├── infini_websearch ├── __init__.py ├── service │ ├── __init__.py │ └── search_service.py ├── actions │ ├── __init__.py │ ├── base_action.py │ ├── action_utils.py │ └── websearch.py ├── utils │ ├── __init__.py │ └── misc.py ├── model │ ├── __init__.py │ ├── postprocessing.py │ └── inference.py └── configs │ ├── server.py │ ├── css_style.py │ ├── __init__.py │ └── prompt.py ├── assets └── websearch_demo.gif ├── requirements.txt ├── setup.py ├── .gitignore ├── .pre-commit-config.yaml ├── README.md ├── README_en.md ├── LICENSE └── gradio_app.py /infini_websearch/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /infini_websearch/service/__init__.py: -------------------------------------------------------------------------------- 1 | -------------------------------------------------------------------------------- /assets/websearch_demo.gif: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/infinigence/InfiniWebSearch/HEAD/assets/websearch_demo.gif -------------------------------------------------------------------------------- /requirements.txt: -------------------------------------------------------------------------------- 1 | fastapi==0.115.6 2 | gradio==5.9.0 3 | gradio_toggle==2.0.2 4 | openai==1.57.4 5 | Requests==2.32.3 6 | selenium==4.27.1 7 | setuptools==68.2.2 8 | transformers==4.46.2 9 | uvicorn==0.32.1 10 | vllm==0.6.3.post1 11 | -------------------------------------------------------------------------------- /infini_websearch/actions/__init__.py: -------------------------------------------------------------------------------- 1 | from infini_websearch.actions.action_utils import parse_function_call_from_model_ouput 2 | from infini_websearch.actions.websearch import GoogleSearch 3 | 4 | __all__ = ["parse_function_call_from_model_ouput", "GoogleSearch"] 5 | -------------------------------------------------------------------------------- /infini_websearch/utils/__init__.py: -------------------------------------------------------------------------------- 1 | from infini_websearch.utils.misc import ( 2 | extract_citations, 3 | format_search_results, 4 | functions2str, 5 | get_datetime_now, 6 | ) 7 | 8 | __all__ = [ 9 | "extract_citations", 10 | "format_search_results", 11 | "functions2str", 12 | "get_datetime_now", 13 | ] 14 | -------------------------------------------------------------------------------- /setup.py: -------------------------------------------------------------------------------- 1 | from setuptools import find_packages, setup 2 | 3 | setup( 4 | name='infini_websearch', 5 | version='0.1.0', 6 | description=( 7 | "A demo built on Megrez-3B-Instruct," 8 | "integrating a web search tool to enhance the model's question-and-answer capabilities." 9 | ), 10 | packages=find_packages(), 11 | ) 12 | -------------------------------------------------------------------------------- /infini_websearch/actions/base_action.py: -------------------------------------------------------------------------------- 1 | from abc import ABC, abstractmethod 2 | from typing import Any, Dict, Optional 3 | 4 | 5 | class BaseAction(ABC): 6 | @property 7 | @abstractmethod 8 | def function_defination(self) -> Optional[Dict]: 9 | pass 10 | 11 | @abstractmethod 12 | def run(self, arguments: Dict) -> Any: 13 | pass 14 | -------------------------------------------------------------------------------- /infini_websearch/model/__init__.py: -------------------------------------------------------------------------------- 1 | from infini_websearch.model.inference import get_vllm_model_output_function 2 | from infini_websearch.model.postprocessing import ( 3 | include_special_tokens, 4 | split_text_by_special_token, 5 | ) 6 | 7 | __all__ = [ 8 | "get_vllm_model_output_function", 9 | "include_special_tokens", 10 | "split_text_by_special_token", 11 | ] 12 | -------------------------------------------------------------------------------- /.gitignore: -------------------------------------------------------------------------------- 1 | # Python 2 | __pycache__ 3 | *.pyc 4 | *.egg-info 5 | dist 6 | .venv 7 | 8 | # Log 9 | *.log 10 | *.log.* 11 | *.json 12 | !playground/deepspeed_config_s2.json 13 | !playground/deepspeed_config_s3.json 14 | 15 | # Editor 16 | .idea 17 | *.swp 18 | 19 | # Other 20 | .DS_Store 21 | wandb 22 | output 23 | checkpoints_flant5_3b 24 | 25 | # Data 26 | *.pkl 27 | *.csv 28 | tests/state_of_the_union.txt 29 | 30 | # Build 31 | build 32 | -------------------------------------------------------------------------------- /infini_websearch/model/postprocessing.py: -------------------------------------------------------------------------------- 1 | from typing import List, Tuple 2 | 3 | 4 | def include_special_tokens(text: str, tokens: List[str]) -> bool: 5 | include_all_tokens = True 6 | for token in tokens: 7 | if token not in text: 8 | include_all_tokens = False 9 | break 10 | return include_all_tokens 11 | 12 | 13 | def split_text_by_special_token(text: str, token: str) -> Tuple[str, str]: 14 | assert token in text, f"{token} not found in {text}" 15 | parts = text.split(token) 16 | return parts[0], parts[1] 17 | -------------------------------------------------------------------------------- /.pre-commit-config.yaml: -------------------------------------------------------------------------------- 1 | repos: 2 | - repo: https://github.com/pre-commit/pre-commit-hooks 3 | rev: v4.6.0 4 | hooks: 5 | - id: trailing-whitespace 6 | - id: end-of-file-fixer 7 | - id: check-added-large-files 8 | - id: check-yaml 9 | - id: check-json 10 | 11 | - repo: https://github.com/pycqa/isort 12 | rev: 5.13.2 13 | hooks: 14 | - id: isort 15 | args: 16 | - "--profile=black" 17 | 18 | - repo: https://github.com/psf/black 19 | rev: 23.7.0 20 | hooks: 21 | - id: black 22 | language_version: python3.10 23 | args: 24 | - "--skip-string-normalization" 25 | 26 | 27 | - repo: https://github.com/PyCQA/flake8 28 | rev: 7.1.1 29 | hooks: 30 | - id: flake8 31 | args: 32 | - --max-line-length=120 33 | additional_dependencies: 34 | - flake8-bugbear 35 | 36 | default_language_version: 37 | python: python3.10 38 | -------------------------------------------------------------------------------- /infini_websearch/configs/server.py: -------------------------------------------------------------------------------- 1 | # gradio 2 | SESSION_WINDOW_SIZE = 4 3 | 4 | # websearch service 5 | SEARCH_SERVER_URL = "http://localhost:8021/search" 6 | NUM_SEARCH_WEBPAGES = 5 7 | WEBPAGE_LOAD_TIMETOUT = 10.0 8 | PROXIES = { 9 | "http": None, 10 | "https": None, 11 | } 12 | 13 | 14 | # model 15 | MODEL_NAME = "megrez" 16 | MODEL_SERVER_URL = "http://localhost:8011/v1/" 17 | STOP_TOKENS = ["<|turn_end|>"] 18 | FUNCTION_START_TOKEN, FUNCTION_END_TOKEN = "<|function_start|>", "<|function_end|>" 19 | MAX_ACTION_TURNS = 1 20 | 21 | # 4k 22 | # WEBPAGE_SUMMARY_MAX_INPUT_TOKENS = 2048 23 | # WEBPAGE_SUMMARY_MAX_OUTPUT_TOKENS = 512 24 | # SESSION_MAX_INPUT_TOKENS = 3072 25 | # CHAT_TEMPERATURE = 0.4 26 | # CHAT_MAX_OUTPUT_TOKENS = 2048 27 | # AGENT_TEMPERATURE = 0.01 28 | # AGENT_MAX_OUTPUT_TOKENS = 512 29 | 30 | # 32k 31 | WEBPAGE_SUMMARY_MAX_INPUT_TOKENS = 2048 32 | WEBPAGE_SUMMARY_MAX_OUTPUT_TOKENS = 512 33 | SESSION_MAX_INPUT_TOKENS = 32768 - 4096 34 | CHAT_TEMPERATURE = 0.4 35 | CHAT_MAX_OUTPUT_TOKENS = 2048 36 | AGENT_TEMPERATURE = 0.01 37 | AGENT_MAX_OUTPUT_TOKENS = 2048 38 | -------------------------------------------------------------------------------- /infini_websearch/utils/misc.py: -------------------------------------------------------------------------------- 1 | import json 2 | import re 3 | from datetime import datetime 4 | 5 | 6 | def functions2str(functions: list) -> str: 7 | return "\n\n".join( 8 | [json.dumps(function, ensure_ascii=False, indent=4) for function in functions] 9 | ) 10 | 11 | 12 | def get_datetime_now(): 13 | current_date = datetime.now() 14 | formatted_time = current_date.strftime("%Y-%m-%d %H:%M:%S") 15 | weekday_id = current_date.weekday() 16 | weekday_names = [ 17 | "Monday", 18 | "Tuesday", 19 | "Wednesday", 20 | "Thursday", 21 | "Friday", 22 | "Saturday", 23 | "Sunday", 24 | ] 25 | return formatted_time, weekday_names[weekday_id] 26 | 27 | 28 | def format_search_results(url_infos: dict): 29 | return "\n".join( 30 | f"[{url_info['title']}]({url_info['link']})" for url_info in url_infos 31 | ) 32 | 33 | 34 | def extract_citations(text): 35 | citations1 = re.findall(r"\[citation:(\d+)\]", text) 36 | citations2 = re.findall(r"\[ citation:(\d+)\]", text) 37 | return citations1 or citations2 38 | -------------------------------------------------------------------------------- /infini_websearch/configs/css_style.py: -------------------------------------------------------------------------------- 1 | CSS_STYLE = """ 2 | .canvas { 3 | # width: 100% !important; 4 | # max-width: 100% !important; 5 | width: 100vh; 6 | } 7 | 8 | .fullheight { 9 | height: 80vh; 10 | } 11 | 12 | .chatbot { 13 | flex-grow: 1; 14 | overflow: auto; 15 | position: relative; 16 | z-index: 100; 17 | } 18 | 19 | .bottom-bar { 20 | position: fixed; 21 | bottom: 0; 22 | left: 50%; 23 | transform: translateX(-50%); 24 | display: flex; 25 | width: 80vh; 26 | z-index: 1000; 27 | } 28 | 29 | .unicode-circle { 30 | font-family: 'Arial Unicode MS', Arial, sans-serif; 31 | font-size: 14px; 32 | border-radius: 50%; 33 | border: 1px solid black; 34 | width: 20px; 35 | height: 20px; 36 | line-height: 20px; 37 | text-align: center; 38 | display: inline-block; 39 | background-color: white; 40 | } 41 | 42 | .circle-link { 43 | font-family: 'Arial Unicode MS', Arial, sans-serif; 44 | text-decoration: none; 45 | color: black; 46 | border-radius: 50%; 47 | border: 1px solid black; 48 | width: 20px; 49 | height: 20px; 50 | line-height: 20px; 51 | text-align: center; 52 | display: inline-block; 53 | background-color: white; 54 | cursor: pointer; 55 | } 56 | 57 | .circle-link:hover { 58 | background-color: #e0e0e0; 59 | } 60 | 61 | .circle-link:hover::after { 62 | content: attr(title); 63 | position: absolute; 64 | white-space: nowrap; 65 | } 66 | """ 67 | -------------------------------------------------------------------------------- /infini_websearch/actions/action_utils.py: -------------------------------------------------------------------------------- 1 | import json 2 | import re 3 | from typing import Dict, List, Optional, Tuple, Union 4 | 5 | 6 | def parse_function_call_from_model_ouput( 7 | output: str, 8 | registered_function_names: Optional[List[str]], 9 | speical_tokens_map: Optional[Dict], 10 | ) -> Tuple[Optional[str], Union[Dict, Optional[str]]]: 11 | if speical_tokens_map is None: 12 | speical_tokens_map = dict( 13 | function_start_token="<|function_start|>", 14 | function_end_token="<|function_end|>", 15 | ) 16 | 17 | function_name, function_arguments = None, None 18 | function_call_texts = re.findall( 19 | f'{re.escape(speical_tokens_map["function_start_token"])}(.*?){re.escape(speical_tokens_map["function_end_token"])}', # noqa: E501 20 | output, 21 | re.DOTALL, 22 | ) 23 | if len(function_call_texts) > 0: 24 | # support only one action per turn, choose the first one 25 | function_call_text = function_call_texts[0].strip() 26 | try: 27 | function_call_dict = json.loads(function_call_text) 28 | function_name = function_call_dict["name"] 29 | function_arguments = function_call_dict["arguments"] 30 | except Exception as e: 31 | print(e) 32 | function_name = None 33 | print("function call json输入格式错误") 34 | print(function_call_text) 35 | 36 | if function_name is not None and function_name not in registered_function_names: 37 | function_arguments = f"{function_name}不在可以使用的工具列表中" 38 | function_name = None 39 | return function_name, function_arguments 40 | -------------------------------------------------------------------------------- /infini_websearch/configs/__init__.py: -------------------------------------------------------------------------------- 1 | from infini_websearch.configs.css_style import CSS_STYLE 2 | from infini_websearch.configs.prompt import ( 3 | FUNCTION_CALLING_PROMPT_TEMPLATE, 4 | OBSERVATION_PROMPT_TEMPLATE, 5 | ROLE_PROMPT, 6 | SUMMARY_PROMPT_TEMPLATE, 7 | TIME_PROMPT_TEMPLATE, 8 | ) 9 | from infini_websearch.configs.server import ( 10 | AGENT_MAX_OUTPUT_TOKENS, 11 | AGENT_TEMPERATURE, 12 | CHAT_MAX_OUTPUT_TOKENS, 13 | CHAT_TEMPERATURE, 14 | FUNCTION_END_TOKEN, 15 | FUNCTION_START_TOKEN, 16 | MAX_ACTION_TURNS, 17 | MODEL_NAME, 18 | MODEL_SERVER_URL, 19 | NUM_SEARCH_WEBPAGES, 20 | PROXIES, 21 | SEARCH_SERVER_URL, 22 | SESSION_MAX_INPUT_TOKENS, 23 | SESSION_WINDOW_SIZE, 24 | STOP_TOKENS, 25 | WEBPAGE_LOAD_TIMETOUT, 26 | WEBPAGE_SUMMARY_MAX_INPUT_TOKENS, 27 | WEBPAGE_SUMMARY_MAX_OUTPUT_TOKENS, 28 | ) 29 | 30 | __all__ = [ 31 | "CSS_STYLE", 32 | "FUNCTION_CALLING_PROMPT_TEMPLATE", 33 | "OBSERVATION_PROMPT_TEMPLATE", 34 | "ROLE_PROMPT", 35 | "SUMMARY_PROMPT_TEMPLATE", 36 | "TIME_PROMPT_TEMPLATE", 37 | "AGENT_MAX_OUTPUT_TOKENS", 38 | "AGENT_TEMPERATURE", 39 | "CHAT_MAX_OUTPUT_TOKENS", 40 | "CHAT_TEMPERATURE", 41 | "FUNCTION_END_TOKEN", 42 | "FUNCTION_START_TOKEN", 43 | "MAX_ACTION_TURNS", 44 | "MODEL_NAME", 45 | "MODEL_SERVER_URL", 46 | "NUM_SEARCH_WEBPAGES", 47 | "PROXIES", 48 | "SEARCH_SERVER_URL", 49 | "STOP_TOKENS", 50 | "WEBPAGE_LOAD_TIMETOUT", 51 | "WEBPAGE_SUMMARY_MAX_INPUT_TOKENS", 52 | "SESSION_MAX_INPUT_TOKENS", 53 | "SESSION_WINDOW_SIZE", 54 | "WEBPAGE_SUMMARY_MAX_OUTPUT_TOKENS", 55 | ] 56 | -------------------------------------------------------------------------------- /infini_websearch/configs/prompt.py: -------------------------------------------------------------------------------- 1 | ROLE_PROMPT = "你是Megrez-3B-Instruct, 将针对用户的问题给出详细的、积极的回答." 2 | 3 | TIME_PROMPT_TEMPLATE = "The current time is {current_time}, {weekday}." 4 | 5 | FUNCTION_CALLING_PROMPT_TEMPLATE = ( 6 | "You have access to the following functions. Use them if required -\n{functions}" 7 | ) 8 | 9 | SUMMARY_PROMPT_TEMPLATE = ( 10 | '从信息中总结能够回答问题的相关内容，要求简明扼要不能完全照搬原文。直接返回总结不要说其他话，如果没有相关内容则返回"无相关内容", 返回内容为中文。\n\n' 11 | "<问题>{question}\n" 12 | "<信息>{context}" 13 | ) 14 | 15 | # this prompt was inspired by 16 | # https://github.com/leptonai/search_with_lepton/blob/main/search_with_lepton.py 17 | OBSERVATION_PROMPT_TEMPLATE = ( 18 | "You will be given a set of related contexts to the question, " 19 | "each starting with a reference number like [[citation:x]], where x is a number. " 20 | "Please use the context and cite the context at the end of each sentence if applicable." 21 | "\n\n" 22 | "Please cite the contexts with the reference numbers, in the format [citation:x]. " 23 | "If a sentence comes from multiple contexts, please list all applicable citations, like [citation:3][citation:5]. " 24 | "If the context does not provide relevant information to answer the question, " 25 | "inform the user that there is no relevant information in the search results and that the question cannot be answered." # noqa: E501 26 | "\n\n" 27 | "Other than code and specific names and citations, your answer must be written in Chinese." 28 | "\n\n" 29 | "Ensure that your response is concise and clearly formatted. " 30 | "Group related content together and use Markdown points or lists where appropriate." 31 | "\n\n" 32 | "Remember, summarize and don't blindly repeat the contexts verbatim. And here is the user question:\n" 33 | "{question}\n" 34 | "Here is the keywords of the question:\n" 35 | "{keywords}" 36 | "\n\n" 37 | "Here are the set of contexts:" 38 | "\n\n" 39 | "{context}" 40 | ) 41 | -------------------------------------------------------------------------------- /infini_websearch/model/inference.py: -------------------------------------------------------------------------------- 1 | from functools import partial 2 | from typing import Callable, Dict, Generator, List, Union 3 | 4 | import openai 5 | 6 | 7 | def get_vllm_model_output_function( 8 | url: str, 9 | model_name: str, 10 | chat_mode: bool, 11 | model_config: Dict, 12 | stream: bool, 13 | buffer_size: int = 20, 14 | timeout: int = 60, 15 | ) -> Callable: 16 | """ 17 | Get model generate function: streaming/non-streaming 18 | """ 19 | data = { 20 | "model": model_name, 21 | "stream": stream, 22 | **model_config, 23 | } 24 | 25 | openai.api_key = "EMPTY" 26 | openai.base_url = url 27 | openai.proxy = "" 28 | chat_func = ( 29 | openai.chat.completions.create 30 | if chat_mode is True 31 | else openai.completions.create 32 | ) 33 | 34 | if stream is True: 35 | return partial( 36 | get_model_streaming_output, 37 | llm_function=chat_func, 38 | model_config=data, 39 | chat_mode=chat_mode, 40 | buffer_size=buffer_size, 41 | timeout=timeout, 42 | ) 43 | else: 44 | return partial( 45 | get_model_output, 46 | llm_function=chat_func, 47 | model_config=data, 48 | chat_mode=chat_mode, 49 | timeout=timeout, 50 | ) 51 | 52 | 53 | def get_model_streaming_output( 54 | messages: List[Union[Dict, str]], 55 | model_config: Dict, 56 | llm_function: Callable, 57 | chat_mode: bool, 58 | buffer_size: int, 59 | timeout: int, 60 | ) -> Generator[str, None, None]: 61 | if chat_mode is True: 62 | model_config["messages"] = messages 63 | else: 64 | model_config["prompt"] = messages 65 | buffer = "" 66 | for chunk in llm_function(**model_config, timeout=timeout): 67 | if chunk.choices[0].delta.content: 68 | buffer += chunk.choices[0].delta.content 69 | if len(buffer) >= buffer_size: 70 | # '[citation:x]' has been truncated? 