识别只返回文本,不投递、不落库:语音支持 mp3/aac/wav/pcm,WebM 先按文件头识别再转 PCM;图片走 qwen-vl-ocr。新增 /health/ai/asr 与 /health/ai/ocr,登录后按用户限流 20 次/分钟,并对 /health/ai/ 前缀放宽请求体上限。
576 lines
22 KiB
JavaScript
576 lines
22 KiB
JavaScript
/**
|
||
* 图像识别适配器:通义千问(DashScope)视觉模型
|
||
*
|
||
* 与 asr.js 的会话接口一致:{ push(buffer), finish() -> Promise<text>, abort() }
|
||
* 图片不做流式推流,分片只用于搬运大文件,`finish()` 时整图一次性提交。
|
||
*
|
||
* 协议(DashScope OpenAI 兼容模式):
|
||
* - POST {baseUrl}/chat/completions
|
||
* - messages 里用 image_url 传 data URL:`data:${mime};base64,...`
|
||
* - 默认 qwen-vl-ocr + transcribe(只转写,解读交给主 agent)
|
||
* - 要求模型输出严格 JSON,这里拼成可读文本再返回
|
||
*
|
||
* 密钥只从配置或环境变量读取,不回显给调用方;原图只在本轮内存里存在,用完即释放。
|
||
*/
|
||
import { randomUUID } from 'node:crypto';
|
||
import logger from '../../../utils/logger.js';
|
||
|
||
/** DashScope OpenAI 兼容入口 */
|
||
export const DASHSCOPE_CHAT_ENDPOINT = 'https://dashscope.aliyuncs.com/compatible-mode/v1/chat/completions';
|
||
|
||
/** 缺省模型:专用 OCR 模型,只转写时速度约为通用视觉模型的 3 倍 */
|
||
export const DEFAULT_VISION_MODEL = 'qwen-vl-ocr';
|
||
|
||
/** 缺省免责声明:随结果返回,由配置覆盖 */
|
||
export const DEFAULT_DISCLAIMER = '本内容为图像文字识别结果,不构成诊断意见,请以医生诊断与医院报告为准。';
|
||
|
||
/** 图片类型取值:首版固定 auto,由模型自判 */
|
||
export const IMAGE_KINDS = ['auto', 'report', 'ecg', 'imaging'];
|
||
|
||
/** 系统提示:只依据图片可见信息作答,输出严格 JSON,解读从简 */
|
||
const SYSTEM_PROMPT = `你是医学图像与医疗文书识别助手。请完整转写图片内容,并给出简短、客观的解读。
|
||
|
||
要求:
|
||
1. 只依据图片中实际可见的信息作答;看不清或不确定写"无法辨认",禁止编造数值、人名、日期与结论。
|
||
2. 数值、单位、参考范围、箭头(↑ ↓)、阴性/阳性标记原样保留,不得四舍五入或换算。
|
||
3. 只输出一个 JSON 对象,不要用 Markdown 代码块包裹,不要输出 JSON 以外的说明。
|
||
4. 解读从简:findings ≤80 字,impression ≤120 字;不要展开病因清单、检查建议或长篇科普。
|
||
5. 发现可能危及生命的征象时写入 urgent 并建议尽快就医;没有则留空字符串。`;
|
||
|
||
/** 按图片类型给出的额外指令:决定模型关注点(首版只有 auto 会命中) */
|
||
const KIND_PROMPTS = {
|
||
report: '这是一份检查/检验报告或病历。逐行转写全部文字,表格列用 | 分隔;facts 只列偏高/偏低/阳性等异常项,正常项不重复。',
|
||
ecg: '这是一张心电图。读出心率、心律、各导联可见的 P/QRS/T 形态、ST 段与 QT 间期等可辨识信息,注明导联数与走纸速度;先客观描述再简短解读。',
|
||
imaging: '这是一张影像图(X 光/CT/MRI 胶片或屏幕照片)。说明部位与体位/投照方式,客观描述可见结构(肺野、心影、纵隔、膈肌、骨骼、有无置管),再给出简短解读;屏幕拍摄的摩尔纹或模糊写进 quality。',
|
||
auto: '先判断图片属于哪一类(检验/检查报告、病历、心电图、影像、处方),再按对应类型完成转写与简短解读。'
|
||
};
|
||
|
||
/** 输出 JSON 的字段契约(限制解读长度、facts 去重,减少生成 token) */
|
||
const SCHEMA_PROMPT = `请只输出以下 JSON 结构(没有的字段留空字符串或空数组):
|
||
{
|
||
"docType": "lab_report | medical_record | ecg | imaging | prescription | other",
|
||
"title": "文档或检查名称",
|
||
"text": "逐行转写的正文,表格用 | 分隔列",
|
||
"facts": [{ "name": "", "value": "", "unit": "", "ref": "", "flag": "偏高|偏低|阳性" }],
|
||
"findings": "客观所见,≤80 字",
|
||
"impression": "通俗解读,≤120 字",
|
||
"urgent": "需要尽快就医的征象,没有则为空",
|
||
"quality": { "readable": true, "issues": [] }
|
||
}
|
||
facts 仅列异常/阳性项,正常项已在 text 中,不要重复;不要输出 disclaimer。`;
|
||
|
||
/**
|
||
* 只转写模式不附加任何提示词。
|
||
*
|
||
* qwen-vl-ocr 是专用 OCR 模型,自带逐行转写模板;实测给它系统提示或 JSON 契约后,
|
||
* 它会把整段文字压成摘要(表格列还会粘成一串),甚至只回坐标不回文字。
