twikoo-func 1.7.13 → 1.7.14

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "twikoo-func",
-  "version": "1.7.13",
+  "version": "1.7.14",
   "description": "A simple comment system.",
   "author": "imaegoo <hello@imaegoo.com> (https://github.com/imaegoo)",
   "license": "MIT",
@@ -25,6 +25,7 @@
     "jsdom": "^16.4.0",
     "marked": "^4.0.12",
     "nodemailer": "^7.0.11",
+    "openai": "^6.45.0",
     "pushoo": "latest",
     "tencentcloud-sdk-nodejs": "^4.0.65",
     "xml2js": "^0.6.0"

package/utils/image.js

@@ -242,13 +242,18 @@ const fn = {
     // 构建对象 key
     const prefix = config.S3_PATH_PREFIX ? config.S3_PATH_PREFIX.replace(/\/$/, '') + '/' : ''
     const key = `${prefix}${Date.now()}-${fileName}`
+    const forcePathStyle = String(config.S3_FORCE_PATH_STYLE).trim().toLowerCase() !== 'false'
     let endpoint
+    let s3Base
     if (config.S3_ENDPOINT) {
-      // 兼容 R2
-      endpoint = `${config.S3_ENDPOINT.replace(/\/$/, '')}/${config.S3_BUCKET}/${key}`
+      // 自定义 S3 Endpoint
+      const endpointBase = config.S3_ENDPOINT.replace(/\/$/, '')
+      s3Base = forcePathStyle ? `${endpointBase}/${config.S3_BUCKET}` : endpointBase
+      endpoint = `${s3Base}/${key}`
     } else {
       // 标准 AWS S3：virtual-hosted-style URL
-      endpoint = `https://${config.S3_BUCKET}.s3.${region}.amazonaws.com/${key}`
+      s3Base = `https://${config.S3_BUCKET}.s3.${region}.amazonaws.com`
+      endpoint = `${s3Base}/${key}`
     }
     const endpointUrl = new URL(endpoint)
     const host = endpointUrl.host
@@ -305,10 +310,8 @@ const fn = {
     let fileUrl
     if (config.S3_CDN_URL) {
       fileUrl = `${config.S3_CDN_URL.replace(/\/$/, '')}/${key}`
-    } else if (config.S3_ENDPOINT) {
-      fileUrl = `${config.S3_ENDPOINT.replace(/\/$/, '')}/${config.S3_BUCKET}/${key}`
     } else {
-      fileUrl = `https://${config.S3_BUCKET}.s3.${region}.amazonaws.com/${key}`
+      fileUrl = `${s3Base}/${key}`
     }
     res.data = { url: fileUrl }
   },

package/utils/lib.js

@@ -71,5 +71,13 @@ module.exports = {
   getXml2js () {
     const xml2js = require('xml2js') // XML 解析
     return xml2js
+  },
+  getOpenAIClient (config) {
+    const OpenAI = require('openai') // OpenAI 的 SDK，用于反垃圾
+    const openaiClient = new OpenAI({
+      apiKey: config.LLM_API_KEY,
+      baseURL: config.LLM_API_ENDPOINT || 'https://api.deepseek.com'
+    })
+    return openaiClient
   }
 }

package/utils/spam.js

@@ -1,7 +1,8 @@
 const {
   getAkismetClient,
   getCryptoJS,
-  getTencentcloud
+  getTencentcloud,
+  getOpenAIClient
 } = require('./lib')
 const {
   equalsMail
@@ -12,6 +13,9 @@ const CryptoJS = getCryptoJS()
 const logger = require('./logger')
 
 let tencentcloud
+let openai
+let _openaiApiKey
+let _openaiEndpoint
 
 function getTencentCloud () {
   if (!tencentcloud) {
@@ -24,6 +28,155 @@ function getTencentCloud () {
   return tencentcloud
 }
 
