app.py

import json
import os
import torch
from sentence_transformers import SentenceTransformer
import faiss
import numpy as np
import gradio as gr
import openai
import re

#openai.api_key = 'xxx'

def load_or_create_model_and_embeddings(model_name, data_file, output_dir):
    model_path = os.path.join(output_dir, 'saved_model')
    embeddings_path = os.path.join(output_dir, 'corpus_embeddings.pt')
    if os.path.exists(model_path) and os.path.exists(embeddings_path):
        model = SentenceTransformer(model_path)
        embeddings = torch.load(embeddings_path)
        with open(data_file, 'r', encoding='utf-8') as f:
            data = json.load(f)
    else:
        model = SentenceTransformer(model_name)
        with open(data_file, 'r', encoding='utf-8') as f:
            data = json.load(f)
        texts = [item['text'] for item in data]
        embeddings = model.encode(texts, convert_to_tensor=True)
        model.save(model_path)
        torch.save(embeddings, embeddings_path)
    return model, embeddings, data


model_name = 'sentence-transformers/all-MiniLM-L6-v2'
data_file = 'labeled_cti_data.json'
output_dir = '.'


model, embeddings, data = load_or_create_model_and_embeddings(model_name, data_file, output_dir)


dimension = embeddings.shape[1]
index = faiss.IndexFlatL2(dimension)
index.add(embeddings.cpu().numpy().astype('float32'))

def get_entity_groups(entities):
    return list(set(entity['entity_group'] for entity in entities))

def get_color_for_entity(entity_group):
    colors = {
        'SamFile': '#EE8434',  # Orange (wheel)
        'Way': '#C95D63',      # Indian red
        'Idus': '#AE8799',     # Mountbatten pink
        'Tool': '#9083AE',     # African Violet
        'Features': '#8181B9', # Tropical indigo
        'HackOrg': '#496DDB',  # Royal Blue (web color)
        'Purp': '#BCD8C1',     # Celadon
        'OffAct': '#D6DBB2',   # Vanilla
        'Org': '#E3D985',      # Flax
        'SecTeam': '#E57A44',  # Orange (Crayola)
        'Time': '#E3D985',     # Dark purple
        'Exp': '#5D76CF',      # Glaucous
        'Area': '#757FC1',     # Another shade of blue
    }
    return colors.get(entity_group, '#000000')  # Default to black if entity group not found

def semantic_search(query, top_k=5):
    query_embedding = model.encode([query], convert_to_tensor=True)
    distances, indices = index.search(query_embedding.cpu().numpy().astype('float32'), top_k)
    
    results = []
    for distance, idx in zip(distances[0], indices[0]):
        similarity_score = 1 - distance / 2  # 將距離轉換為相似度分數
        if similarity_score >= 0.45:  # 只添加相似度大於等於0.45的結果
            results.append({
                'text': data[idx]['text'],
                'entities': data[idx]['entities'],
                'similarity_score': similarity_score,
                'entity_groups': get_entity_groups(data[idx]['entities'])
            })
    
    return results

def search_and_format(query):
    results = semantic_search(query)
    
    if not results:
        return "<div class='search-result'><p>查無相關資訊。</p></div>"
    
    formatted_results = """
    <style>
    .search-result {
        font-size: 24px;
        line-height: 1.6;
    }
    .search-result h2 {
        font-size: 24px;
        color: #333;
    }
    .search-result h3 {
        font-size: 24px;
        color: #444;
    }
    .search-result p {
        margin-bottom: 24px;
    }
    .result-separator {
        border-top: 2px solid #ccc;
        margin: 20px 0;
    }
    </style>
    <div class="search-result">
    """
    for i, result in enumerate(results, 1):
        if i > 1:
            formatted_results += '<div class="result-separator"></div>'
        formatted_results += f"<p><strong>相似度分數:</strong> {result['similarity_score']:.4f}</p>"
        formatted_results += f"<p><strong>情資:</strong> {format_text_with_entities_markdown(result['text'], result['entities'])}</p>"
        formatted_results += f"<p><strong>命名實體:</strong> {'、'.join(result['entity_groups'])}</p>"
    formatted_results += "</div>"
    return formatted_results

def format_text_with_entities_markdown(text, entities):
    # 將實體按照起始位置排序
    entity_spans = sorted(entities, key=lambda x: x['start'])
    
