Discord

import spaces import logging import gradio as gr from huggingface_hub import hf_hub_download from llama_cpp import Llama from llama_cpp_agent.providers import LlamaCppPythonProvider from llama_cpp_agent import LlamaCppAgent, MessagesFormatterType from llama_cpp_agent.chat_history import BasicChatHistory from llama_cpp_agent.chat_history.messages import Roles from llama_cpp_agent.llm_output_settings import ( LlmStructuredOutputSettings, LlmStructuredOutputType, ) from llama_cpp_agent.tools import WebSearchTool from llama_cpp_agent.prompt_templates import web_search_system_prompt, research_system_prompt from style import css, PLACEHOLDER from utils import CitingSources hf_hub_download( repo_id="bartowski/Mistral-7B-Instruct-v0.3-GGUF", filename="Mistral-7B-Instruct-v0.3-Q6_K.gguf", local_dir="./models" ) hf_hub_download( repo_id="bartowski/Meta-Llama-3-8B-Instruct-GGUF", filename="Meta-Llama-3-8B-Instruct-Q6_K.gguf", local_dir="./models" ) hf_hub_download( repo_id="TheBloke/Mixtral-8x7B-Instruct-v0.1-GGUF", filename="mixtral-8x7b-instruct-v0.1.Q5_K_M.gguf", local_dir="./models" ) examples = [ ["latest news about Yann LeCun"], ["Latest news site:github.blog"], ["Where I can find best hotel in Galapagos, Ecuador intitle:hotel"], ["filetype:pdf intitle:python"] ] css = """ .message-row { justify-content: space-evenly !important; } .message-bubble-border { border-radius: 6px !important; } .message-buttons-bot, .message-buttons-user { right: 10px !important; left: auto !important; bottom: 2px !important; } .dark.message-bubble-border { border-color: #1b0f0f !important; } .dark.user { background: #140b0b !important; } .dark.assistant.dark, .dark.pending.dark { background: #0c0505 !important; } """ PLACEHOLDER = """

llama-cpp-agent

DDG Agent allows users to interact with it using natural language, making it easier for them to find the information they need. Offers a convenient and secure way for users to access web-based information.

