLES_LANGUAGES[0],
+                label='YouTube subtitle language',
+                )
+
+        load_documents_btn = gr.Button(value='Upload documents and initialize database')
+        load_docs_log = gr.Textbox(label='Status of loading and splitting documents', interactive=False)
+
+        load_documents_btn.click(
+            fn=load_documents_and_create_db,
+            inputs=[upload_files, web_links, subtitles_lang, chunk_size, chunk_overlap, embed_model],
+            outputs=[documents, db, load_docs_log],
+        ).success(
+            fn=get_rag_mode_component,
+            inputs=[db],
+            outputs=[rag_mode],
+        )
+
+        gr.HTML("""<h3 style='text-align: center'>
+        <a href="https://github.com/sergey21000/chatbot-rag" target='_blank'>GitHub Repository</a></h3>
+        """)
+
+
+
+    # ================= VIEW PAGE FOR ALL DOCUMENTS =================
+
+    with gr.Tab(label='View documents'):
+        view_documents_btn = gr.Button(value='Show downloaded text chunks')
+        view_documents_textbox = gr.Textbox(
+            lines=1,
+            placeholder='To view chunks, load documents in the Load documents tab',
+            label='Uploaded chunks',
+            )
+        sep = '=' * 20
+        view_documents_btn.click(
+            lambda documents: f'\n{sep}\n\n'.join([doc.page_content for doc in documents]),
+            inputs=[documents],
+            outputs=[view_documents_textbox],
+        )
+
+
+    # ============== GGUF MODELS DOWNLOAD PAGE =====================
+
+    with gr.Tab('Load LLM model'):
+        new_llm_model_repo = gr.Textbox(
+            value='',
+            label='Add repository',
+            placeholder='Link to repository of HF models in GGUF format',
+            )
+        new_llm_model_repo_btn = gr.Button('Add repository')
+        curr_llm_model_repo = gr.Dropdown(
+            choices=LLM_MODEL_REPOS,
+            value=None,
+            label='HF Model Repository',
+            )
+        curr_llm_model_path = gr.Dropdown(
+            choices=[],
+            value=None,
+            label='GGUF model file',
+            )
+        load_llm_model_btn = gr.Button('Loading and initializing model')
+        load_llm_model_log = gr.Textbox(
+            value=f'Model {LLM_MODEL_REPOS[0]} loaded at application startup',
+            label='Model loading status',
+            lines=6,
+            )
+
+        with gr.Group():
+            gr.Markdown('Free up disk space by deleting all models except the currently selected one')
+            clear_llm_folder_btn = gr.Button('Clear folder')
+
+        new_llm_model_repo_btn.click(
+            fn=add_new_model_repo,
+            inputs=[new_llm_model_repo, llm_model_repos],
+            outputs=[curr_llm_model_repo, load_llm_model_log],
+        ).success(
+            fn=lambda: '',
+            inputs=None,
+            outputs=[new_llm_model_repo],
+        )
+
+        curr_llm_model_repo.change(
+            fn=get_gguf_model_names,
+            inputs=[curr_llm_model_repo],
+            outputs=[curr_llm_model_path],
+        )
+
+        load_llm_model_btn.click(
+            fn=load_llm_model,
+            inputs=[curr_llm_model_repo, curr_llm_model_path],
+            outputs=[llm_model, support_system_role, load_llm_model_log],
+            queue=True,
+        ).success(
+            fn=lambda log: log + get_memory_usage(),
+            inputs=[load_llm_model_log],
+            outputs=[load_llm_model_log],
+        ).then(
+            fn=get_system_prompt_component,
