app.py

from langchain.agents import AgentExecutor, create_openai_functions_agent, create_react_agent
from langchain_community.tools.tavily_search import TavilySearchResults
from langchain_openai import ChatOpenAI
import os
from langchain_core.output_parsers import StrOutputParser, PydanticToolsParser
from langchain_core.prompts import ChatPromptTemplate
from serpapi import GoogleSearch
from langchain.agents import initialize_agent, Tool
import gradio as gr
from css import *
import time
from openai import OpenAI
from PIL import Image
from langchain.schema import (
    HumanMessage,
    SystemMessage,
)
import base64
import io

os.environ["OPENAI_API_KEY"] = "sb-6a683cb3bd63a9b72040aa2dd08feff8b68f08a0e1d959f5"
os.environ['OPENAI_BASE_URL'] = "https://api.openai-sb.com/v1/"
os.environ["SERPAPI_API_KEY"] = "dcc98b22d5f7d413979a175ff7d75b721c5992a3ee1e2363020b2bbdf4f82404"
# os.environ['TAVILY_API_KEY'] = "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" #you
os.environ['TAVILY_API_KEY'] = "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" #zeng
os.environ["REPLICATE_API_TOKEN"] = "r8_IYJpjwjrxegcUfBeBbyUxErJXXsnHDM4AlSQQ"

llm = ChatOpenAI(model="gpt-3.5-turbo-0125", temperature=0.6)

def search(query) -> str:
    params = {
        "engine": "baidu_news",
        "q": query,
        "ct": "1",
        "api_key": "dcc98b22d5f7d413979a175ff7d75b721c5992a3ee1e2363020b2bbdf4f82404"
    }
    search = GoogleSearch(params)
    results = search.get_dict()
    organic_results = results.get("organic_results", [])
    if not organic_results:
        return ""
    final_output = "\n\n".join([
        f"Title: {news.get('title', 'No Title')}\nSnippet: {news.get('snippet', 'No Snippet')}\nDate: {news.get('date', 'No Date')}\nSource:{news.get('source', 'No Source')}"
        for news in organic_results
    ])
    return final_output

def img_size(image):
    # img = Image.open(image)
    img = Image.fromarray(image.astype("uint8"))
    width, height = img.size
    while width >= 500 or height >= 400:
        width = width * 0.8
        height = height * 0.8
    width = int(width)
    height = int(height)
    resized_img = img.resize((width, height))
    # resized_img.save(image_path)
    return resized_img

def search_image(query) -> str:
    params = {
        "engine": "google_images",
        "q": query,
        "gl": "cn",
    }
    search = GoogleSearch(params)
    results = search.get_dict()
    suggested_searches = results.get('suggested_searches', [])
    if suggested_searches:
        thumbnails = [search['thumbnail'] for search in suggested_searches][:3]
    else:
        thumbnails = []
    # thumbnails = [search['thumbnail'] for search in results['suggested_searches']][:3]
    information = ""
    client = OpenAI()
    for idx, thumbnail in enumerate(thumbnails):
        response = client.chat.completions.create(
            model="gpt-4o-mini",
            messages=[
                {
                    "role": "user",
                    "content": [
                        {"type": "text", "text": "What’s in this image?Give a brief answer"},
                        {
                            "type": "image_url",
                            "image_url": {
                                "url": thumbnail,
                            },
                        },
                    ],
                }
            ],
            max_tokens=100,
        )
        response_content = response.choices[0].message.content
        information = information + str(idx + 1) + ": " + response_content + "\n"
    return (information)

def encode_image(image):
    image = Image.fromarray(image.astype("uint8"))
    buffered = io.BytesIO()
    image.save(buffered, format="PNG")
    return base64.b64encode(buffered.getvalue()).decode("utf-8")
    # with open(image_path, "rb") as image_file:
    #     return base64.b64encode(image_file.read()).decode("utf-8")


def image_summarize(img_base64, prompt):
    chat = ChatOpenAI(model="gpt-4o", max_tokens=256)
    msg = chat.invoke(
        [
            HumanMessage(
                content=[
                    {"type": "text", "text": prompt},
                    {
                        "type": "image_url",
                        "image_url": {"url": f"data:image/jpeg;base64,{img_base64}"},
                    },
                ]
            )
        ]
    )
    return msg.content


def generate_img_summaries(image):
    # img_size(path)
    image_summaries = []
    prompt = """You are an assistant responsible for compiling images for retrieval\
        These abstracts will be embedded and used to retrieve the original images\
        Provide a detailed summary of the optimized images for retrieval."""
    base64_image = encode_image(image)
    image_summaries.append(image_summarize(base64_image, prompt))
    return image_summaries

search_tool = Tool(
    name="Baidu News Search",  # 工具名称
    func=search,  # 引用search函数
    description="备用搜索引擎，当主要搜索引擎不返回结果时调用，输入是检索query"  # 工具描述
)

search_image = Tool(
    name="Image Search",  # 工具名称
    func=search_image,  # 引用search函数
    description="搜索图片引擎，当你需要检索相关图片的时候调用，输入是检索query，输出是与query有关的图片的信息"  # 工具描述
)
en_prompt = """
Remember, you are an AI called Anti-Faker, used for misinformation detection.
You should asses the authenticity of every detail in the input information as thoroughly as possible.
Your goal is to use the provided tools :{tools} and information about news related images:{image_information} to evaluate the truthfulness of the claims in the news article.