71 | if buffer.rfind("]") < buffer.rfind("["): 72 | yield buffer[: buffer.rfind("[")] 73 | buffer = buffer[buffer.rfind("[") :] # noqa: E203 74 | else: 75 | yield buffer 76 | buffer = "" 77 | if buffer: 78 | yield buffer 79 | 80 | 81 | def get_model_output( 82 | messages: List[Union[Dict, str]], 83 | model_config: Dict, 84 | llm_function: Callable, 85 | chat_mode: bool, 86 | timeout: int, 87 | ) -> str: 88 | if chat_mode is True: 89 | model_config["messages"] = messages 90 | else: 91 | model_config["prompt"] = messages 92 | response = llm_function(**model_config, timeout=timeout) 93 | return response 94 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # InfiniWebSearch 2 | 3 | 基于[Megrez-3B-Instruct](https://huggingface.co/Infinigence/Megrez-3B-Instruct)搭建的demo, 接入网络搜索工具增强模型的问答能力. 4 | 5 | Read this in [English](README_en.md) 6 | 7 |
8 | Example GIF 9 |
10 | 11 | ## 特性 12 | 13 | 1. **意图识别**: LLM自动决定搜索工具调用时机 14 | 2. **上下文理解**: 根据多轮对话生成合理搜索关键词 15 | 3. **模型回答包含引用**: 回答内容可查明出处 16 | 4. **即插即用**: 通过system prompt控制WebSearch功能开启与否 17 | 18 | ## 项目简介 19 | 20 | 本工程包含如下几部分: 21 | 22 | - gradio服务 ([gradio_app.py](gradio_app.py)): 定义了整个app的workflow 23 | - 网络搜索服务 ([search_service.py](infini_websearch/service/search_service.py)): 搜索网页, 加载网页 24 | - 模型服务: 聊天问答, 工具调用, 总结网页内容 25 | 26 | ## 快速上手 27 | 28 | ### 安装 29 | 30 | ```shell 31 | git clone https://github.com/infinigence/InfiniWebSearch 32 | cd InfiniWebSearch 33 | conda create -n infini_websearch python=3.10 -y 34 | conda activate infini_websearch 35 | pip install -r requirements.txt 36 | pip install -e . 37 | ``` 38 | 39 | ### 运行demo 40 | 41 | #### 1. 启动网络搜索服务 42 | 43 | 以Ubuntu环境为例, 从[Google Chrome Labs](https://googlechromelabs.github.io/chrome-for-testing/)下载**chrome**和**chromedriver**并解压到本地. 44 | 45 | ``` 46 | cd infini_websearch/service 47 | wget https://storage.googleapis.com/chrome-for-testing-public/128.0.6613.86/linux64/chrome-linux64.zip 48 | wget https://storage.googleapis.com/chrome-for-testing-public/128.0.6613.86/linux64/chromedriver-linux64.zip 49 | unzip chrome-linux64.zip 50 | unzip chromedriver-linux64.zip 51 | ``` 52 | 53 | 安装依赖 54 | 55 | ```shell 56 | sudo apt-get update 57 | sudo apt-get install -y \ 58 | libatk-bridge2.0-0 \ 59 | libatk1.0-0 \ 60 | libgconf-2-4 \ 61 | libnss3 \ 62 | libxss1 \ 63 | libappindicator1 \ 64 | libindicator7 \ 65 | libasound2 \ 66 | libxcomposite1 \ 67 | libxcursor1 \ 68 | libxdamage1 \ 69 | libxi6 \ 70 | libxtst6 \ 71 | libglib2.0-0 \ 72 | libpango1.0-0 \ 73 | libcups2 \ 74 | libxrandr2 \ 75 | libxrandr-dev \ 76 | libxkbcommon0 \ 77 | libgbm1 78 | ``` 79 | 80 | 在[Serper](https://serper.dev/)注册账户获得`SERPER_API_KEY`, 并添加到环境变量中. 81 | 启动网络搜索服务并指定端口号. 82 | 在[server.py](infini_websearch/configs/server.py)设置`SEARCH_SERVER_URL`为 http://localhost:8021/search . 83 | 84 | ```shell 85 | export SERPER_API_KEY=$YOUR_API_KEY 86 | cd infini_websearch/service 87 | python search_service.py --port 8021 --chrome ./chrome-linux64/chrome --chromedriver ./chromedriver-linux64/chromedriver 88 | ``` 89 | 90 | #### 2. 启动模型服务 91 | 92 | 使用vllm.entrypoints.openai.api_server启动服务并制定端口号, `--served-model-name`设置为megrez, `--max-seq-len`设置为32768. 93 | 在[server.py](infini_websearch/configs/server.py)设置`MODEL_SERVER_URL`, 默认为 http://localhost:8011/v1/ . 设置`MODEL_NAME`为"megrez". 94 | 95 | ```shell 96 | python -m vllm.entrypoints.openai.api_server --served-model-name megrez --model $MODEL_PATH --port 8011 --max-seq-len 32768 --trust_remote_code --gpu-memory-utilization 0.8 97 | ``` 98 | 99 | #### 3. 启动gradio服务 100 | 101 | 运行[gradio_app.py](infini_websearch/gradio_app.py), 指定模型路径和端口号. 102 | ```shell 103 | export no_proxy="localhost,127.0.0.1" 104 | python gradio_app.py -m $MODEL_PATH --port 7860 105 | ``` 106 | 107 | 成功启动之后, 访问 http://localhost:7860/ 即可使用 108 | 109 | ## 说明 110 | 111 | 1. 我们提供了`WEBPAGE_SUMMARY_MAX_INPUT_TOKENS`, `WEBPAGE_SUMMARY_MAX_OUTPUT_TOKENS`, `SESSION_MAX_INPUT_TOKENS`, `CHAT_MAX_OUTPUT_TOKENS`, `AGENT_MAX_OUTPUT_TOKENS`来控制模型的输入和输出长度. 使用`SESSION_WINDOW_SIZE`来保留最近的几轮对话历史.你可以在[server.py](infini_websearch/configs/server.py)中按需修改. 112 | 2. 已经开始首轮对话后，点击websearch toggle切换状态会在后端清空对话历史, 但前端显示依然保留对话历史. 113 | 3. 如果搜索服务出现异常(例如: 网页加载超时或服务器异常), 工具调用的observation会返回预定义好的信息(例如: "搜索页面加载超时, 请重试").你可以在[websearch.py](infini_websearch/actions/websearch.py)和[search_service.py](infini_websearch/service/search_service.py)中自定义边界条件的后处理逻辑. 114 | 4. 使用[Serper](https://serper.dev/)时([search_service.py](infini_websearch/service/search_service.py)), 我们设置"hl"参数为"zh-CN"来尽可能得到中文搜索结果. 如果搜索结果英文网页太多, 可能导致模型用英文回答. 115 | 5. 如果网页搜索成功了, 但是模型表示搜索结果中没有包含回答问题的相关信息, 可以检查控制台打印的各网页的摘要信息. 如果摘要信息显示"无相关内容", 代表原网页没有与问题相关的信息, 或者模型提取与问题相关信息失败. 116 | 117 | ## 协议 118 | 119 | - 本开源仓库的代码则遵循 [Apache 2.0](LICENSE) 协议. 120 | -------------------------------------------------------------------------------- /README_en.md: -------------------------------------------------------------------------------- 1 | # InfiniWebSearch 2 | 3 | A demo built on [Megrez-3B-Instruct](https://huggingface.co/Infinigence/Megrez-3B-Instruct), integrating a web search tool to enhance the model's question-and-answer capabilities. 4 | 5 | [中文阅读](README.md). 6 | 7 |
8 | Example GIF 9 |
10 | 11 | ## Features 12 | 13 | 1. **Intent Recognition**: LLM automatically determines when to invoke search tool calls. 14 | 2. **Context Understanding**: Generates reasonable search keywords based on multi-turn dialogue. 15 | 3. **Model Responses Include Citation Links**: The content of the responses can be traced back to their sources. 16 | 4. **Plug-and-Play**: Control the activation or deactivation of WebSearch functionality through system prompts. 17 | 18 | ## Project Introduction 19 | 20 | This project consists of the following parts: 21 | 22 | - Gradio service ([gradio_app.py](gradio_app.py)): Defines the entire app's workflow. 23 | - Web search service ([search_service.py](infini_websearch/service/search_service.py)): Searches for web pages and loads web pages. 24 | - Model service: Chat, function calling and summarize web page. 25 | 26 | ## Quick start 27 | 28 | ### Install 29 | 30 | ```shell 31 | git clone https://github.com/infinigence/InfiniWebSearch 32 | cd InfiniWebSearch 33 | conda create -n infini_websearch python=3.10 -y 34 | conda activate infini_websearch 35 | pip install -r requirements.txt 36 | pip install -e . 37 | ``` 38 | 39 | ### Running Demo 40 | 41 | #### 1. Starting Web Search Service 42 | 43 | As an example for the Ubuntu environment, download **chrome** and **chromedriver** from [Google Chrome Labs](https://googlechromelabs.github.io/chrome-for-testing/) and unzip them to the local directory. 44 | 45 | ``` 46 | cd infini_websearch/service 47 | wget https://storage.googleapis.com/chrome-for-testing-public/128.0.6613.86/linux64/chrome-linux64.zip 48 | wget https://storage.googleapis.com/chrome-for-testing-public/128.0.6613.86/linux64/chromedriver-linux64.zip 49 | unzip chrome-linux64.zip 50 | unzip chromedriver-linux64.zip 51 | ``` 52 | 53 | Install dependencies 54 | 55 | ```shell 56 | sudo apt-get update 57 | sudo apt-get install -y \ 58 | libatk-bridge2.0-0 \ 59 | libatk1.0-0 \ 60 | libgconf-2-4 \ 61 | libnss3 \ 62 | libxss1 \ 63 | libappindicator1 \ 64 | libindicator7 \ 65 | libasound2 \ 66 | libxcomposite1 \ 67 | libxcursor1 \ 68 | libxdamage1 \ 69 | libxi6 \ 70 | libxtst6 \ 71 | libglib2.0-0 \ 72 | libpango1.0-0 \ 73 | libcups2 \ 74 | libxrandr2 \ 75 | libxrandr-dev \ 76 | libxkbcommon0 \ 77 | libgbm1 78 | ``` 79 | 80 | Register an account on [Serper](https://serper.dev/) to obtain `SERPER_API_KEY` and add it to the environment variable. 81 | Start the web search service and specify the port number. 82 | Set the `SEARCH_SERVER_URL` to http://localhost:8021/search in the file [server.py](infini_websearch/configs/server.py). 