+            inputs=[support_system_role],
+            outputs=[system_prompt],
+        )
+
+        clear_llm_folder_btn.click(
+            fn=clear_llm_folder,
+            inputs=[curr_llm_model_path],
+            outputs=None,
+        ).success(
+            fn=lambda model_path: f'Models other than {model_path} removed',
+            inputs=[curr_llm_model_path],
+            outputs=None,
+        )
+
+
+    # ============== EMBEDDING MODELS DOWNLOAD PAGE =============
+
+    with gr.Tab('Load embed model'):
+        new_embed_model_repo = gr.Textbox(
+            value='',
+            label='Add repository',
+            placeholder='Link to HF model repository',
+            )
+        new_embed_model_repo_btn = gr.Button('Add repository')
+        curr_embed_model_repo = gr.Dropdown(
+            choices=EMBED_MODEL_REPOS,
+            value=None,
+            label='HF model repository',
+            )
+
+        load_embed_model_btn = gr.Button('Loading and initializing model')
+        load_embed_model_log = gr.Textbox(
+            value=f'Model {EMBED_MODEL_REPOS[0]} loaded at application startup',
+            label='Model loading status',
+            lines=7,
+            )
+        with gr.Group():
+            gr.Markdown('Free up disk space by deleting all models except the currently selected one')
+            clear_embed_folder_btn = gr.Button('Clear folder')
+
+        new_embed_model_repo_btn.click(
+            fn=add_new_model_repo,
+            inputs=[new_embed_model_repo, embed_model_repos],
+            outputs=[curr_embed_model_repo, load_embed_model_log],
+        ).success(
+            fn=lambda: '',
+            inputs=None,
+            outputs=new_embed_model_repo,
+        )
+
+        load_embed_model_btn.click(
+            fn=load_embed_model,
+            inputs=[curr_embed_model_repo],
+            outputs=[embed_model, load_embed_model_log],
+        ).success(
+            fn=lambda log: log + get_memory_usage(),
+            inputs=[load_embed_model_log],
+            outputs=[load_embed_model_log],
+        )
+
+        clear_embed_folder_btn.click(
+            fn=clear_embed_folder,
+            inputs=[curr_embed_model_repo],
+            outputs=None,
+        ).success(
+            fn=lambda model_repo: f'Models other than {model_repo} removed',
+            inputs=[curr_embed_model_repo],
+            outputs=None,
+        )
+
+
+interface.launch(server_name='0.0.0.0', server_port=7860)  # debug=True

config.py

@@ -0,0 +1,105 @@
+from pathlib import Path
+
+# document loaders
+from langchain_community.document_loaders import (
+    CSVLoader,
+    PDFMinerLoader,
+    PyPDFLoader,
+    TextLoader,
+    UnstructuredHTMLLoader,
+    UnstructuredMarkdownLoader,
+    UnstructuredPowerPointLoader,
+    UnstructuredWordDocumentLoader,
+    WebBaseLoader,
+    YoutubeLoader,
+    DirectoryLoader,
+)
+
+
+# langchain classes for extracting text from various sources
+LOADER_CLASSES = {
+    '.csv': CSVLoader,
+    '.doc': UnstructuredWordDocumentLoader,
+    '.docx': UnstructuredWordDocumentLoader,
+    '.html': UnstructuredHTMLLoader,
+    '.md': UnstructuredMarkdownLoader,
+    '.pdf': PDFMinerLoader,
+    '.ppt': UnstructuredPowerPointLoader,
+    '.pptx': UnstructuredPowerPointLoader,
+    '.txt': TextLoader,
+    'web': WebBaseLoader,
+    'directory': DirectoryLoader,
+    'youtube': YoutubeLoader,
+}
+
+# languages ​​for youtube subtitles
+SUBTITLES_LANGUAGES = ['ru', 'en']
+
+# prom template subject to context
+CONTEXT_TEMPLATE = '''Ответь на вопрос при условии контекста.