|
||
* 只传图片时输出才是逐行完整原文,类型与标题在本文件里按文本兜底判断。
|
||
*/
|
||
|
||
/**
|
||
* 构造一次识别请求的消息体
|
||
* @param {Object} params
|
||
* @param {string} [params.kind] - auto / report / ecg / imaging
|
||
* @param {string} [params.mime] - 图片 MIME
|
||
* @param {string} [params.base64] - 图片 base64(不含前缀)
|
||
* @param {number} [params.width] - 像素宽(供模型判断清晰度)
|
||
* @param {number} [params.height] - 像素高
|
||
* @param {string} [params.profile] - full(转写+解读)/ transcribe(只转写)
|
||
* @returns {Array<Object>} OpenAI 兼容的 messages
|
||
*/
|
||
export function buildMessages({
|
||
kind = 'auto',
|
||
mime = 'image/jpeg',
|
||
base64 = '',
|
||
width = 0,
|
||
height = 0,
|
||
profile = 'transcribe'
|
||
} = {}) {
|
||
const imagePart = { type: 'image_url', image_url: { url: `data:${mime};base64,${base64}` } };
|
||
|
||
// 只转写:不挂系统提示、不加结构化契约,交给模型按原生 OCR 模板逐行输出
|
||
if (profile === 'transcribe') {
|
||
return [
|
||
{
|
||
role: 'user',
|
||
content: [imagePart]
|
||
}
|
||
];
|
||
}
|
||
|
||
const sizeNote = width && height ? `图片尺寸约 ${width}×${height} 像素。` : '';
|
||
const kindPrompt = KIND_PROMPTS[kind] || KIND_PROMPTS.auto;
|
||
const instruction = `${kindPrompt}\n${sizeNote}\n${SCHEMA_PROMPT}`;
|
||
|
||
return [
|
||
{ role: 'system', content: SYSTEM_PROMPT },
|
||
{
|
||
role: 'user',
|
||
content: [imagePart, { type: 'text', text: instruction }]
|
||
}
|
||
];
|
||
}
|
||
|
||
function tryParseJson(text) {
|
||
try {
|
||
return JSON.parse(text);
|
||
} catch {
|
||
return undefined;
|
||
}
|
||
}
|
||
|
||
/**
|
||
* 从模型输出里取出 JSON 主体:模型可能加 Markdown 代码块或前后说明。
|
||
* 依次尝试「代码块内容 / 整段 / 花括号区间 / 方括号区间」,都解析不了返回 undefined。
|
||
* 纯文本转写一律走 undefined 分支,避免把正文里的方括号当成数组截断。
|
||
*/
|
||
function parseStructured(raw) {
|
||
const text = (raw || '').trim();
|
||
if (!text) return undefined;
|
||
|
||
const candidates = [];
|
||
const fenced = text.match(/```(?:json)?\s*([\s\S]*?)```/i);
|
||
if (fenced) candidates.push(fenced[1].trim());
|
||
candidates.push(text);
|
||
|
||
const objStart = text.indexOf('{');
|
||
if (objStart !== -1) {
|
||
const end = text.lastIndexOf('}');
|
||
if (end > objStart) candidates.push(text.slice(objStart, end + 1));
|
||
}
|
||
const arrStart = text.indexOf('[');
|
||
if (arrStart !== -1) {
|
||
const end = text.lastIndexOf(']');
|
||
if (end > arrStart) candidates.push(text.slice(arrStart, end + 1));
|
||
}
|
||
|
||
for (const candidate of candidates) {
|
||
const parsed = tryParseJson(candidate);
|
||
if (parsed !== undefined) return parsed;
|
||
}
|
||
return undefined;
|
||
}
|
||
|
||
/**
|
||
* 原生 OCR 行结果([{ rotate_rect, text }])拼成逐行正文。
|
||
* 有的模型版本会逐行返回坐标 + 文字,这里只取文字并按给出的行序排列。
|
||
* @param {Array} items
|
||
* @returns {string} 空串表示该结果里没有可用文字
|
||
*/
|
||
function joinOcrLines(items) {
|
||
if (!Array.isArray(items)) return '';
|
||