    # 創建一個字典來存儲每個單詞的實體
    word_entities = {}
    
    # 使用正則表達式分割文本為單詞
    words = re.findall(r'\S+|\s+', text)
    current_pos = 0
    
    for word in words:
        word_start = current_pos
        word_end = current_pos + len(word)
        word_entities[word] = []
        
        # 檢查每個實體是否與當前單詞重疊
        for entity in entity_spans:
            if entity['start'] < word_end and entity['end'] > word_start:
                word_entities[word].append(entity['entity_group'])
        
        current_pos = word_end
    
    # 處理每個單詞
    formatted_text = []
    for word in words:
        if word_entities[word]:
            unique_entity_groups = list(dict.fromkeys(word_entities[word]))  # 去除重複的實體
            entity_tags = []
            for i, group in enumerate(unique_entity_groups):
                entity_tag = f'<sup style="color: {get_color_for_entity(group)}; font-size: 14px;">{group}</sup>'
                if i > 0:  # 如果不是第一個標籤，添加逗號分隔符
                    entity_tags.append('<sup style="font-size: 14px;">、</sup>')
                entity_tags.append(entity_tag)
            formatted_word = f'<strong>{word}</strong>{"".join(entity_tags)}'
        else:
            formatted_word = word
        formatted_text.append(formatted_word)
    
    return ''.join(formatted_text)

def transcribe_audio(audio):
    try:
        with open(audio, "rb") as audio_file:
            transcript = openai.Audio.transcribe("whisper-1", audio_file)
        return transcript.text
    except Exception as e:
        return f"轉錄時發生錯誤: {str(e)}"

def audio_to_search(audio):
    transcription = transcribe_audio(audio)
    search_results = search_and_format(transcription)
    combined_output = f""
    return combined_output, transcription

# 範例問題
example_queries = [
    "Tell me about recent cyber attacks from Russia",
    "What APT groups are targeting Ukraine?",
    "Explain the Log4j vulnerability",
    "Chinese state-sponsored hacking activities",
    "Toilet?",
    "Latest North Korean hacker",
    "Describe the SolarWinds supply chain attack",
    "What is the Lazarus Group known for?",
    "Common attack vectors used against critical infrastructure",
    "pls rick roll me"
]

# 自定義 CSS
custom_css = """
.container {display: flex; flex-direction: row;}
.input-column {flex: 1; padding-right: 20px;}
.output-column {flex: 2;}
.examples-list {display: flex; flex-wrap: wrap; gap: 10px;}
.examples-list > * {flex-basis: calc(50% - 5px);}
footer {display:none !important}
.gradio-container {font-size: 16px;}
"""

# 創建Gradio界面
with gr.Blocks(css=custom_css) as iface:
    gr.Markdown("# AskCTI", elem_classes=["text-3xl"])
    gr.Markdown("使用文字或使用語音輸入問題或關鍵字查詢相關威脅情資，右側會顯示前 5 個最相關的情資報告。", elem_classes=["text-xl"])
    
    with gr.Row(equal_height=True):
        with gr.Column(scale=1, min_width=300):
            query_input = gr.Textbox(lines=3, label="", elem_classes=["text-lg"])
            with gr.Row():
                submit_btn = gr.Button("查詢", elem_classes=["text-lg"])
                audio_input = gr.Audio(type="filepath", label="語音輸入")
                #audio_input = gr.Audio(sources="microphone", label="錄音", elem_classes="small-button")
            
            gr.Markdown("### 範例查詢", elem_classes=["text-xl"])
            for i in range(0, len(example_queries), 2):
                with gr.Row():
                    for j in range(2):
                        if i + j < len(example_queries):
                            gr.Button(example_queries[i+j], elem_classes=["text-lg"]).click(
                                lambda x: x, inputs=[gr.Textbox(value=example_queries[i+j], visible=False)], outputs=[query_input]
                            )
        
        with gr.Column(scale=2):
            output = gr.HTML(elem_classes=["text-lg"])

    submit_btn.click(search_and_format, inputs=[query_input], outputs=[output])
    audio_input.change(audio_to_search, inputs=[audio_input], outputs=[output, query_input])

# 啟動Gradio界面
iface.launch()