hulai-website/scripts/import-review-screenshots.mjs

#!/usr/bin/env node
/**
 * 批量 OCR 评价截图 → 匹配 DB 评价 → 写入 screenshot 字段
 *
 * 流程：
 *  1. 扫描 public/uploads/reviews/ 的 jpg
 *  2. tesseract chi_sim OCR
 *  3. 过滤：包含手机号（1[3-9]\d{9}）或负面关键词的截图丢弃
 *  4. 规整文本，按 n-gram 交集相似度匹配 82 条 DB 评价
 *  5. 命中 → UPDATE reviews SET screenshot=/uploads/reviews/xxx.jpg
 *  6. 未命中 → 报告供人工处理
 *
 * 不自动创建新评价：避免 OCR 噪声/差评混入库。
 */
import Database from 'better-sqlite3'
import { execFileSync } from 'node:child_process'
import { readdirSync, existsSync, writeFileSync } from 'node:fs'
import { join, dirname } from 'node:path'
import { fileURLToPath } from 'node:url'

const __dirname = dirname(fileURLToPath(import.meta.url))
const ROOT = join(__dirname, '..')
const DB_PATH = join(ROOT, 'server/database/hulai.db')
const SHOTS_DIR = join(ROOT, 'public/uploads/reviews')
const REL_PREFIX = '/uploads/reviews/'

const PHONE_RE = /\b1[3-9]\d{9}\b/
const NEG_KEYWORDS = ['差评', '投诉', '退款', '退一赔', '不满意', '很失望', '骗人', '被骗', '黑心', '垃圾', '坑人', '太烂', '烂透', '糟糕', '欺诈', '维权', '曝光']

function ocr(imagePath) {
  try {
    const out = execFileSync('tesseract', [imagePath, 'stdout', '-l', 'chi_sim'], {
      encoding: 'utf8',
      stdio: ['ignore', 'pipe', 'ignore'],
      maxBuffer: 10 * 1024 * 1024,
    })
    return out
  } catch {
    return ''
  }
}

// 清理 OCR 文本：去空白/标点/数字/半角字母，只留中文
function canonical(s) {
  return (s || '').replace(/[^\u4e00-\u9fa5]/g, '')
}

// 生成 bigram 集合
function bigrams(s) {
  const set = new Set()
  for (let i = 0; i < s.length - 1; i++) set.add(s.slice(i, i + 2))
  return set
}

// bigram Jaccard 相似度（0~1）
function similarity(a, b) {
  if (!a || !b) return 0
  const A = bigrams(a), B = bigrams(b)
  if (A.size === 0 || B.size === 0) return 0
  let inter = 0
  for (const x of A) if (B.has(x)) inter++
  return inter / Math.min(A.size, B.size)  // use min instead of union: 截图文本往往比润色后的 DB 内容多噪声
}

// 最长公共子串长度（粗粒度，用于 tie-break）
function lcsLen(a, b) {
  // 限制长度避免 O(n²) 爆炸
  if (a.length > 500) a = a.slice(0, 500)
  if (b.length > 500) b = b.slice(0, 500)
  const n = a.length, m = b.length
  if (!n || !m) return 0
  let prev = new Array(m + 1).fill(0)
  let curr = new Array(m + 1).fill(0)
  let best = 0
  for (let i = 1; i <= n; i++) {
    for (let j = 1; j <= m; j++) {
      curr[j] = a[i - 1] === b[j - 1] ? prev[j - 1] + 1 : 0
      if (curr[j] > best) best = curr[j]
    }
    ;[prev, curr] = [curr, prev]
  }
  return best
}

const db = new Database(DB_PATH)
const reviews = db.prepare('SELECT id, author, content, screenshot FROM reviews').all()
console.log(`DB reviews: ${reviews.length}`)

const reviewsIdx = reviews.map(r => ({ ...r, canon: canonical(r.content) }))

const files = readdirSync(SHOTS_DIR).filter(f => /\.(jpe?g|png)$/i.test(f))
console.log(`screenshots: ${files.length}\n`)

const SIM_THRESHOLD = 0.35
const LCS_MIN = 10

const results = { matched: [], filteredPhone: [], filteredNeg: [], unmatched: [], ocrFailed: [] }
const assignedReviewIds = new Set()

let i = 0
for (const file of files) {
  i++
  const abs = join(SHOTS_DIR, file)
  process.stdout.write(`[${i}/${files.length}] ${file} ... `)
  const text = ocr(abs)
  if (!text || text.length < 20) {
    console.log('OCR 失败/太短')
    results.ocrFailed.push(file)
    continue
  }
  if (PHONE_RE.test(text)) {
    console.log('含手机号 → 跳过')
    results.filteredPhone.push({ file, snippet: text.match(PHONE_RE)[0] })
    continue
  }
  const hitNeg = NEG_KEYWORDS.find(k => text.includes(k))
  if (hitNeg) {
    console.log(`负面关键词「${hitNeg}」→ 跳过`)
    results.filteredNeg.push({ file, keyword: hitNeg })
    continue
  }
  const canon = canonical(text)
  if (canon.length < 30) {
    console.log('有效中文太少')
    results.ocrFailed.push(file)
    continue
  }

  // 评分
  let best = null
  for (const r of reviewsIdx) {
    const sim = similarity(canon, r.canon)
    if (!best || sim > best.sim) best = { ...r, sim }
  }
  if (best && best.sim >= SIM_THRESHOLD) {
    const lcs = lcsLen(canon, best.canon)
    if (lcs >= LCS_MIN) {
      if (assignedReviewIds.has(best.id)) {
        // 已分配给更高分的图
        console.log(`评价#${best.id} 已占，放入未匹配`)
        results.unmatched.push({ file, bestId: best.id, sim: best.sim.toFixed(3), reason: 'dup' })
        continue
      }
      assignedReviewIds.add(best.id)
      results.matched.push({ file, id: best.id, author: best.author, sim: best.sim.toFixed(3), lcs })
      console.log(`→ #${best.id} ${best.author}  sim=${best.sim.toFixed(3)} lcs=${lcs}`)
      continue
    }
  }
  console.log(`未匹配 (best sim=${best?.sim.toFixed(3)})`)
  results.unmatched.push({ file, bestId: best?.id, sim: best?.sim.toFixed(3) })
}

// 写入 DB
const update = db.prepare('UPDATE reviews SET screenshot = ? WHERE id = ?')
const tx = db.transaction(rows => rows.forEach(r => update.run(REL_PREFIX + r.file, r.id)))
tx(results.matched)

// 过滤掉手机/负面的截图文件直接删除
import('node:fs').then(({ unlinkSync }) => {
  for (const { file } of [...results.filteredPhone, ...results.filteredNeg]) {
    try { unlinkSync(join(SHOTS_DIR, file)) } catch {}
  }
})

writeFileSync(join(ROOT, 'scripts/.screenshot-import-report.json'), JSON.stringify(results, null, 2))

console.log('\n===== 汇总 =====')
console.log(`命中 DB 并写入:       ${results.matched.length}`)
console.log(`含手机号已跳过删除:   ${results.filteredPhone.length}`)
console.log(`负面关键词已跳过删除: ${results.filteredNeg.length}`)
console.log(`未匹配:               ${results.unmatched.length}`)
console.log(`OCR 失败:             ${results.ocrFailed.length}`)
console.log(`\n详细报告: scripts/.screenshot-import-report.json`)