+
+Контекст:
+{context}
+
+Вопрос:
+{user_message}
+
+Ответ:'''
+
+# dictionary for text generation config
+GENERATE_KWARGS = dict(
+    temperature=0.2,
+    top_p=0.95,
+    top_k=40,
+    repeat_penalty=1.0,
+    )
+
+# paths to LLM and embeddings models 
+LLM_MODELS_PATH = Path('models')
+EMBED_MODELS_PATH = Path('embed_models')
+LLM_MODELS_PATH.mkdir(exist_ok=True)
+EMBED_MODELS_PATH.mkdir(exist_ok=True)
+
+# available when running the LLM application models in GGUF format
+LLM_MODEL_REPOS = [
+    # https://huggingface.co/bartowski/gemma-2-2b-it-GGUF
+    'bartowski/gemma-2-2b-it-GGUF',
+    # https://huggingface.co/bartowski/Qwen2.5-3B-Instruct-GGUF
+    'bartowski/Qwen2.5-3B-Instruct-GGUF',
+    # https://huggingface.co/bartowski/Qwen2.5-1.5B-Instruct-GGUF
+    'bartowski/Qwen2.5-1.5B-Instruct-GGUF',
+    # https://huggingface.co/bartowski/openchat-3.6-8b-20240522-GGUF
+    'bartowski/openchat-3.6-8b-20240522-GGUF',
+    # https://huggingface.co/bartowski/Mistral-7B-Instruct-v0.3-GGUF
+    'bartowski/Mistral-7B-Instruct-v0.3-GGUF',
+    # https://huggingface.co/bartowski/Llama-3.2-3B-Instruct-GGUF
+    'bartowski/Llama-3.2-3B-Instruct-GGUF',
+]
+
+# Embedding models available at application startup
+EMBED_MODEL_REPOS = [
+    # https://huggingface.co/sergeyzh/rubert-tiny-turbo  # 117 MB
+    'sergeyzh/rubert-tiny-turbo',
+    # https://huggingface.co/cointegrated/rubert-tiny2  # 118 MB
+    'cointegrated/rubert-tiny2',
+    # https://huggingface.co/cointegrated/LaBSE-en-ru  # 516 MB
+    'cointegrated/LaBSE-en-ru',
+    # https://huggingface.co/sergeyzh/LaBSE-ru-turbo  # 513 MB
+    'sergeyzh/LaBSE-ru-turbo',
+    # https://huggingface.co/intfloat/multilingual-e5-large  # 2.24 GB
+    'intfloat/multilingual-e5-large',
+    # https://huggingface.co/intfloat/multilingual-e5-base  # 1.11 GB
+    'intfloat/multilingual-e5-base',
+    # https://huggingface.co/intfloat/multilingual-e5-small  # 471 MB
+    'intfloat/multilingual-e5-small',
+    # https://huggingface.co/intfloat/multilingual-e5-large-instruct  # 1.12 GB
+    'intfloat/multilingual-e5-large-instruct',
+    # https://huggingface.co/sentence-transformers/all-mpnet-base-v2  # 438 MB
+    'sentence-transformers/all-mpnet-base-v2',
+    # https://huggingface.co/sentence-transformers/paraphrase-multilingual-mpnet-base-v2  # 1.11 GB
+    'sentence-transformers/paraphrase-multilingual-mpnet-base-v2',
+    # https://huggingface.co/ai-forever?search_models=ruElectra  # 356 MB
+    'ai-forever/ruElectra-medium',
+    # https://huggingface.co/ai-forever/sbert_large_nlu_ru  # 1.71 GB
+    'ai-forever/sbert_large_nlu_ru',
+]

requirements-base.txt

@@ -0,0 +1,9 @@
+gradio==4.44.1
+huggingface-hub==0.24.7
+langchain==0.3.1
+langchain-community==0.3.1
+langchain-huggingface==0.1.0
+pdfminer.six==20240706
+youtube-transcript-api==0.6.2
+psutil==6.0.0
+faiss-cpu==1.8.0.post1

requirements-cpu.txt

@@ -0,0 +1,5 @@
+--extra-index-url https://download.pytorch.org/whl/cpu
+--extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cpu
+torch==2.4.1
+llama_cpp_python==0.2.88
+-r requirements-base.txt

utils.py

@@ -0,0 +1,495 @@
+import csv
+from pathlib import Path
+from shutil import rmtree
+from typing import List, Tuple, Dict, Union, Optional, Any, Iterable
+from tqdm import tqdm
+
+import psutil
+import requests
+from requests.exceptions import MissingSchema
+
+import torch
+import gradio as gr
+
+from llama_cpp import Llama
+from youtube_transcript_api import YouTubeTranscriptApi, NoTranscriptFound, TranscriptsDisabled
+from huggingface_hub import hf_hub_download, list_repo_tree, list_repo_files, repo_info, repo_exists, snapshot_download
+
+from langchain.text_splitter import RecursiveCharacterTextSplitter, CharacterTextSplitter