Use the following format:
Question: the input information (e.g., claim, post, news) you must assess.

Thought: to evaluate first detail in input information, you should always think about what to do.

Action: the action to take, should be one of [{tool_names}].

Action Input: the input to the action.

Observation: the result of the action.

Thought: to evaluate second detail in input information, you should always think about what to do.

Action: the action to take, should be one of [{tool_names}].

Action Input: the input to the action.

Observation: the result of the action.

... (this Thought/Action/Action Input/Observation can repeat N times)

Thought: I now know the final answer.

Final Answer: The definitive response, including reasons on all details and a conclusion, for assessing the original input information.

Additional Notes:
1. If the input information contains multiple statements or details, you'd better to assess them one by one.
2. The results of tools may also be incorrect, pay attention to conflicts and distinguishing them base on your knowledge.
3. The conclusion in the Final Answer should include the detected labels,and you need to give a clear label without any possible label results.
( For information lacking clear evidence (e.g., rumors), candidate labels are: [Highly Likely Correct, Highly Likely Incorrect, Slightly Incorrect, Slightly Correct, Neutral]; 
For information with clear evidence (e.g., verified by specific proof), candidate labels are: [Completely Incorrect, Mostly Incorrect, Mixed Authenticity, Mostly Correct, Completely Correct].)
4. You should answer in English.
5. If you feel the need to search for image information. You can use the tools in {tools} to obtain information
Begin!
Question: {input}
image_information: {image_information}
Thought:{agent_scratchpad}
"""

ch_prompt = '''
记住, 你是一个人工智能工具, 名叫'鉴伪者', 用于虚假信息检测.
请尽可能详细地评估 输入信息 中每个细节的真实性.
您的目标是使用提供的工具：{tools}和有关新闻相关图像的信息：{image_information}来评估新闻文章中声明的真实性。

使用以下格式：
Question: 您必须评估的 输入信息(例如, 声明、帖子、新闻).

Thought: 要评估输入信息中的第一个细节, 你应该怎么做.

Action: 只能是此工具列表({tool_names})之一的工具名，不能有其他内容

Action Input: 行为的输入.

Observation: 行动的执行结果.

Thought: 要评估输入信息中的第二个细节, 你应该怎么做.

Action: 只能是此工具列表({tool_names})之一的工具名，不能有其他内容

Action Input: 动作的输入.

Observation: 行动的结果.

... (这个 'Thought/Action/Action Input/Observation' 逻辑可以最多重复5次)

Thought: 我现在知道最后的答案了.

Final Answer: 对输入信息的评估结果 (包括对每个细节的推理验证或者预测, 整体评估结果以及相关URL).

补充说明：
1.如果输入信息包含多个陈述或细节, 你最好逐一评估 ,请注意对时间,地点的判断.
2.检索工具的结果也可能不正确, 请注意其中的冲突, 并根据你的知识进行区分.
3.Final Answer的评估结果应包含一个检测标签, 你需要给出一个明确的标签，没有任何“可能的标签结果”。
对于缺乏明确证据进行严格验证的信息(如谣言), 需预测标签, 候选标签为: 极有可能正确、极有可能不正确、可能错误、可能正确、中性;
对于有明确证据进行验证的信息, 候选标签为: 完全不正确、大部分不正确、真假参半、大部分正确、完全正确.
4.你的Final Answer需要用中文回答，必须给出明确的标签和解释.
5.如果你觉得需要搜索图像信息。您可以使用｛tools｝中的工具来获取信息
开始！
Question: {input}
image_information: {image_information}
Thought: {agent_scratchpad}
'''

# origin_prompt = ChatPromptTemplate.from_template(template)
# chain = origin_prompt | llm | StrOutputParser()

selected_language = "ch"

def SelectLanguage(option):
    global selected_language
    if option == "英文":
        selected_language = "en"
    else:
        selected_language = "ch"
language_option = ["中文","英文"]

def react(input,image_information):
    tools = [TavilySearchResults(max_result=1), search_image]
    if selected_language == "en":
        origin_prompt = ChatPromptTemplate.from_template(template_en)
        prompt = ChatPromptTemplate.from_template(en_prompt)
    elif selected_language == "ch":
        prompt = ChatPromptTemplate.from_template(ch_prompt)
        origin_prompt = ChatPromptTemplate.from_template(template_zh)
    chain = origin_prompt | llm | StrOutputParser()
    ori_result = chain.invoke(input)
    agent = create_react_agent(llm, tools, prompt)
    agent_executor = AgentExecutor(agent=agent, tools=tools, verbose=True, handle_parsing_errors=True)
    result = agent_executor.invoke({"input": input, "image_information": image_information})
    return result['output'], ori_result