return items
|
||
.map((item) => {
|
||
if (typeof item === 'string') return item.trim();
|
||
return typeof item?.text === 'string' ? item.text.trim() : '';
|
||
})
|
||
.filter(Boolean)
|
||
.join('\n');
|
||
}
|
||
|
||
/**
|
||
* 解析模型输出:兼容结构化 JSON、原生 OCR 行数组与纯文本转写。
|
||
* 三种形态最终都会拼成带【图像识别】【原文转写】的可读文本,
|
||
* 保证「原文」段就是模型看到的全部文字,而不是摘要。
|
||
* @param {string} raw - 模型返回的 content
|
||
* @param {string} [disclaimer] - 免责声明(JSON 未提供时兜底)
|
||
* @returns {{ data: Object|null, text: string }}
|
||
*/
|
||
export function parseVisionResult(raw, disclaimer = DEFAULT_DISCLAIMER) {
|
||
const rawText = (raw || '').trim();
|
||
if (!rawText) return { data: null, text: '' };
|
||
|
||
let data = null;
|
||
const parsed = parseStructured(rawText);
|
||
if (Array.isArray(parsed)) {
|
||
const text = joinOcrLines(parsed);
|
||
// 行数组里只有坐标没有文字:属于模型输出异常,不能当正文展示
|
||
if (!text) return { data: null, text: '' };
|
||
data = { text };
|
||
} else if (parsed && typeof parsed === 'object') {
|
||
data = parsed;
|
||
} else if (typeof parsed === 'string') {
|
||
data = { text: parsed.trim() };
|
||
}
|
||
|
||
// 纯文本输出(专用 OCR 模型最常见的形态)同样按原文转写落盘
|
||
if (!data) data = { text: rawText };
|
||
|
||
return { data, text: composeDocumentText(data, disclaimer) };
|
||
}
|
||
|
||
/** 文档类型关键词:标题命中优先,正文兜底(\bCT\b 避免把 PCT、CRP 当成 CT 检查) */
|
||
const DOC_TYPE_RULES = [
|
||
['ecg', /心电图|ECG|EKG/i],
|
||
['prescription', /处方|Rp\./i],
|
||
['imaging', /\bCT\b|\bMRI\b|核磁|磁共振|X\s*线|X\s*光|超声|B\s*超|彩超|影像/],
|
||
['lab_report', /检验|化验|血常规|生化|尿常规/]
|
||
];
|
||
|
||
/** 专用 OCR 模型常把 CT 报告判成 lab_report,这里用标题/正文关键词兜底纠正 */
|
||
function refineDocType(data) {
|
||
const title = String(data.title || pickTitle(data.text) || '');
|
||
const sample = `${title}\n${data.text || ''}`.slice(0, 1200);
|
||
for (const [type, pattern] of DOC_TYPE_RULES) {
|
||
if (pattern.test(title)) return type;
|
||
}
|
||
for (const [type, pattern] of DOC_TYPE_RULES) {
|
||
if (pattern.test(sample)) return type;
|
||
}
|
||
return data.docType;
|
||
}
|
||
|
||
/** 原生转写没有单独标题字段时,从首行取出文档名(只在像标题时才用) */
|
||
function pickTitle(text) {
|
||
const first = String(text || '')
|
||
.split('\n')
|
||
.map(line => line.trim())
|
||
.find(Boolean) || '';
|
||
if (!first || first.length > 40) return '';
|
||
return /报告|病历|处方|检验|化验|检查|影像|超声|心电|CT|MRI|X\s*光|门诊|入院|出院/i.test(first)
|
||
? first
|
||
: '';
|
||
}
|
||
|
||
/** 原生转写列之间可能是制表符、多个半角空格或全角空格;markdown 会折叠空白,需要还原成表格 */
|
||
const COLUMN_GAP_RE = /\t+|[ ]{2,}|[\u3000\u00a0\u2002\u2003\u2009]+/;
|
||
const TABLE_HEADER_RE = /项目|检验项目|检查项目|名称|指标|结果|数值|单位|参考范围|参考值|正常值|提示|状态|结论|日期|时间|药物|剂量|用法|频次/;
|
||
|
||
function splitColumns(line) {
|
||
return line
|
||
.trim()
|
||
.split(COLUMN_GAP_RE)
|
||
.map(cell => cell.trim())
|
||
.filter(Boolean);
|
||
}
|
||
|
||
/** 表头至少 3 列,且多数列名能命中报告表头关键词,避免把「姓名 性别 年龄」当成表格 */