83 | 84 | ```shell 85 | export SERPER_API_KEY=$YOUR_API_KEY 86 | cd infini_websearch/service 87 | python search_service.py --port 8021 --chrome ./chrome-linux64/chrome --chromedriver ./chromedriver-linux64/chromedriver 88 | ``` 89 | 90 | #### 2. Starting Model Service 91 | 92 | Use vllm.entrypoints.openai.api_server to start the service and specify the port number. Set the `--served-model-name` to "megrez" and `--max-seq-len` to 32768. 93 | Set the `MODEL_SERVER_URL` in the file [server.py](infini_websearch/configs/server.py) with a default value of http://localhost:8011/v1/. Also, set the `MODEL_NAME` to "megrez". 94 | 95 | ```shell 96 | python -m vllm.entrypoints.openai.api_server --served-model-name megrez --model $MODEL_PATH --port 8011 --max-seq-len 32768 --trust_remote_code --gpu-memory-utilization 0.8 97 | ``` 98 | 99 | #### 3. Starting Gradio Service 100 | 101 | Run [gradio_app.py](infini_websearch/gradio_app.py), specify the model path and port number. 102 | 103 | ```shell 104 | export no_proxy="localhost,127.0.0.1" 105 | python gradio_app.py -m $MODEL_PATH --port 7860 106 | ``` 107 | 108 | After successful startup, you can use it by visiting http://localhost:7860/. 109 | 110 | ## Notes 111 | 112 | 1. We provide `WEBPAGE_SUMMARY_MAX_INPUT_TOKENS`, `WEBPAGE_SUMMARY_MAX_OUTPUT_TOKENS`, `SESSION_MAX_INPUT_TOKENS`, `CHAT_MAX_OUTPUT_TOKENS`, `AGENT_MAX_OUTPUT_TOKENS` to control the input and output lengths of the model. Use `SESSION_WINDOW_SIZE` to retain the most recent dialogue history. You can modify these settings as needed in [server.py](infini_websearch/configs/server.py). 113 | 2. After starting the first round of dialogue, toggling the websearch state will clear the dialogue history on the backend, but the frontend will still display the dialogue history. 114 | 3. If there is an exception with the search service (e.g. webpage loading timeout or server error), the observation from the tool call will return predefined messages (e.g. "The search page loading timed out, please try again"). You can customize the post-processing logic for boundary conditions in [websearch.py](infini_websearch/actions/websearch.py) and [search_service.py](infini_websearch/service/search_service.py). 115 | 4. When using [Serper](https://serper.dev/) ([search_service.py](infini_websearch/service/search_service.py)), we set the "hl" parameter to "zh-CN" to obtain Chinese search results as much as possible. If there are too many English webpages in the search results, it may lead to the model responding in English. 116 | 5. If the web search is successful but the model indicates that the search results do not contain relevant information to answer the question, you can check the summary information of each webpage printed in the console. If the summary shows "No relevant content", it means either the original webpage does not contain information related to the question, or the model failed to extract relevant information from the webpage. 117 | 118 | ## License 119 | 120 | The code in this open-source repository follows the [Apache 2.0](LICENSE) license. 121 | -------------------------------------------------------------------------------- /infini_websearch/service/search_service.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | import json 3 | import os 4 | import time 5 | from concurrent.futures import ThreadPoolExecutor, as_completed 6 | from typing import Dict, Generator, Optional, Tuple, Union 7 | 8 | import requests 9 | from fastapi import FastAPI, HTTPException, Request 10 | from fastapi.responses import StreamingResponse 11 | from selenium import webdriver 12 | from selenium.common.exceptions import TimeoutException 13 | from selenium.webdriver.chrome.options import Options 14 | from selenium.webdriver.chrome.service import Service 15 | 16 | parser = argparse.ArgumentParser() 17 | parser.add_argument("--chrome", type=str) 18 | parser.add_argument("--chromedriver", type=str) 19 | parser.add_argument("--port", type=int) 20 | 21 | args = parser.parse_args() 22 | 23 | app = FastAPI() 24 | 25 | SERPER_API_KEY = os.environ.get("SERPER_API_KEY") 26 | 27 | 28 | def get_webpage_content(url: str, chrome_path: str, chromedriver_path: str) -> str: 29 | """ 30 | Load the content of web pages by chromedriver. 31 | """ 32 | options = Options() 33 | options.add_argument("--no-sandbox") 34 | options.add_argument("--disable-dev-shm-usage") 35 | options.add_argument("--disable-extensions") 36 | options.add_argument("--disable-gpu") 37 | options.add_argument("--headless") 38 | options.add_argument("--disable-infobars") 39 | options.add_argument("--disable-browser-side-navigation") 40 | options.add_argument("--disable-features=VizDisplayCompositor") 41 | options.add_argument("--no-first-run") 42 | options.add_argument("--no-default-browser-check") 43 | options.add_argument("--disable-popup-blocking") 44 | options.add_argument("--disable-application-cache") 45 | options.add_argument("--dns-prefetch-disable") 46 | options.add_argument("--no-proxy-server") 47 | options.add_argument("--blink-settings=imagesEnabled=false") 48 | options.add_argument("--enable-http2") 49 | options.add_argument("--disable-quic") 50 | options.binary_location = chrome_path 51 | options.page_load_strategy = "eager" 52 | prefs = { 53 | "profile.managed_default_content_settings.images": 2, 54 | "profile.default_content_setting_values.notifications": 2, 55 | "download_restrictions": 3, 56 | } 57 | options.add_experimental_option("prefs", prefs) 58 | service = Service(executable_path=chromedriver_path) 59 | driver = webdriver.Chrome(options=options, service=service) 60 | try: 61 | timeout = 10 62 | start = time.time() 63 | driver.set_page_load_timeout(timeout) 64 | try: 65 | driver.get(url) 66 | except TimeoutException: 67 | print(f"页面加载超时（{timeout}秒）") 68 | return "搜索页面加载超时, 请重试" 69 | end = time.time() 70 | print(f"读取网页内容耗时: {end - start}s") 71 | content = driver.execute_script("return document.body.innerText;") 72 | return content 73 | except Exception as e: 74 | print(e) 75 | return "" 76 | finally: 77 | driver.quit() 78 | 79 | 80 | def serper_search( 81 | search_term: str, search_type: Optional[str] = "search", timeout: int = 5, **kwargs 82 | ) -> Tuple[int, Union[Dict, str]]: 83 | """ 84 | Get google search results by serper api (https://serper.dev/). 85 | """ 86 | headers = { 87 | "X-API-KEY": SERPER_API_KEY, 88 | "Content-Type": "application/json", 89 | } 90 | params = { 91 | "q": search_term, 92 | "gl": "cn", # country 93 | "sort": "date", 94 | "hl": "zh-CN", 95 | **{key: value for key, value in kwargs.items() if value is not None}, 96 | } 97 | try: 98 | response = requests.post( 99 | f"https://google.serper.dev/{search_type}", 100 | headers=headers, 101 | params=params, 102 | proxies=None, 103 | timeout=timeout, 104 | ) 105 | except Exception as e: 106 | return -1, str(e) 107 | return response.status_code, response.json() 108 | 109 | 110 | def streaming_fetch_webpage_content( 111 | results: dict, num_search_pages: int, chrome_path: str, chromedriver_path: str 112 | ) -> Generator[Tuple[str, str], None, None]: 113 | url_infos = results["organic"][:num_search_pages] 114 | 115 | with ThreadPoolExecutor(max_workers=len(url_infos)) as executor: 116 | future_to_url = { 117 | executor.submit( 118 | get_webpage_content, url_info["link"], chrome_path, chromedriver_path 119 | ): url_info 120 | for url_info in url_infos 121 | } 122 | for future in as_completed(future_to_url): 123 | url_info = future_to_url[future] 124 | try: 125 | result = future.result() 126 | yield url_info, result 127 | except Exception as exc: 128 | print(f'{url_info["link"]} generated an exception: {exc}') 129 | yield url_info, "" 130 | 131 | 132 | @app.post("/search") 133 | async def search(request: Request): 134 | data = await request.json() 135 | print(data) 136 | 137 | start = time.time() 138 | status_code, response = serper_search(data["query"], timeout=10) 139 | end = time.time() 140 | print(f"搜索网页耗时: {end - start}s") 141 | 142 | if status_code != 200: 143 | raise HTTPException(status_code=500, detail="搜索网页超时, 请重试") 144 | 145 | def html_docs_text_generator(): 146 | start = time.time() 147 | for url_info, content in streaming_fetch_webpage_content( 148 | response, 149 | num_search_pages=data["num_search_pages"], 150 | chrome_path=args.chrome, 151 | chromedriver_path=args.chromedriver, 152 | ): 153 | yield json.dumps( 154 | { 155 | "search_status_code": status_code, 156 | "search_response": response, 157 | "url_info": url_info, 158 | "html_content": content, 159 | }, 160 | ensure_ascii=False, 161 | ) + "\n" 162 | end = time.time() 163 | print(f"解析网页耗时: {end - start}s") 164 | 165 | return StreamingResponse(html_docs_text_generator(), media_type="application/json") 166 | 167 | 168 | if __name__ == "__main__": 169 | import uvicorn 170 | 171 | uvicorn.run("search_service:app", host="0.0.0.0", port=args.port, reload=True) 172 | -------------------------------------------------------------------------------- /infini_websearch/actions/websearch.py: -------------------------------------------------------------------------------- 1 | import json 2 | from typing import Callable, Dict, Generator, List, Optional 3 | 4 | import requests 5 | from transformers import AutoTokenizer 6 | 7 | from infini_websearch.actions.base_action import BaseAction 8 | 9 | 10 | class GoogleSearch(BaseAction): 11 | def __init__( 12 | self, 13 | server_url: str, 14 | summary_prompt_template: str, 15 | observation_prompt_template: str, 16 | num_search_webpages: int = 5, 17 | webpage_summary_max_input_tokens: int = 2048, 18 | webpage_load_timetout: float = 10.0, 19 | proxies: Optional[Dict] = None, 20 | ) -> None: 21 | self.server_url = server_url 22 | self.summary_prompt_template = summary_prompt_template 23 | self.observation_prompt_template = observation_prompt_template 24 | self.num_search_webpages = num_search_webpages 25 | self.webpage_summary_max_input_tokens = webpage_summary_max_input_tokens 26 | self.webpage_load_timetout = webpage_load_timetout 27 | if proxies is None: 28 | proxies = {"http": None, "https": None} 29 | self.proxies = proxies 30 | 31 | @property 32 | def function_defination(self) -> Optional[Dict]: 33 | return { 34 | "name": "googleWebSearch", 35 | "description": ( 36 | "A Google Search Engine. " 37 | "Useful when you need to search information you don't know such as weather, " 38 | "exchange rate, current events." 39 | "Never ever use this tool when user want to translate" 40 | ), 41 | "parameters": { 42 | "type": "object", 43 | "properties": { 44 | "query": { 45 | "type": "string", 46 | "description": ( 47 | "Content that users want to search for, such as 'weather', 'current events', etc." 48 | "If special characters such as '\n' appear in the search, " 49 | "these special characters must be ignored.\n" 50 | "Chinese characters are preferred." 51 | ), 52 | } 53 | }, 54 | "required": ["query"], 55 | }, 56 | } 57 | 58 | def run( 59 | self, 60 | user_question: str, 61 | arguments: Dict, 62 | llm_completion_funcion: Callable, 63 | tokenizer: AutoTokenizer, 64 | return_webpage_details: bool, 65 | ) -> Generator[Dict, None, None]: 66 | if "query" not in arguments: 67 | return {"observation": "调用工具失败, 缺乏必要输入参数, 请重试"} 68 | 69 | # get webpage content 70 | webpage_detail_list = [] 71 | try: 72 | for webpage_detail in self.streaming_fetch_search_results( 73 | self.server_url, 74 | { 75 | "query": arguments["query"], 76 | "num_search_pages": self.num_search_webpages, 77 | }, 78 | self.proxies, 79 | ): 80 | webpage_detail_list.append(webpage_detail) 81 | if return_webpage_details: 82 | yield webpage_detail 83 | except Exception as e: 84 | print(e) 85 | yield {"observation": '输出"websearch server发生错误, 请重试"'} 86 | return 87 | 88 | webpage_texts = [ 89 | webpage_detail["html_content"] for webpage_detail in webpage_detail_list 90 | ] 91 | 92 | no_webpages_loaded = True 93 | for webpage_text in webpage_texts: 94 | if webpage_text != "搜索页面加载超时, 请重试": 95 | no_webpages_loaded = False 96 | break 97 | 98 | # all web pages are timing out when loading 99 | if no_webpages_loaded: 100 | summaries = webpage_texts 101 | yield {"observation": "搜索页面加载超时, 请重试"} 102 | else: 103 | summary_prompts = self.make_summary_tasks( 104 | query=arguments["query"], 105 | webpage_texts=webpage_texts, 106 | summary_prompt_template=self.summary_prompt_template, 107 | tokenizer=tokenizer, 108 | webpage_summary_max_input_tokens=self.webpage_summary_max_input_tokens, 109 | ) 110 | print("#######summary prompts[start]######") 111 | print(summary_prompts) 112 | print("#######summary prompts[end]######") 113 | response_message = llm_completion_funcion(messages=summary_prompts) 114 | summaries = [choice.text for choice in response_message.choices] 115 | context = "\n".join( 116 | [ 117 | f"[[citation:{str(i+1)}]]\n{summary}" 118 | for i, summary in enumerate(summaries) 119 | ] 120 | ) 121 | yield { 122 | "observation": self.observation_prompt_template.format( 123 | context=context, question=user_question, keywords=arguments["query"] 124 | ) 125 | } 126 | return 127 | 128 | @staticmethod 129 | def make_summary_tasks( 130 | query: str, 131 | webpage_texts: List[str], 132 | summary_prompt_template: str, 133 | tokenizer: AutoTokenizer, 134 | webpage_summary_max_input_tokens: int = 2048, 135 | ) -> List[str]: 136 | messages_all = [] 137 | for webpage_text in webpage_texts: 138 | if len(webpage_text) > 0: 139 | webpage_tokens = tokenizer.encode(webpage_text) 140 | webpage_text = tokenizer.decode( 141 | webpage_tokens[:webpage_summary_max_input_tokens] 142 | ) 143 | messages = [ 144 | {"role": "system", "content": "You are a helpful assistant."}, 145 | { 146 | "role": "user", 147 | "content": summary_prompt_template.format( 148 | question=query, context=webpage_text 149 | ), 150 | }, 151 | ] 152 | messages_all.append( 153 | tokenizer.apply_chat_template( 154 | messages, tokenize=False, add_generation_prompt=True 155 | ) 156 | ) 157 | return messages_all 158 | 159 | @staticmethod 160 | def streaming_fetch_search_results( 161 | url: str, content: Dict, proxies: Dict 162 | ) -> Generator[Dict, None, str]: 163 | try: 164 | with requests.post( 165 | url, json=content, stream=True, proxies=proxies 166 | ) as response: 167 | response.raise_for_status() 168 | for line in response.iter_lines(): 169 | if line: 170 | yield json.loads(line) 171 | except requests.exceptions.HTTPError as error: 172 | print(f"HTTP error occurred: {error}") 173 | return "网页加载超时" 174 | -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | Apache License 2 | Version 2.0, January 2004 3 | http://www.apache.org/licenses/ 4 | 5 | TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION 6 | 7 | 1. Definitions. 8 | 9 | "License" shall mean the terms and conditions for use, reproduction, 10 | and distribution as defined by Sections 1 through 9 of this document. 11 | 12 | "Licensor" shall mean the copyright owner or entity authorized by 13 | the copyright owner that is granting the License. 14 | 15 | "Legal Entity" shall mean the union of the acting entity and all 16 | other entities that control, are controlled by, or are under common 17 | control with that entity. For the purposes of this definition, 18 | "control" means (i) the power, direct or indirect, to cause the 19 | direction or management of such entity, whether by contract or 20 | otherwise, or (ii) ownership of fifty percent (50%) or more of the 21 | outstanding shares, or (iii) beneficial ownership of such entity. 22 | 23 | "You" (or "Your") shall mean an individual or Legal Entity 24 | exercising permissions granted by this License. 25 | 26 | "Source" form shall mean the preferred form for making modifications, 27 | including but not limited to software source code, documentation 28 | source, and configuration files. 29 | 30 | "Object" form shall mean any form resulting from mechanical 31 | transformation or translation of a Source form, including but 32 | not limited to compiled object code, generated documentation, 33 | and conversions to other media types. 34 | 35 | "Work" shall mean the work of authorship, whether in Source or 36 | Object form, made available under the License, as indicated by a 37 | copyright notice that is included in or attached to the work 38 | (an example is provided in the Appendix below). 39 | 40 | "Derivative Works" shall mean any work, whether in Source or Object 41 | form, that is based on (or derived from) the Work and for which the 42 | editorial revisions, annotations, elaborations, or other modifications 43 | represent, as a whole, an original work of authorship. For the purposes 44 | of this License, Derivative Works shall not include works that remain 45 | separable from, or merely link (or bind by name) to the interfaces of, 46 | the Work and Derivative Works thereof. 