+function getOpenAI (config) {
+  if (isConfigChanged(config)) {
+    _openaiApiKey = config.LLM_API_KEY || ''
+    _openaiEndpoint = config.LLM_API_ENDPOINT || ''
+    openai = getOpenAIClient(config)
+  }
+  return openai
+}
+
+function isConfigChanged (config) {
+  return !openai ||
+    _openaiApiKey !== (config.LLM_API_KEY || '') ||
+    _openaiEndpoint !== (config.LLM_API_ENDPOINT || '')
+}
+
+// 提取json结构的函数
+function extractJson (rawText) {
+  if (!rawText) return ''
+  const trimmed = rawText.trim()
+  const match = trimmed.match(/\{[\s\S]*\}/)
+  return match ? match[0] : trimmed
+}
+
+// 移除多余的字符
+function repairJson (jsonStr) {
+  if (!jsonStr) return ''
+  let cleaned = jsonStr.trim()
+  cleaned = cleaned.replace(/,\s*([}\]])/g, '$1')
+  return cleaned
+}
+
+// validateJson 返回的 error 信息会被拼接进 prompt
+function validateJson (jsonStr) {
+  try {
+    const obj = JSON.parse(jsonStr)
+    if (typeof obj !== 'object' || obj === null) {
+      return { valid: false, error: 'Parsed JSON is not an object' }
+    }
+    if (!('spam' in obj)) {
+      return { valid: false, error: 'Missing required key "spam"' }
+    }
+    if (typeof obj.spam !== 'boolean') {
+      return { valid: false, error: 'Key "spam" must be a boolean value' }
+    }
+    return { valid: true, data: obj }
+  } catch (err) {
+    return { valid: false, error: `JSON Parse failed: ${err.message}` }
+  }
+}
+
+// 生成提示词的函数: system 为管理员指令，user 为待审核的评论数据
+// errorMsg 用于重试，customPrompt 替代默认的 system 指令
+function buildMessages (commentData, errorMsg = '', customPrompt = '') {
+  // 1. system 指令（管理员自定义或内置默认）
+  const systemContent = customPrompt || `You are a blog comment moderation assistant. Analyze ALL fields below and determine if this submission is spam or ham.
+
+Spam includes ANY of the following in ANY field:
+- Commercial ads, promotions, or buying/selling offers (e.g., "代开发票", "加微信", "兼职", "办证", "AI中转站").
+- Special case: If the comment text is harmless, but the nickname is suspicious (e.g., contains ads or promotions) AND a website link is provided, treat it as SPAM.
+- Meaningless gibberish or spammy repetition (e.g., "顶顶顶", "111111", "asdfgh", "好" repeated).
+- Abusive language, insults, or offensive Chinese slang.
+- Suspicious links or SEO spam in the website field.
+- Bot-like automated greetings.
+
+Ham includes:
+- Genuine questions, constructive feedback, technical discussions, or normal greetings in Chinese/English.
+
+Strictly follow these rules:
+1. If ANY field contains spam content, output exactly {"spam": true}.
+2. If ALL fields are legitimate, output exactly {"spam": false}.
+3. Do not include any explanations, introduction, punctuation, or extra spaces. Output only the JSON object.
+
+Your response MUST be a single valid JSON object, like:
+{"spam": true} or {"spam": false}`
+
+  // 2. user 数据（始终是待审核的评论内容）
+  let userContent = `Comment: ${commentData.comment}
+Nickname: ${commentData.nick || ''}
+Website: ${commentData.link || ''}`
+
+  // 3. 如果有错误信息，追加到用户消息末尾
+  if (errorMsg) {
+    userContent += `\n\n[ERROR FROM PREVIOUS ATTEMPT]: Your last response failed verification with error: "${errorMsg}". Please correct your output format and make sure to return exactly valid JSON.`
+  }
+
+  // 4. 返回 messages 数组
+  const finalPrompt = [
+    { role: 'system', content: systemContent },
+    { role: 'user', content: userContent }
+  ]
+  logger.log('提示词是：', finalPrompt)
+  return finalPrompt
+}
+
+async function checkByLLM (comment, config) {
+  const maxRetries = Number(config.LLM_MAX_RETRIES) || 3
+  let lastError = ''
+
+  const openai = getOpenAI(config)
+
+  // 网络/Provider 异常或者格式校验不通过会进入重试逻辑
+  for (let attempt = 1; attempt <= maxRetries; attempt++) {
+    if (attempt > 1) {
+      await new Promise(resolve => setTimeout(resolve, 1000))
+    }
+    try {
+      let messages
+      // 自定义提示词和内置提示词的 message 构建逻辑全交给 buildMessages 函数
+      if (config.LLM_SPAM_PROMPT) {
+        messages = buildMessages(comment, lastError, config.LLM_SPAM_PROMPT)
+      } else {
+        messages = buildMessages(comment, lastError)
+      }
+
+      const chatCompletion = await openai.chat.completions.create({
+        model: config.LLM_MODEL || 'deepseek-v4-pro',
+        response_format: { type: 'json_object' },
+        messages
+      })
+
+      const rawText = chatCompletion.choices[0].message.content || ''
+
+      const extracted = extractJson(rawText)
+      const repaired = repairJson(extracted)
+      const validation = validateJson(repaired)
+
+      // 校验通过返回，不通过继续循环
+      if (validation.valid) {
+        const isSpam = validation.data.spam
+        if (isSpam) {
+          logger.info(`LLM 判定为 SPAM (尝试第 ${attempt} 次): id="${comment.id}" nick="${comment.nick}"`)
+        } else {
+          logger.log(`LLM 判定为 HAM (尝试第 ${attempt} 次): id="${comment.id}" nick="${comment.nick}"`)
+        }
+        return isSpam
+      } else {
+        lastError = validation.error
+        logger.warn(`LLM 返回校验失败 (尝试第 ${attempt}/${maxRetries} 次), 错误: ${lastError}. 原始返回: "${rawText}"`)
+      }
+    } catch (error) {
+      lastError = error.message
+      logger.error(`LLM 请求异常 (尝试第 ${attempt}/${maxRetries} 次), 错误: ${lastError}`)
+    }
+  }
+
+  logger.error(`LLM 垃圾评论检测历经 ${maxRetries} 次尝试均失败，执行终极放行兜底（返回 false）`)
+  return false
+}
+
 const fn = {
   // 后垃圾评论检测
   async postCheckSpam (comment, config) {
@@ -77,6 +230,9 @@ const fn = {
           comment_author_url: comment.link,
           comment_content: comment.comment
         })
+      } else if (config.LLM_API_KEY) {
+        // 大语言模型检测
+        isSpam = await checkByLLM(comment, config)
       }
       logger.log('垃圾评论检测结果：', isSpam)
       return isSpam