memind/health-document-ocr.mjs

export function buildHealthDocumentOcrPrompt() {
  return [
    '你是医疗报告 OCR 助手。从体检单/化验单/影像报告照片中提取结构化信息。',
    '禁止诊断，只提取可见文字与数值。',
    '只输出严格 JSON，不要 Markdown：',
    '{"doc_type":"checkup|lab|imaging|prescription|discharge|other","report_date":"YYYY-MM-DD|null","institution":"","title":"","metrics":[{"name":"","value":"","unit":"","flag":""}],"summary":"","confidence":0.0}',
    '无法识别时 metrics 为空数组，summary 说明原因。',
  ].join('\n');
}

export function parseHealthDocumentOcrJson(raw) {
  const text = Array.isArray(raw)
    ? raw.map((item) => (typeof item === 'string' ? item : item?.text ?? '')).join('')
    : String(raw ?? '');
  const unfenced = text.trim().replace(/^```(?:json)?\s*/i, '').replace(/\s*```$/, '');
  const start = unfenced.indexOf('{');
  const end = unfenced.lastIndexOf('}');
  if (start < 0 || end <= start) return null;
  try {
    return JSON.parse(unfenced.slice(start, end + 1));
  } catch {
    return null;
  }
}

export function normalizeHealthDocumentOcr(parsed = {}) {
  const docType = String(parsed.doc_type ?? 'other').trim() || 'other';
  const metrics = Array.isArray(parsed.metrics)
    ? parsed.metrics
        .filter((item) => item && (item.name || item.value))
        .slice(0, 30)
        .map((item) => ({
          name: String(item.name ?? '').trim(),
          value: String(item.value ?? '').trim(),
          unit: String(item.unit ?? '').trim() || null,
          flag: String(item.flag ?? '').trim() || null,
        }))
    : [];
  const summary = String(parsed.summary ?? '').trim();
  const confidence = Number(parsed.confidence ?? 0);
  const hasSignal = metrics.length > 0 || summary.length > 20;
  return {
    ok: hasSignal,
    docType,
    reportDate: parsed.report_date ?? null,
    institution: String(parsed.institution ?? '').trim() || null,
    title: String(parsed.title ?? '').trim() || '健康报告',
    extractedMetrics: metrics,
    ocrText: summary,
    extractionStatus: metrics.length > 0 ? 'partial' : summary ? 'partial' : 'failed',
    confidence: Number.isFinite(confidence) ? confidence : null,
  };
}

export async function extractHealthDocumentFromVision({
  buffer,
  mimeType = 'image/jpeg',
  analyzeImagesWithVision,
  buildVisionThumbnailBuffer = null,
} = {}) {
  if (!analyzeImagesWithVision || !Buffer.isBuffer(buffer) || !buffer.length) {
    return { ok: false, error: 'vision_unavailable', message: '报告 OCR 暂不可用' };
  }

  let visionBuffer = buffer;
  let visionMimeType = mimeType;
  if (buildVisionThumbnailBuffer) {
    try {
      visionBuffer = await buildVisionThumbnailBuffer(buffer, mimeType);
      visionMimeType = 'image/jpeg';
    } catch {
      visionBuffer = buffer;
      visionMimeType = mimeType;
    }
  }

  let raw;
  try {
    raw = await analyzeImagesWithVision(
      [{
        mimeType,
        visionMimeType,
        data: visionBuffer.toString('base64'),
      }],
      buildHealthDocumentOcrPrompt(),
    );
  } catch (error) {
    return {
      ok: false,
      error: 'vision_failed',
      message: error instanceof Error ? error.message : '报告 OCR 失败',
    };
  }

  const parsed = parseHealthDocumentOcrJson(raw);
  if (!parsed) {
    return { ok: false, error: 'invalid_response', message: '未能解析报告内容' };
  }
  const normalized = normalizeHealthDocumentOcr(parsed);
  if (!normalized.ok) {
    return {
      ok: false,
      error: 'no_content',
      message: '未能从报告中提取有效文字，原图已归档，关键数值请手输确认。',
    };
  }
  return { ok: true, ...normalized };
}

export async function extractHealthDocumentFromUrl(options = {}) {
  const { fetchHealthImageBuffer } = await import('./health-image-extract.mjs');
  const { buffer, mimeType } = await fetchHealthImageBuffer(options);
  return extractHealthDocumentFromVision({
    buffer,
    mimeType,
    analyzeImagesWithVision: options.analyzeImagesWithVision,
    buildVisionThumbnailBuffer: options.buildVisionThumbnailBuffer,
  });
}