|
||
function looksLikeTableHeader(cells) {
|
||
if (cells.length < 3) return false;
|
||
return cells.filter(cell => TABLE_HEADER_RE.test(cell)).length >= 2;
|
||
}
|
||
|
||
function escapeTableCell(value) {
|
||
return String(value)
|
||
.replace(/\\/g, '\\\\')
|
||
.replace(/\|/g, '\\|')
|
||
.replace(/\*/g, '\\*')
|
||
.replace(/_/g, '\\_')
|
||
.replace(/`/g, '\\`')
|
||
.replace(/\[/g, '\\[')
|
||
.replace(/\]/g, '\\]');
|
||
}
|
||
|
||
function toMarkdownTable(header, rows) {
|
||
const width = header.length;
|
||
const normalize = (cells) => {
|
||
const list = cells.slice(0, width).map(cell => cell.trim());
|
||
// 某行列数超出表头时并进最后一列,避免丢文字
|
||
if (cells.length > width) {
|
||
list[width - 1] = `${list[width - 1] || ''} ${cells.slice(width - 1).join(' ')}`.trim();
|
||
}
|
||
while (list.length < width) list.push('');
|
||
return list.map(escapeTableCell);
|
||
};
|
||
const lines = [
|
||
`| ${normalize(header).join(' | ')} |`,
|
||
`| ${Array.from({ length: width }, () => '---').join(' | ')} |`
|
||
];
|
||
for (const row of rows) lines.push(`| ${normalize(row).join(' | ')} |`);
|
||
return lines.join('\n');
|
||
}
|
||
|
||
/**
|
||
* 把原生 OCR 正文排版成 markdown,尽量贴近原图布局:
|
||
* - 连续的空格分列块还原成 markdown 表格(手机端表格会按列换行,不会像纯文本那样被折叠);
|
||
* - 其余行各自成段,保留原有换行顺序,不被 markdown 合并成一段。
|
||
* @param {string} text
|
||
* @returns {string}
|
||
*/
|
||
export function formatTranscription(text) {
|
||
const lines = String(text || '').replace(/\r\n/g, '\n').split('\n');
|
||
const blocks = [];
|
||
const paragraph = [];
|
||
|
||
const flushParagraph = () => {
|
||
if (!paragraph.length) return;
|
||
blocks.push(paragraph.join('\n\n'));
|
||
paragraph.length = 0;
|
||
};
|
||
|
||
for (let i = 0; i < lines.length; i++) {
|
||
const line = lines[i].trimEnd();
|
||
if (!line.trim()) {
|
||
flushParagraph();
|
||
continue;
|
||
}
|
||
|
||
const cells = splitColumns(line);
|
||
if (looksLikeTableHeader(cells)) {
|
||
const rows = [];
|
||
let j = i + 1;
|
||
while (j < lines.length) {
|
||
const next = lines[j].trimEnd();
|
||
if (!next.trim()) break;
|
||
const nextCells = splitColumns(next);
|
||
// 少于 3 列是普通说明行;又出现表头则是下一张表
|
||
if (nextCells.length < 3 || looksLikeTableHeader(nextCells)) break;
|
||
rows.push(nextCells);
|
||
j += 1;
|
||
}
|
||
if (rows.length) {
|
||
flushParagraph();
|
||
blocks.push(toMarkdownTable(cells, rows));
|
||
i = j - 1;
|
||
continue;
|
||
}
|
||
}
|
||
|
||
paragraph.push(line.trim());
|
||
}
|
||
|
||
flushParagraph();
|
||
return blocks.join('\n\n');
|
||
}
|
||
|
||
/** 把结构化结果拼成 agent 易读的纯文本(输入框回填 / 投递的人类消息就是这段) */
|
||
export function composeDocumentText(data, disclaimer = DEFAULT_DISCLAIMER) {
|
||
const typeMap = {
|
||
lab_report: '检验报告',
|
||
medical_record: '病历',
|
||
ecg: '心电图',
|
||
imaging: '影像',
|
||
prescription: '处方',