47 | 48 | "Contribution" shall mean any work of authorship, including 49 | the original version of the Work and any modifications or additions 50 | to that Work or Derivative Works thereof, that is intentionally 51 | submitted to Licensor for inclusion in the Work by the copyright owner 52 | or by an individual or Legal Entity authorized to submit on behalf of 53 | the copyright owner. For the purposes of this definition, "submitted" 54 | means any form of electronic, verbal, or written communication sent 55 | to the Licensor or its representatives, including but not limited to 56 | communication on electronic mailing lists, source code control systems, 57 | and issue tracking systems that are managed by, or on behalf of, the 58 | Licensor for the purpose of discussing and improving the Work, but 59 | excluding communication that is conspicuously marked or otherwise 60 | designated in writing by the copyright owner as "Not a Contribution." 61 | 62 | "Contributor" shall mean Licensor and any individual or Legal Entity 63 | on behalf of whom a Contribution has been received by Licensor and 64 | subsequently incorporated within the Work. 65 | 66 | 2. Grant of Copyright License. Subject to the terms and conditions of 67 | this License, each Contributor hereby grants to You a perpetual, 68 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable 69 | copyright license to reproduce, prepare Derivative Works of, 70 | publicly display, publicly perform, sublicense, and distribute the 71 | Work and such Derivative Works in Source or Object form. 72 | 73 | 3. Grant of Patent License. Subject to the terms and conditions of 74 | this License, each Contributor hereby grants to You a perpetual, 75 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable 76 | (except as stated in this section) patent license to make, have made, 77 | use, offer to sell, sell, import, and otherwise transfer the Work, 78 | where such license applies only to those patent claims licensable 79 | by such Contributor that are necessarily infringed by their 80 | Contribution(s) alone or by combination of their Contribution(s) 81 | with the Work to which such Contribution(s) was submitted. If You 82 | institute patent litigation against any entity (including a 83 | cross-claim or counterclaim in a lawsuit) alleging that the Work 84 | or a Contribution incorporated within the Work constitutes direct 85 | or contributory patent infringement, then any patent licenses 86 | granted to You under this License for that Work shall terminate 87 | as of the date such litigation is filed. 88 | 89 | 4. Redistribution. You may reproduce and distribute copies of the 90 | Work or Derivative Works thereof in any medium, with or without 91 | modifications, and in Source or Object form, provided that You 92 | meet the following conditions: 93 | 94 | (a) You must give any other recipients of the Work or 95 | Derivative Works a copy of this License; and 96 | 97 | (b) You must cause any modified files to carry prominent notices 98 | stating that You changed the files; and 99 | 100 | (c) You must retain, in the Source form of any Derivative Works 101 | that You distribute, all copyright, patent, trademark, and 102 | attribution notices from the Source form of the Work, 103 | excluding those notices that do not pertain to any part of 104 | the Derivative Works; and 105 | 106 | (d) If the Work includes a "NOTICE" text file as part of its 107 | distribution, then any Derivative Works that You distribute must 108 | include a readable copy of the attribution notices contained 109 | within such NOTICE file, excluding those notices that do not 110 | pertain to any part of the Derivative Works, in at least one 111 | of the following places: within a NOTICE text file distributed 112 | as part of the Derivative Works; within the Source form or 113 | documentation, if provided along with the Derivative Works; or, 114 | within a display generated by the Derivative Works, if and 115 | wherever such third-party notices normally appear. The contents 116 | of the NOTICE file are for informational purposes only and 117 | do not modify the License. You may add Your own attribution 118 | notices within Derivative Works that You distribute, alongside 119 | or as an addendum to the NOTICE text from the Work, provided 120 | that such additional attribution notices cannot be construed 121 | as modifying the License. 122 | 123 | You may add Your own copyright statement to Your modifications and 124 | may provide additional or different license terms and conditions 125 | for use, reproduction, or distribution of Your modifications, or 126 | for any such Derivative Works as a whole, provided Your use, 127 | reproduction, and distribution of the Work otherwise complies with 128 | the conditions stated in this License. 129 | 130 | 5. Submission of Contributions. Unless You explicitly state otherwise, 131 | any Contribution intentionally submitted for inclusion in the Work 132 | by You to the Licensor shall be under the terms and conditions of 133 | this License, without any additional terms or conditions. 134 | Notwithstanding the above, nothing herein shall supersede or modify 135 | the terms of any separate license agreement you may have executed 136 | with Licensor regarding such Contributions. 137 | 138 | 6. Trademarks. This License does not grant permission to use the trade 139 | names, trademarks, service marks, or product names of the Licensor, 140 | except as required for reasonable and customary use in describing the 141 | origin of the Work and reproducing the content of the NOTICE file. 142 | 143 | 7. Disclaimer of Warranty. Unless required by applicable law or 144 | agreed to in writing, Licensor provides the Work (and each 145 | Contributor provides its Contributions) on an "AS IS" BASIS, 146 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or 147 | implied, including, without limitation, any warranties or conditions 148 | of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A 149 | PARTICULAR PURPOSE. You are solely responsible for determining the 150 | appropriateness of using or redistributing the Work and assume any 151 | risks associated with Your exercise of permissions under this License. 152 | 153 | 8. Limitation of Liability. In no event and under no legal theory, 154 | whether in tort (including negligence), contract, or otherwise, 155 | unless required by applicable law (such as deliberate and grossly 156 | negligent acts) or agreed to in writing, shall any Contributor be 157 | liable to You for damages, including any direct, indirect, special, 158 | incidental, or consequential damages of any character arising as a 159 | result of this License or out of the use or inability to use the 160 | Work (including but not limited to damages for loss of goodwill, 161 | work stoppage, computer failure or malfunction, or any and all 162 | other commercial damages or losses), even if such Contributor 163 | has been advised of the possibility of such damages. 164 | 165 | 9. Accepting Warranty or Additional Liability. While redistributing 166 | the Work or Derivative Works thereof, You may choose to offer, 167 | and charge a fee for, acceptance of support, warranty, indemnity, 168 | or other liability obligations and/or rights consistent with this 169 | License. However, in accepting such obligations, You may act only 170 | on Your own behalf and on Your sole responsibility, not on behalf 171 | of any other Contributor, and only if You agree to indemnify, 172 | defend, and hold each Contributor harmless for any liability 173 | incurred by, or claims asserted against, such Contributor by reason 174 | of your accepting any such warranty or additional liability. 175 | 176 | END OF TERMS AND CONDITIONS 177 | 178 | APPENDIX: How to apply the Apache License to your work. 179 | 180 | To apply the Apache License to your work, attach the following 181 | boilerplate notice, with the fields enclosed by brackets "[]" 182 | replaced with your own identifying information. (Don't include 183 | the brackets!) The text should be enclosed in the appropriate 184 | comment syntax for the file format. We also recommend that a 185 | file or class name and description of purpose be included on the 186 | same "printed page" as the copyright notice for easier 187 | identification within third-party archives. 