+from langchain_community.vectorstores import FAISS
+from langchain_huggingface import HuggingFaceEmbeddings
+
+# imports for annotations
+from langchain.docstore.document import Document
+from langchain_core.embeddings import Embeddings
+from langchain_core.vectorstores import VectorStore
+
+from config import (
+    LLM_MODELS_PATH,
+    EMBED_MODELS_PATH,
+    GENERATE_KWARGS,
+    LOADER_CLASSES,
+    CONTEXT_TEMPLATE,
+)
+
+
+# type annotations
+CHAT_HISTORY = List[Tuple[Optional[str], Optional[str]]]
+LLM_MODEL_DICT = Dict[str, Llama]
+EMBED_MODEL_DICT = Dict[str, Embeddings]
+
+
+# ===================== ADDITIONAL FUNCS =======================
+
+# getting the amount of free memory on disk, CPU and GPU
+def get_memory_usage() -> str:
+    print_memory = ''
+
+    memory_type = 'Disk'
+    psutil_stats = psutil.disk_usage('.')
+    memory_total = psutil_stats.total / 1024**3
+    memory_usage = psutil_stats.used / 1024**3
+    print_memory += f'{memory_type} Menory Usage: {memory_usage:.2f} / {memory_total:.2f} GB\n'
+
+    memory_type = 'CPU'
+    psutil_stats = psutil.virtual_memory()
+    memory_total = psutil_stats.total / 1024**3
+    memory_usage =  memory_total - (psutil_stats.available / 1024**3)
+    print_memory += f'{memory_type} Menory Usage: {memory_usage:.2f} / {memory_total:.2f} GB\n'
+
+    if torch.cuda.is_available():
+        memory_type = 'GPU'
+        memory_free, memory_total = torch.cuda.mem_get_info()
+        memory_usage = memory_total - memory_free
+        print_memory += f'{memory_type} Menory Usage: {memory_usage / 1024**3:.2f} / {memory_total:.2f} GB\n'
+
+    print_memory = f'---------------\n{print_memory}---------------'
+    return print_memory
+
+
+# clearing the list of documents
+def clear_documents(documents: Iterable[Document]) -> Iterable[Document]:
+    def clear_text(text: str) -> str:
+        lines = text.split('\n')
+        lines = [line for line in lines if len(line.strip()) > 2]
+        text = '\n'.join(lines).strip()
+        return text
+
+    output_documents = []
+    for document in documents:
+        text = clear_text(document.page_content)
+        if len(text) > 10:
+            document.page_content = text
+            output_documents.append(document)
+    return output_documents
+
+
+# ===================== INTERFACE FUNCS =============================
+
+
+# ------------- LLM AND EMBEDDING MODELS LOADING ------------------------
+
+# функция для загрузки файла по URL ссылке и отображением прогресс баров tqdm и gradio
+def download_file(file_url: str, file_path: Union[str, Path]) -> None:
+    response = requests.get(file_url, stream=True)
+    if response.status_code != 200:
+        raise Exception(f'The file is not available for download at the link: {file_url}')
+    total_size = int(response.headers.get('content-length', 0))
+    progress_tqdm = tqdm(desc='Loading GGUF file', total=total_size, unit='iB', unit_scale=True)
+    progress_gradio = gr.Progress()
+    completed_size = 0
+    with open(file_path, 'wb') as file:
+        for data in response.iter_content(chunk_size=4096):
+            size = file.write(data)
+            progress_tqdm.update(size)
+            completed_size += size
+            desc = f'Loading GGUF file, {completed_size/1024**3:.3f}/{total_size/1024**3:.3f} GB'
+            progress_gradio(completed_size/total_size, desc=desc)
+
+
+# loading and initializing the GGUF model
+def load_llm_model(model_repo: str, model_file: str) -> Tuple[LLM_MODEL_DICT, str, str]:
+    llm_model = None
+    load_log = ''
+    support_system_role = False
+
+    if isinstance(model_file, list):
+        load_log += 'No model selected\n'
+        return llm_model, load_log
+    if '(' in model_file:
+        model_file = model_file.split('(')[0].rstrip()
+
+    progress = gr.Progress()
+    progress(0.3, desc='Step 1/2: Download the GGUF file')
+    model_path = LLM_MODELS_PATH / model_file
+
+    if model_path.is_file():
+        load_log += f'Model {model_file} already loaded, reinitializing\n'
+    else:
+        try:
+            gguf_url = f'https://huggingface.co/{model_repo}/resolve/main/{model_file}'
+            download_file(gguf_url, model_path)
+            load_log += f'Model {model_file} loaded\n'
+        except Exception as ex:
+            model_path = ''
+            load_log += f'Error loading model, error code:\n{ex}\n'
+
+    if model_path:
+        progress(0.7, desc='Step 2/2: Initialize the model')
+        try:
+            llm_model = Llama(model_path=str(model_path), n_gpu_layers=-1, verbose=False)
+            support_system_role = 'System role not supported' not in llm_model.metadata['tokenizer.chat_template']
+            load_log += f'Model {model_file} initialized, max context size is {llm_model.n_ctx()} tokens\n'
+        except Exception as ex:
+            load_log += f'Error initializing model, error code:\n{ex}\n'
+
+    llm_model = {'model': llm_model}
+    return llm_model, support_system_role, load_log
+
+
+# loading and initializing the embedding model
+def load_embed_model(model_repo: str) -> Tuple[Dict[str, HuggingFaceEmbeddings], str]:
+    embed_model = None
+    load_log = ''
+
+    if isinstance(model_repo, list):
+        load_log = 'No model selected'
+        return embed_model, load_log
+
+    progress = gr.Progress()
+    folder_name = model_repo.replace('/', '_')
+    folder_path = EMBED_MODELS_PATH / folder_name
+    if Path(folder_path).is_dir():
+        load_log += f'Reinitializing model {model_repo} \n'
+    else:
+        progress(0.5, desc='Step 1/2: Download model repository')
+        snapshot_download(
+            repo_id=model_repo,
+            local_dir=folder_path,
+            ignore_patterns='*.h5',
+        )
+        load_log += f'Model {model_repo} loaded\n'
+
+    progress(0.7, desc='Шаг 2/2: Инициализация модели')
+    model_kwargs = {'device': 'cuda' if torch.cuda.is_available() else 'cpu'}
+    embed_model = HuggingFaceEmbeddings(
+        model_name=str(folder_path), 
+        model_kwargs=model_kwargs,
+        # encode_kwargs={'normalize_embeddings': True},
+        )
+    load_log += f'Embeddings model {model_repo} initialized\n'
+    load_log += f'Please upload documents and initialize database again\n'
+    embed_model = {'embed_model': embed_model}
+    return embed_model, load_log
+
+
+# adding a new HF repository new_model_repo to the current list of model_repos
+def add_new_model_repo(new_model_repo: str, model_repos: List[str]) -> Tuple[gr.Dropdown, str]:
+    load_log = ''
+    repo = new_model_repo.strip()
+    if repo:
+        repo = repo.split('/')[-2:]
+        if len(repo) == 2:
+            repo = '/'.join(repo).split('?')[0]
+            if repo_exists(repo) and repo not in model_repos:
+                model_repos.insert(0, repo)
+                load_log += f'Model repository {repo} successfully added\n'
+            else:
+                load_log += 'Invalid HF repository name or model already in the list\n'
+        else:
+            load_log += 'Invalid link to HF repository\n'
+    else:
+        load_log += 'Empty line in HF repository field\n'
+    model_repo_dropdown = gr.Dropdown(choices=model_repos, value=model_repos[0])
+    return model_repo_dropdown, load_log
+
+
+# get list of GGUF models from HF repository
+def get_gguf_model_names(model_repo: str) -> gr.Dropdown:
+    repo_files = list(list_repo_tree(model_repo))
+    repo_files = [file for file in repo_files if file.path.endswith('.gguf')]