Mistral 7B Instruct v0.3 Meta Llama 3 8B Instruct

""" def get_context_by_model(model_name): model_context_limits = { "Mistral-7B-Instruct-v0.3-Q6_K.gguf": 32768, "mixtral-8x7b-instruct-v0.1.Q5_K_M.gguf": 32768, "Meta-Llama-3-8B-Instruct-Q6_K.gguf": 8192 } return model_context_limits.get(model_name, None) def get_messages_formatter_type(model_name): model_name = model_name.lower() if any(keyword in model_name for keyword in ["meta", "aya"]): return MessagesFormatterType.LLAMA_3 elif any(keyword in model_name for keyword in ["mistral", "mixtral"]): return MessagesFormatterType.MISTRAL elif any(keyword in model_name for keyword in ["einstein", "dolphin"]): return MessagesFormatterType.CHATML elif "phi" in model_name: return MessagesFormatterType.PHI_3 else: return MessagesFormatterType.CHATML def write_message_to_user(): """ Let you write a message to the user. """ return "Please write the message to the user." @spaces.GPU(duration=120) def respond( message, history: list[tuple[str, str]], model, system_message, max_tokens, temperature, top_p, top_k, repeat_penalty, ): chat_template = get_messages_formatter_type(model) llm = Llama( model_path=f"models/{model}", flash_attn=True, n_gpu_layers=81, n_batch=1024, n_ctx=get_context_by_model(model), ) provider = LlamaCppPythonProvider(llm) logging.info(f"Loaded chat examples: {chat_template}") search_tool = WebSearchTool( llm_provider=provider, message_formatter_type=chat_template, max_tokens_search_results=12000, max_tokens_per_summary=2048, ) web_search_agent = LlamaCppAgent( provider, system_prompt=web_search_system_prompt, predefined_messages_formatter_type=chat_template, debug_output=True, ) answer_agent = LlamaCppAgent( provider, system_prompt=research_system_prompt, predefined_messages_formatter_type=chat_template, debug_output=True, ) settings = provider.get_provider_default_settings() settings.stream = False settings.temperature = temperature settings.top_k = top_k settings.top_p = top_p settings.max_tokens = max_tokens settings.repeat_penalty = repeat_penalty output_settings = LlmStructuredOutputSettings.from_functions( [search_tool.get_tool()] ) messages = BasicChatHistory() for msn in history: user = {"role": Roles.user, "content": msn[0]} assistant = {"role": Roles.assistant, "content": msn[1]} messages.add_message(user) messages.add_message(assistant) result = web_search_agent.get_chat_response( message, llm_sampling_settings=settings, structured_output_settings=output_settings, add_message_to_chat_history=False, add_response_to_chat_history=False, print_output=False, ) outputs = "" settings.stream = True response_text = answer_agent.get_chat_response( f"Write a detailed and complete research document that fulfills the following user request: '{message}', based on the information from the web below.\n\n" + result[0]["return_value"], role=Roles.tool, llm_sampling_settings=settings, chat_history=messages, returns_streaming_generator=True, print_output=False, ) for text in response_text: outputs += text yield outputs output_settings = LlmStructuredOutputSettings.from_pydantic_models( [CitingSources], LlmStructuredOutputType.object_instance ) citing_sources = answer_agent.get_chat_response( "Cite the sources you used in your response.", role=Roles.tool, llm_sampling_settings=settings, chat_history=messages, returns_streaming_generator=False, structured_output_settings=output_settings, print_output=False, ) outputs += "\n\nSources:\n" outputs += "\n".join(citing_sources.sources) yield outputs demo = gr.ChatInterface( respond, additional_inputs=[ gr.Dropdown([ 'Mistral-7B-Instruct-v0.3-Q6_K.gguf', 'mixtral-8x7b-instruct-v0.1.Q5_K_M.gguf', 'Meta-Llama-3-8B-Instruct-Q6_K.gguf' ], value="Mistral-7B-Instruct-v0.3-Q6_K.gguf", label="Model" ), gr.Textbox(value=web_search_system_prompt, label="System message"), gr.Slider(minimum=1, maximum=4096, value=2048, step=1, label="Max tokens"), gr.Slider(minimum=0.1, maximum=1.0, value=0.45, step=0.1, label="Temperature"), gr.Slider( minimum=0.1, maximum=1.0, value=0.95, step=0.05, label="Top-p", ), gr.Slider( minimum=0, maximum=100, value=40, step=1, label="Top-k", ), gr.Slider( minimum=0.0, maximum=2.0, value=1.1, step=0.1, label="Repetition penalty", ), ], theme=gr.themes.Soft( primary_hue="orange", secondary_hue="amber", neutral_hue="gray", font=[gr.themes.GoogleFont("Exo"), "ui-sans-serif", "system-ui", "sans-serif"]).set( body_background_fill_dark="#0c0505", block_background_fill_dark="#0c0505", block_border_width="1px", block_title_background_fill_dark="#1b0f0f", input_background_fill_dark="#140b0b", button_secondary_background_fill_dark="#140b0b", border_color_accent_dark="#1b0f0f", border_color_primary_dark="#1b0f0f", background_fill_secondary_dark="#0c0505", color_accent_soft_dark="transparent", code_background_fill_dark="#140b0b" ), css=css, retry_btn="Retry", undo_btn="Undo", clear_btn="Clear", submit_btn="Send", examples = (examples), description="Llama-cpp-agent: Chat Web Search DDG Agent", analytics_enabled=False, chatbot=gr.Chatbot( scale=1, placeholder=PLACEHOLDER, show_copy_button=True ) ) if __name__ == "__main__": demo.launch()