    # result = react(
    #     "On September 23rd, the leaders of the United States, Japan, India, and Australia discussed and issued a joint statement on issues such as quadripartite security cooperation, regional and international situations",
    #     "en")
    # print(result['output'])

    # result = result['output']
    #
    # translator = Translator(from_lang="EN", to_lang="ZH")
    # if len(result) >= 500:
    #     result_cn = []
    #     while len(result) > 300:
    #         pos = min(result[300:].find('.'), result[300:].find(','))
    #         sub_result = result[:300+pos]
    #         result_cn = str(result_cn) + str(translator.translate(sub_result))
    #         result = result[300+pos:]
    #     result_cn = str(result_cn) + str(translator.translate(result))
    # else:
    #     result_cn = translator.translate(result)
    # print(result_cn[2:])

    # with gr.Blocks(css=css1) as demo:
    #     gr.Markdown("<h1 style='font-size: 36px; text-align: center; color: #333333; margin-bottom: 20px;'>虚假信息检测</h1>")
    #
    #     with gr.Row():
    #         with gr.Column(scale=3):
    #             input_text = gr.Textbox(label="Enter your news", elem_classes="gradio-input")
    #             language_text = gr.Textbox(label="language", elem_classes="gradio-input")
    #             greet_button = gr.Button("提交", elem_classes="gradio-button")
    #         with gr.Column(scale=2):
    #             output_text = gr.Textbox(label="结果", elem_classes="gradio-output", lines=5)

# result1,result2 = react("On September 23rd, the leaders of the United States, Japan, India, and Australia discussed and issued a joint statement on issues such as quadripartite security cooperation, regional and international situations")
# print(result1)
# print(result2)
title = "# 虚假信息检测"

with (gr.Blocks(css=css1) as demo):
    gr.Markdown(title, elem_id="title")
    with gr.Row():
        with gr.Column(scale=4):
            chatbot = gr.Chatbot(elem_classes="gradio-output")
            language_select = gr.Dropdown(choices = language_option, elem_classes="gradio-input", label= "请选择要使用的语言")
            # input_box = gr.Textbox(label="输入", elem_classes="gradio-input", placeholder="请输入要判断的新闻", lines=3)
            img_info = gr.Textbox(label="提取到的信息", lines=5, elem_classes="gradio-output")
            # ans_box = gr.Textbox(label="gpt-4o", lines=5, elem_classes="gradio-output")


        with gr.Column(scale=2):
            # ans_box = gr.Textbox(label="gpt-4o", lines=5, elem_classes="gradio-output")
            ans_box = gr.Textbox(label="gpt-4o", lines=5, elem_classes="gradio-output")
            img_input = gr.Image(label="上传图像", type="numpy")
            img_output = gr.Image(label="处理后的图像", type="numpy", visible=False)
            react_box = gr.Textbox(label="React", lines=5, elem_classes="gradio-output", visible=False)
            input_box = gr.Textbox(label="输入", elem_classes="gradio-input", placeholder="请输入要判断的新闻", lines=3)
            with gr.Row():
                # with gr.Column(scale=1):
                clear = gr.Button("清空页面", elem_classes="gradio-button", scale=1)
                # with gr.Column(scale=1):
                submit_btn = gr.Button("提交", elem_classes="gradio-button", scale=1)
            # clear = gr.Button("清空页面", elem_classes="gradio-button")
            # submit_btn = gr.Button("提交", elem_classes="gradio-button")
        language_select.change(SelectLanguage, language_select)

    def user(user_input, history):
        if history is None:
            history = []
        return user_input, history + [[user_input, None]]

    def bot(history, rag_box):
        if history is None or len(history) == 0:
            return
        bot_rag = str(rag_box)
        history[-1][1] = ""
        for character in bot_rag:
            history[-1][1] += character
            time.sleep(0.01)
            yield history

    submit_btn.click(img_size, img_input, img_output).then(
        generate_img_summaries, img_output, img_info
    ).then(
        react, [input_box,img_info], [react_box, ans_box]
    ).then(
        user, [input_box, chatbot], [input_box, chatbot]
    ).then(
        bot, [chatbot, react_box], chatbot
    )
    clear.click(lambda: (None, None, None, None, None, None), inputs=None, outputs=[chatbot, ans_box, react_box, img_input, img_info, img_output])

if __name__ == "__main__":
    demo.launch()