|
||
other: '其他'
|
||
};
|
||
const refinedType = refineDocType(data);
|
||
const docType = typeMap[refinedType] || refinedType || '未识别类型';
|
||
const title = data.title || pickTitle(data.text);
|
||
const lines = [`【图像识别】${docType}${title ? `|${title}` : ''}`];
|
||
|
||
// 原文转写用空行和标题分隔:markdown 会把单换行合并成一段,标题会和正文粘在一起
|
||
if (data.text) lines.push(`【原文转写】\n\n${formatTranscription(data.text)}`);
|
||
|
||
const facts = Array.isArray(data.facts) ? data.facts.filter(item => item && item.name) : [];
|
||
if (facts.length) {
|
||
const rows = facts.map((item) => {
|
||
const value = [item.value, item.unit].filter(Boolean).join(' ');
|
||
const ref = item.ref ? `(参考 ${item.ref})` : '';
|
||
const flag = item.flag && item.flag !== '未知' ? item.flag : '';
|
||
return `- ${item.name}:${value || '—'}${ref}${flag}`;
|
||
});
|
||
lines.push(`【结构化要点】\n${rows.join('\n')}`);
|
||
}
|
||
|
||
if (data.findings) lines.push(`【客观所见】\n${data.findings}`);
|
||
if (data.impression) lines.push(`【解读】\n${data.impression}`);
|
||
if (data.urgent) lines.push(`【需尽快就医】\n${data.urgent}`);
|
||
|
||
const quality = data.quality || {};
|
||
if (Array.isArray(quality.issues) && quality.issues.length) {
|
||
lines.push(`【图像质量】${quality.issues.join('、')}${quality.readable === false ? '(识别可能不完整)' : ''}`);
|
||
}
|
||
|
||
lines.push(`【说明】${data.disclaimer || disclaimer}`);
|
||
return lines.join('\n\n').trim();
|
||
}
|
||
|
||
/**
|
||
* 创建一轮图像识别会话
|
||
*
|
||
* 与 ASR 会话同签名,调用方无需区分模态:
|
||
* push(buffer) 只累积(图片不流式),finish() 提交整图,abort() 丢弃缓冲。
|
||
*
|
||
* @param {Object} [options]
|
||
* @param {string} [options.endpoint] - chat/completions 地址
|
||
* @param {string} [options.baseUrl] - OpenAI 兼容 baseUrl(自动补 /chat/completions)
|
||
* @param {string} [options.apiKey] - 显式密钥(优先于环境变量)
|
||
* @param {string} [options.apiKeyEnv] - 密钥环境变量名
|
||
* @param {string} [options.model] - 模型名(默认 qwen-vl-ocr)
|
||
* @param {string} [options.kind] - auto / report / ecg / imaging
|
||
* @param {string} [options.profile] - full(转写+解读)/ transcribe(只转写,默认)
|
||
* @param {string} [options.mime] - 图片 MIME
|
||
* @param {number} [options.width] - 像素宽
|
||
* @param {number} [options.height] - 像素高
|
||
* @param {string} [options.turnId] - 本轮标识(仅日志关联用)
|
||
* @param {Object} [options.parameters] - 追加到请求体的供应商参数
|
||
* @param {string} [options.disclaimer] - 免责声明(拼装时兜底)
|
||
* @param {number} [options.requestTimeoutMs] - 单次请求超时
|
||
* @param {Object} [options.logger] - 日志实例
|
||
* @param {Function} [options.fetchImpl] - 测试注入的 fetch
|
||
* @returns {{ push: Function, finish: Function, abort: Function }}
|
||
*/
|
||
export function createQwenVisionSession({
|
||
endpoint = DASHSCOPE_CHAT_ENDPOINT,
|
||
baseUrl = '',
|
||
apiKey = '',
|
||
apiKeyEnv = '',
|
||
model = DEFAULT_VISION_MODEL,
|
||
kind = 'auto',
|
||