188 | 189 | Copyright [yyyy] [name of copyright owner] 190 | 191 | Licensed under the Apache License, Version 2.0 (the "License"); 192 | you may not use this file except in compliance with the License. 193 | You may obtain a copy of the License at 194 | 195 | http://www.apache.org/licenses/LICENSE-2.0 196 | 197 | Unless required by applicable law or agreed to in writing, software 198 | distributed under the License is distributed on an "AS IS" BASIS, 199 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 200 | See the License for the specific language governing permissions and 201 | limitations under the License. 202 | -------------------------------------------------------------------------------- /gradio_app.py: -------------------------------------------------------------------------------- 1 | import argparse 2 | import os 3 | from typing import Dict, Generator, List, Optional, Tuple 4 | 5 | import gradio as gr 6 | from gradio_toggle import Toggle 7 | from transformers import AutoTokenizer 8 | 9 | from infini_websearch.actions import GoogleSearch, parse_function_call_from_model_ouput 10 | from infini_websearch.configs import ( 11 | AGENT_MAX_OUTPUT_TOKENS, 12 | AGENT_TEMPERATURE, 13 | CHAT_MAX_OUTPUT_TOKENS, 14 | CHAT_TEMPERATURE, 15 | CSS_STYLE, 16 | FUNCTION_CALLING_PROMPT_TEMPLATE, 17 | FUNCTION_END_TOKEN, 18 | FUNCTION_START_TOKEN, 19 | MAX_ACTION_TURNS, 20 | MODEL_NAME, 21 | MODEL_SERVER_URL, 22 | NUM_SEARCH_WEBPAGES, 23 | OBSERVATION_PROMPT_TEMPLATE, 24 | PROXIES, 25 | ROLE_PROMPT, 26 | SEARCH_SERVER_URL, 27 | SESSION_MAX_INPUT_TOKENS, 28 | SESSION_WINDOW_SIZE, 29 | STOP_TOKENS, 30 | SUMMARY_PROMPT_TEMPLATE, 31 | TIME_PROMPT_TEMPLATE, 32 | WEBPAGE_LOAD_TIMETOUT, 33 | WEBPAGE_SUMMARY_MAX_INPUT_TOKENS, 34 | WEBPAGE_SUMMARY_MAX_OUTPUT_TOKENS, 35 | ) 36 | from infini_websearch.model import ( 37 | get_vllm_model_output_function, 38 | include_special_tokens, 39 | split_text_by_special_token, 40 | ) 41 | from infini_websearch.utils import ( 42 | extract_citations, 43 | format_search_results, 44 | functions2str, 45 | get_datetime_now, 46 | ) 47 | 48 | parser = argparse.ArgumentParser() 49 | parser.add_argument("--model-path", "-m", type=str) 50 | parser.add_argument("--port", type=int, default=7860) 51 | 52 | args = parser.parse_args() 53 | 54 | MODEL_PATH = args.model_path 55 | SERVER_PORT = args.port 56 | 57 | # tokenizer 58 | os.environ["TOKENIZERS_PARALLELISM"] = "false" 59 | TOKENIZER = AutoTokenizer.from_pretrained(MODEL_PATH, trust_remote_code=True) 60 | 61 | 62 | # function name -> action 63 | ACTIONS_MAP = { 64 | "googleWebSearch": GoogleSearch( 65 | server_url=SEARCH_SERVER_URL, 66 | num_search_webpages=NUM_SEARCH_WEBPAGES, 67 | summary_prompt_template=SUMMARY_PROMPT_TEMPLATE, 68 | observation_prompt_template=OBSERVATION_PROMPT_TEMPLATE, 69 | webpage_summary_max_input_tokens=WEBPAGE_SUMMARY_MAX_INPUT_TOKENS, 70 | webpage_load_timetout=WEBPAGE_LOAD_TIMETOUT, 71 | proxies=PROXIES, 72 | ), 73 | } 74 | # tool -> function name 75 | TOOLS_TO_ACTION_NAMES = { 76 | "websearch": "googleWebSearch", 77 | } 78 | 79 | 80 | def get_system_prompt(functions: Optional[List] = None) -> str: 81 | """ 82 | Get system prompt for current conversation. 83 | """ 84 | if functions is None: 85 | functions = [] 86 | current_time, weekday = get_datetime_now() 87 | time_info = TIME_PROMPT_TEMPLATE.format(current_time=current_time, weekday=weekday) 88 | system_prompt = ROLE_PROMPT + "\n" + time_info 89 | if len(functions) > 0: 90 | system_prompt += "\n" + FUNCTION_CALLING_PROMPT_TEMPLATE.format( 91 | functions=functions2str(functions) 92 | ) 93 | return system_prompt 94 | 95 | 96 | def user( 97 | user_message: str, history: List[Dict], session_state: gr.State 98 | ) -> Tuple[str, List[Dict], gr.State]: 99 | """ 100 | Add user input message to history. 101 | """ 102 | session_state["messages"] += [{"role": "user", "content": user_message}] 103 | return "", history + [{"role": "user", "content": user_message}], session_state 104 | 105 | 106 | def bot( 107 | history: List[Dict], 108 | websearch: bool, 109 | session_state: gr.State, 110 | ) -> Generator[List[Dict], None, None]: 111 | """ 112 | Main workflow. 113 | """ 114 | # get registered tools 115 | registered_tools = [] 116 | if websearch is True: 117 | registered_tools.append("websearch") 118 | temperature = AGENT_TEMPERATURE 119 | max_gen_length = AGENT_MAX_OUTPUT_TOKENS 120 | else: 121 | temperature = CHAT_TEMPERATURE 122 | max_gen_length = CHAT_MAX_OUTPUT_TOKENS 123 | registered_function_names = [ 124 | TOOLS_TO_ACTION_NAMES[tool] for tool in registered_tools 125 | ] 126 | registered_functions = [ 127 | ACTIONS_MAP[function_name] for function_name in registered_function_names 128 | ] 129 | functions = [ 130 | function.function_defination 131 | for function in registered_functions 132 | if function.function_defination is not None 133 | ] 134 | 135 | # get system prompt 136 | system_prompt = get_system_prompt(functions=functions) 137 | # get model streaming output function 138 | llm_streaming_output_func = get_vllm_model_output_function( 139 | url=MODEL_SERVER_URL, 140 | model_name=MODEL_NAME, 141 | chat_mode=True, 142 | stream=True, 143 | model_config={ 144 | "temperature": temperature, 145 | "max_tokens": max_gen_length, 146 | "stop": STOP_TOKENS, 147 | }, 148 | ) 149 | 150 | input_dict = { 151 | "temperature": temperature, 152 | "max_gen_length": max_gen_length, 153 | "use_websearch": "websearch" in registered_tools, 154 | "MODEL_NAME": MODEL_NAME, 155 | } 156 | 157 | for _ in range(MAX_ACTION_TURNS * 2): 158 | # ASSISTANT answers two times per action turn 159 | # answer1: <|function_start|>xxx<|function_end|> 160 | # answer2: observation -> final answer 161 | 162 | """ 163 | retain SESSION_WINDOW_SIZE turns (max_sequence_length is short -> 4096) 164 | """ 165 | messages_truncated = truncate_messages( 166 | messages=session_state["messages"], 167 | tokenizer=TOKENIZER, 168 | session_window_size=SESSION_WINDOW_SIZE, 169 | max_input_tokens=SESSION_MAX_INPUT_TOKENS, 170 | system_prompt=system_prompt, 171 | ) 172 | 173 | messages_input = [ 174 | {"role": "system", "content": system_prompt} 175 | ] + messages_truncated 176 | input_dict.update(dict(messages=messages_input)) 177 | # print input prompt 178 | print( 179 | TOKENIZER.apply_chat_template( 180 | messages_input, tokenize=False, add_generation_prompt=True 181 | ) 182 | ) 183 | 184 | response_raw = "" 185 | response_gradio = "" 186 | """ 187 | streaming output status: 188 | 1. [chat]: generating chat message 189 | 2. [function start]: start generating function calling information ([chat] -> [function]) 190 | 3. [function]: generating function calling information 191 | 4. [function end]: end generating function calling information ([function] -> [chat]) 192 | """ 193 | # in [function] status? 194 | function_status = False 195 | chunk_buffer = "" 196 | for chunk in llm_streaming_output_func(messages=input_dict["messages"]): 197 | chunk_buffer += chunk 198 | # '<|function_start|>' and '<|function_end|>' appear to be truncated ? 199 | if chunk_buffer.rfind("|>") < chunk_buffer.rfind("<|"): 200 | continue 201 | # [function start] status: ([chat] -> [function]) 202 | if function_status is False and include_special_tokens( 203 | chunk_buffer, [FUNCTION_START_TOKEN] 204 | ): 205 | function_status = True 206 | chat_part, tool_part = split_text_by_special_token( 207 | chunk_buffer, FUNCTION_START_TOKEN 208 | ) 209 | tool_part = FUNCTION_START_TOKEN + tool_part 210 | # add chat message to history 211 | if len(response_gradio + chat_part) > 0: 212 | history.append( 213 | {"role": "assistant", "content": response_gradio + chat_part} 214 | ) 215 | response_gradio = tool_part 216 | response_raw += chat_part + tool_part 217 | chunk_buffer = "" 218 | yield history + [ 219 | { 220 | "role": "assistant", 221 | "content": response_gradio, 222 | "metadata": {"title": "tool parameters"}, 223 | } 224 | ] 225 | # [function end] status: ([function] -> [chat]) 226 | elif function_status is True and include_special_tokens( 227 | chunk_buffer, [FUNCTION_END_TOKEN] 228 | ): 229 | chat_part, _ = split_text_by_special_token( 230 | chunk_buffer, FUNCTION_END_TOKEN 231 | ) 232 | response_gradio += chat_part + FUNCTION_END_TOKEN 233 | response_raw += chat_part + FUNCTION_END_TOKEN 234 | chunk_buffer = "" 235 | history.append( 236 | { 237 | "role": "assistant", 238 | "content": response_gradio, 239 | "metadata": {"title": "tool parameters"}, 240 | } 241 | ) 242 | yield history 243 | break 244 | # [function] status 245 | elif function_status is True: 246 | response_gradio += chunk_buffer 247 | response_raw += chunk_buffer 248 | chunk_buffer = "" 249 | yield history + [ 250 | { 251 | "role": "assistant", 252 | "content": response_gradio, 253 | "metadata": {"title": "tool parameters"}, 254 | } 255 | ] 256 | # [chat] status 257 | elif function_status is False: 258 | citations = extract_citations(chunk_buffer) 259 | if len(citations) > 0 and len(session_state["url_infos"]) > 0: 260 | chunk_new = chunk_buffer 261 | for citation in citations: 262 | url_ind = int(citation) - 1 263 | # hardcoding for out-of-bounds 264 | if url_ind < 0: 265 | url_ind = 0 266 | elif url_ind >= len(session_state["url_infos"]): 267 | url_ind = len(session_state["url_infos"]) - 1 268 | # Add a space before the tag to prevent rendering 269 | # errors when multiple tags are adjacent to 270 | # each other. 