+    model_paths = [f'{file.path} ({file.size / 1000 ** 3:.2f}G)' for file in repo_files]
+    model_paths_dropdown = gr.Dropdown(
+        choices=model_paths,
+        value=model_paths[0],
+        label='GGUF model file',
+        )
+    return model_paths_dropdown
+
+
+# delete model files and folders to clear space except for the current model gguf_filename
+def clear_llm_folder(gguf_filename: str) -> None:
+    if gguf_filename is None:
+        gr.Info(f'The name of the model file that does not need to be deleted is not selected.')
+        return
+    if '(' in gguf_filename:
+        gguf_filename = gguf_filename.split('(')[0].rstrip()
+    for path in LLM_MODELS_PATH.iterdir():
+        if path.name == gguf_filename:
+            continue
+        if path.is_file():
+            path.unlink(missing_ok=True)
+    gr.Info(f'All files removed from directory {LLM_MODELS_PATH} except {gguf_filename}')
+
+
+# delete model folders to clear space except for the current model model_folder_name
+def clear_embed_folder(model_repo: str) -> None:
+    if model_repo is None:
+        gr.Info(f'The name of the model that does not need to be deleted is not selected.')
+        return
+    model_folder_name = model_repo.replace('/', '_')
+    for path in EMBED_MODELS_PATH.iterdir():
+        if path.name == model_folder_name:
+            continue
+        if path.is_dir():
+            rmtree(path, ignore_errors=True)
+    gr.Info(f'All directories have been removed from the {EMBED_MODELS_PATH} directory except {model_folder_name}')
+
+
+# ------------------------ YOUTUBE ------------------------
+
+# function to check availability of subtitles, if manual or automatic are available - returns True and logs
+# if subtitles are not available - returns False and logs
+def check_subtitles_available(yt_video_link: str, target_lang: str) -> Tuple[bool, str]:
+    video_id = yt_video_link.split('watch?v=')[-1].split('&')[0]
+    load_log = ''
+    available = True
+    try:
+        transcript_list = YouTubeTranscriptApi.list_transcripts(video_id)
+        try:
+            transcript = transcript_list.find_transcript([target_lang])
+            if transcript.is_generated:
+                load_log += f'Automatic subtitles will be loaded, manual ones are not available for video {yt_video_link}\n'
+            else:
+                load_log += f'Manual subtitles will be downloaded for the video {yt_video_link}\n'
+        except NoTranscriptFound:
+            load_log += f'Subtitle language {target_lang} is not available for video {yt_video_link}\n'
+            available = False
+    except TranscriptsDisabled:
+        load_log += f'No subtitles for video {yt_video_link}\n'
+        available = False
+    return available, load_log
+
+
+# ------------- UPLOADING DOCUMENTS FOR RAG ------------------------
+
+# extract documents (in langchain Documents format) from downloaded files
+def load_documents_from_files(upload_files: List[str]) -> Tuple[List[Document], str]:
+    load_log = ''
+    documents = []
+    for upload_file in upload_files:
+        file_extension = f".{upload_file.split('.')[-1]}"
+        if file_extension in LOADER_CLASSES:
+            loader_class = LOADER_CLASSES[file_extension]
+            loader_kwargs = {}
+            if file_extension == '.csv':
+                with open(upload_file) as csvfile:
+                    delimiter = csv.Sniffer().sniff(csvfile.read(4096)).delimiter
+                loader_kwargs = {'csv_args': {'delimiter': delimiter}}
+            try:
+                load_documents = loader_class(upload_file, **loader_kwargs).load()
+                documents.extend(load_documents)
+            except Exception as ex:
+                load_log += f'Error uploading file {upload_file}\n'
+                load_log += f'Error code: {ex}\n'
+                continue
+        else:
+            load_log += f'Unsupported file format {upload_file}\n'
+            continue
+    return documents, load_log
+
+
+# extracting documents (in langchain Documents format) from WEB links
+def load_documents_from_links(
+        web_links: str,
+        subtitles_lang: str,
+        ) -> Tuple[List[Document], str]:
+
+    load_log = ''
+    documents = []
+    loader_class_kwargs = {}
+    web_links = [web_link.strip() for web_link in web_links.split('\n') if web_link.strip()]
+    for web_link in web_links:
+        if 'youtube.com' in web_link:
+            available, log = check_subtitles_available(web_link, subtitles_lang)
+            load_log += log
+            if not available:
+                continue
+            loader_class = LOADER_CLASSES['youtube'].from_youtube_url
+            loader_class_kwargs = {'language': subtitles_lang}
+        else:
+            loader_class = LOADER_CLASSES['web']
+            
+        try:
+            if requests.get(web_link).status_code != 200:
+                load_log += f'Ссылка недоступна для Python requests: {web_link}\n'
+                continue
+            load_documents = loader_class(web_link, **loader_class_kwargs).load()
+            if len(load_documents) == 0:
+                load_log += f'No text chunks were found at the link: {web_link}\n'
+                continue
+            documents.extend(load_documents)
+        except MissingSchema:
+            load_log += f'Invalid link: {web_link}\n'
+            continue
+        except Exception as ex:
+            load_log += f'Error loading data by web loader at link: {web_link}\n'
+            load_log += f'Error code: {ex}\n'
+            continue
+    return documents, load_log
+
+
+# uploading files and generating documents and databases
+def load_documents_and_create_db(
+        upload_files: Optional[List[str]],
+        web_links: str,
+        subtitles_lang: str,
+        chunk_size: int,
+        chunk_overlap: int,
+        embed_model_dict: EMBED_MODEL_DICT,
+        ) -> Tuple[List[Document], Optional[VectorStore], str]:
+
+    load_log = ''
+    all_documents = []
+    db = None
+    progress = gr.Progress()
+
+    embed_model = embed_model_dict.get('embed_model')
+    if embed_model is None:
+        load_log += 'Embeddings model not initialized, DB cannot be created'
+        return all_documents, db, load_log
+
+    if upload_files is None and not web_links:
+        load_log = 'No files or links selected'
+        return all_documents, db, load_log
+
+    if upload_files is not None:
+        progress(0.3, desc='Step 1/2: Upload documents from files')
+        docs, log = load_documents_from_files(upload_files)
+        all_documents.extend(docs)
+        load_log += log
+
+    if web_links:
+        progress(0.3 if upload_files is None else 0.5, desc='Step 1/2: Upload documents via links')
+        docs, log = load_documents_from_links(web_links, subtitles_lang)
+        all_documents.extend(docs)
+        load_log += log
+
+    if len(all_documents) == 0:
+        load_log += 'Download was interrupted because no documents were extracted\n'
+        load_log += 'RAG mode cannot be activated'
+        return all_documents, db, load_log
+
+    load_log += f'Documents loaded: {len(all_documents)}\n'
+    text_splitter = RecursiveCharacterTextSplitter(
+        chunk_size=chunk_size,
+        chunk_overlap=chunk_overlap,
+    )
+    documents = text_splitter.split_documents(all_documents)
+    documents = clear_documents(documents)
+    load_log += f'Documents are divided, number of text chunks: {len(documents)}\n'