271 | chunk_new = chunk_new.replace( 272 | f"[citation:{citation}]", 273 | f' {citation}', # noqa: E501 274 | ) 275 | response_gradio += chunk_new 276 | else: 277 | response_gradio += chunk_buffer 278 | response_raw += chunk_buffer 279 | 280 | chunk_buffer = "" 281 | yield history + [{"role": "assistant", "content": response_gradio}] 282 | 283 | if session_state["stop_generation"] is True: 284 | session_state["stop_generation"] = False 285 | break 286 | 287 | # if streaming ends with [chat] status, add response to history 288 | if not include_special_tokens(response_gradio, FUNCTION_END_TOKEN): 289 | history.append({"role": "assistant", "content": response_gradio}) 290 | 291 | session_state["messages"].append({"role": "assistant", "content": response_raw}) 292 | 293 | # no tool registered, end this turn 294 | if len(registered_tools) == 0: 295 | break 296 | 297 | function_name, function_arguments = parse_function_call_from_model_ouput( 298 | response_raw, 299 | registered_function_names, 300 | speical_tokens_map=dict( 301 | function_start_token=FUNCTION_START_TOKEN, 302 | function_end_token=FUNCTION_END_TOKEN, 303 | ), 304 | ) 305 | 306 | # no tool use this turn, end this turn 307 | if function_arguments is None: 308 | break 309 | 310 | # something is wrong, use function_arguments as observation (error 311 | # message) 312 | if function_name is None and isinstance(function_arguments, str): 313 | history.append({"role": "observation", "content": function_arguments}) 314 | session_state["messages"].append( 315 | {"role": "observation", "content": function_arguments} 316 | ) 317 | continue 318 | 319 | url_infos, html_contents = [], [] 320 | latest_tool_response = None 321 | action = ACTIONS_MAP[function_name] 322 | observation = None 323 | if function_name == "googleWebSearch": 324 | observation_genrator = action.run( 325 | user_question=session_state["messages"][-2]["content"], 326 | arguments=function_arguments, 327 | llm_completion_funcion=get_vllm_model_output_function( 328 | url=MODEL_SERVER_URL, 329 | model_name=MODEL_NAME, 330 | chat_mode=False, 331 | stream=False, 332 | model_config={ 333 | "temperature": temperature, 334 | "max_tokens": WEBPAGE_SUMMARY_MAX_OUTPUT_TOKENS, 335 | "stop": STOP_TOKENS, 336 | }, 337 | ), 338 | tokenizer=TOKENIZER, 339 | return_webpage_details=True, 340 | ) 341 | for item in gr.Progress().tqdm(observation_genrator, desc="summarizing..."): 342 | if isinstance(item, dict): 343 | if "observation" in item: 344 | observation = item["observation"] 345 | break 346 | url_infos.append(item["url_info"]) 347 | html_contents.append(item["html_content"]) 348 | yield history + [ 349 | { 350 | "role": "assistant", 351 | "content": format_search_results(url_infos), 352 | "metadata": {"title": "tool results"}, 353 | } 354 | ] 355 | # error message 356 | elif isinstance(item, str): 357 | latest_tool_response = item 358 | yield history + [ 359 | { 360 | "role": "assistant", 361 | "content": latest_tool_response, 362 | "metadata": {"title": "tool results"}, 363 | } 364 | ] 365 | else: 366 | raise NotImplementedError 367 | 368 | if len(url_infos) > 0: 369 | history.append( 370 | { 371 | "role": "assistant", 372 | "content": format_search_results(url_infos), 373 | "metadata": {"title": "tool results"}, 374 | } 375 | ) 376 | else: 377 | history.append( 378 | { 379 | "role": "assistant", 380 | "content": latest_tool_response, 381 | "metadata": {"title": "tool results"}, 382 | } 383 | ) 384 | 385 | # update url_infos 386 | session_state["url_infos"] = url_infos 387 | else: 388 | observation = action.run(function_arguments) 389 | 390 | assert observation is not None 391 | history.append({"role": "observation", "content": observation}) 392 | session_state["messages"].append( 393 | {"role": "observation", "content": observation} 394 | ) 395 | 396 | 397 | def truncate_messages( 398 | messages: List[Dict], 399 | tokenizer: AutoTokenizer, 400 | session_window_size: int, 401 | max_input_tokens: int, 402 | system_prompt: str, 403 | ) -> List[Dict]: 404 | """ 405 | truncate messages for model input by session_window_size and max_input_tokens 406 | """ 407 | # get parts for each turn 408 | turn_start_inds = [] 409 | for ind, message in enumerate(messages): 410 | if message["role"] == "user": 411 | turn_start_inds.append(ind) 412 | # only latest turns are used as input 413 | turn_start_inds_used = turn_start_inds[-session_window_size:] 414 | messages_parts = [] 415 | for i in range(len(turn_start_inds_used)): 416 | turn_start_ind = turn_start_inds_used[i] 417 | turn_end_ind = ( 418 | len(messages) 419 | if i + 1 >= len(turn_start_inds_used) 420 | else turn_start_inds_used[i + 1] 421 | ) 422 | messages_parts.append(messages[turn_start_ind:turn_end_ind]) 423 | # truncate by max_input_tokens 424 | messages_truncated = [] 425 | for i, messages_part in enumerate(reversed(messages_parts)): 426 | if ( 427 | i == 0 428 | or len( 429 | tokenizer.apply_chat_template( 430 | [{"role": "system", "content": system_prompt}] 431 | + messages_truncated 432 | + messages_part, 433 | tokenize=True, 434 | ) 435 | ) 436 | < max_input_tokens 437 | ): 438 | messages_truncated = messages_part + messages_truncated 439 | else: 440 | break 441 | return messages_truncated 442 | 443 | 444 | def stop_response(session_state: gr.State) -> gr.State: 445 | session_state["stop_generation"] = True 446 | return session_state 447 | 448 | 449 | def clear(history: List[Dict], session_state: gr.State) -> Tuple[List[Dict], gr.State]: 450 | session_state["messages"] = [] 451 | session_state["url_infos"] = [] 452 | session_state["stop_generation"] = False 453 | return [], session_state 454 | 455 | 456 | def toggle_change(session_state: gr.State) -> gr.State: 457 | session_state["messages"] = [] 458 | return session_state 459 | 460 | 461 | with gr.Blocks( 462 | css=CSS_STYLE, fill_height=True, elem_classes="canvas", theme=gr.themes.Monochrome() 463 | ) as demo: 464 | # chatbot interface 465 | with gr.Row(equal_height=False, variant="compact"): 466 | with gr.Column(scale=1.0, elem_classes="fullheight"): 467 | chatbot = gr.Chatbot( 468 | type="messages", 469 | elem_classes="chatbot", 470 | label="infini-websearch", 471 | autoscroll=True, 472 | ) 473 | 474 | # conversation state vars 475 | session_state = gr.State( 476 | dict( 477 | messages=[], 478 | url_infos=[], 479 | stop_generation=False, 480 | ) 481 | ) 482 | toggle_is_interactive = gr.State(value=True) 483 | 484 | # bottom bar 485 | with gr.Group(elem_classes="bottom-bar") as bottom_bar: 486 | msg = gr.Textbox(label="question") 487 | with gr.Row(): 488 | clear_btn = gr.Button("Clear") 489 | stop_btn = gr.Button("Stop") 490 | 491 | # toggle 492 | with gr.Group() as toggle_group: 493 | websearch = Toggle( 494 | label="websearch", 495 | value=True, 496 | interactive=True, 497 | ) 498 | 499 | websearch.change(toggle_change, [session_state], [session_state]) 500 | msg.submit( 501 | user, [msg, chatbot, session_state], outputs=[msg, chatbot, session_state] 502 | ).then( 503 | bot, 504 | [chatbot, websearch, session_state], 505 | outputs=[chatbot], 506 | concurrency_limit=2, 507 | ) 508 | clear_btn.click(clear, [chatbot, session_state], outputs=[chatbot, session_state]) 509 | stop_btn.click(stop_response, [session_state], outputs=[session_state], queue=False) 510 | 511 | 512 | if __name__ == "__main__": 513 | demo.launch(share=False, server_port=SERVER_PORT) 514 | --------------------------------------------------------------------------------