+
+    progress(0.7, desc='Step 2/2: Initialize DB')
+    db = FAISS.from_documents(documents=documents, embedding=embed_model)
+    load_log += 'DB is initialized, RAG mode is activated and can be activated in the Chatbot tab'
+    return documents, db, load_log
+
+
+# ------------------ ФУНКЦИИ ЧАТ БОТА ------------------------
+
+# adding a user message to the chat bot window
+def user_message_to_chatbot(user_message: str, chatbot: CHAT_HISTORY) -> Tuple[str, CHAT_HISTORY]:
+    chatbot.append([user_message, None])
+    return '', chatbot
+
+
+# formatting prompt with adding context if DB is available and RAG mode is enabled
+def update_user_message_with_context(
+        chatbot: CHAT_HISTORY,
+        rag_mode: bool,
+        db: VectorStore,
+        k: Union[int, str],
+        score_threshold: float,
+        ) -> Tuple[str, CHAT_HISTORY]:
+
+    user_message = chatbot[-1][0]
+    user_message_with_context = ''
+    if db is not None and rag_mode and user_message.strip():
+        if k == 'all':
+            k = len(db.docstore._dict)
+        docs_and_distances = db.similarity_search_with_relevance_scores(
+            user_message,
+            k=k,
+            score_threshold=score_threshold,
+            )
+        if len(docs_and_distances) > 0:
+            retriever_context = '\n\n'.join([doc[0].page_content for doc in docs_and_distances])
+            user_message_with_context = CONTEXT_TEMPLATE.format(
+                user_message=user_message,
+                context=retriever_context,
+                )
+    return user_message_with_context
+
+
+# model response generation
+def get_llm_response(
+        chatbot: CHAT_HISTORY,
+        llm_model_dict: LLM_MODEL_DICT,
+        user_message_with_context: str,
+        rag_mode: bool,
+        system_prompt: str,
+        support_system_role: bool,
+        history_len: int,
+        do_sample: bool,
+        *generate_args,
+        ) -> CHAT_HISTORY:
+
+    user_message = chatbot[-1][0]
+    if not user_message.strip():
+        yield chatbot[:-1]
+        return None
+
+    if rag_mode:
+        if user_message_with_context:
+            user_message = user_message_with_context
+        else:
+            gr.Info((
+                f'No documents relevant to the query were found, generation in RAG mode is not possible.\n'
+                f'Try reducing searh_score_threshold or disable RAG mode for normal generation'
+                ))
+            yield chatbot[:-1]
+            return None
+
+    llm_model = llm_model_dict.get('model')
+    gen_kwargs = dict(zip(GENERATE_KWARGS.keys(), generate_args))
+    gen_kwargs['top_k'] = int(gen_kwargs['top_k'])
+    if not do_sample:
+        gen_kwargs['top_p'] = 0.0
+        gen_kwargs['top_k'] = 1
+        gen_kwargs['repeat_penalty'] = 1.0
+
+    messages = []
+    if support_system_role and system_prompt:
+        messages.append({'role': 'system', 'content': system_prompt})
+
+    if history_len != 0:
+        for user_msg, bot_msg in chatbot[:-1][-history_len:]:
+            messages.append({'role': 'user', 'content': user_msg})
+            messages.append({'role': 'assistant', 'content': bot_msg})
+
+    messages.append({'role': 'user', 'content': user_message})
+    stream_response = llm_model.create_chat_completion(
+        messages=messages,
+        stream=True,
+        **gen_kwargs,
+        )
+    try:
+        chatbot[-1][1] = ''
+        for chunk in stream_response:
+            token = chunk['choices'][0]['delta'].get('content')
+            if token is not None:
+                chatbot[-1][1] += token
+                yield chatbot
+    except Exception as ex:
+        gr.Info(f'Error generating response, error code: {ex}')
+        yield chatbot