profile = 'transcribe',
|
||
mime = 'image/jpeg',
|
||
width = 0,
|
||
height = 0,
|
||
turnId = '',
|
||
parameters = {},
|
||
disclaimer = DEFAULT_DISCLAIMER,
|
||
requestTimeoutMs = 60000,
|
||
logger: customLogger = null,
|
||
fetchImpl = fetch
|
||
} = {}) {
|
||
const _logger = customLogger || logger;
|
||
const key = apiKey || (apiKeyEnv ? process.env[apiKeyEnv] : '') || '';
|
||
const pickedModel = model || DEFAULT_VISION_MODEL;
|
||
const target = !endpoint && baseUrl
|
||
? `${baseUrl.replace(/\/$/, '')}/chat/completions`
|
||
: endpoint;
|
||
const requestId = randomUUID().replace(/-/g, '').slice(0, 12);
|
||
|
||
const chunks = [];
|
||
let aborted = false;
|
||
let finished = false;
|
||
/** 进行中的请求控制器:abort() 时一并中止,避免轮次已回收还在计费 */
|
||
let activeController = null;
|
||
|
||
return {
|
||
/**
|
||
* 推入一个图片分片(原始字节)
|
||
* @param {Buffer} buffer - 分片
|
||
*/
|
||
push(buffer) {
|
||
if (aborted || finished || !buffer?.length) return;
|
||
chunks.push(buffer);
|
||
},
|
||
|
||
/**
|
||
* 结束收集并提交识别
|
||
* @returns {Promise<string>} 可投递文本(结构化结果已拼成可读文本)
|
||
*/
|
||
async finish() {
|
||
if (aborted) return '';
|
||
finished = true;
|
||
|
||
// 耗时拆解:encoding(base64+消息体)、request(等模型返回)、parse(读响应)
|
||
const startedAt = Date.now();
|
||
const image = Buffer.concat(chunks);
|
||
const imageBytes = image.length;
|
||
chunks.length = 0;
|
||
|
||
if (imageBytes === 0) return '';
|
||
if (!key) throw new Error('ocr api key missing');
|
||
if (!/^https?:\/\//i.test(target || '')) throw new Error(`invalid ocr endpoint: ${target}`);
|
||
|
||
const encodeStartedAt = Date.now();
|
||
const body = {
|
||
model: pickedModel,
|
||
temperature: 0,
|
||
messages: buildMessages({ kind, profile, mime, base64: image.toString('base64'), width, height }),
|
||
...parameters
|
||
};
|
||
const encodeMs = Date.now() - encodeStartedAt;
|
||
|
||
const controller = new AbortController();
|
||
activeController = controller;
|
||
const timer = setTimeout(() => controller.abort(), requestTimeoutMs);
|
||
let response = null;
|
||
const requestStartedAt = Date.now();
|
||
try {
|
||
response = await fetchImpl(target, {
|
||
method: 'POST',
|
||
headers: {
|
||
'Content-Type': 'application/json',
|
||
Authorization: `Bearer ${key}`
|
||
},
|
||
body: JSON.stringify(body),
|
||
signal: controller.signal
|
||
});
|
||
} catch (err) {
|
||
_logger.warn('ocr request failed', {
|
||
turnId,
|
||
requestId,
|
||
model: pickedModel,
|
||
kind,
|
||
bytes: imageBytes,
|
||
encodeMs,
|
||
requestMs: Date.now() - requestStartedAt,
|
||
elapsedMs: Date.now() - startedAt,
|
||
error: err.message
|
||
});
|
||
throw err.name === 'AbortError'
|
||
? new Error(aborted ? 'ocr aborted' : 'ocr request timeout')
|
||
: err;
|
||
} finally {
|
||
clearTimeout(timer);
|
||
activeController = null;
|
||
}
|
||
|
||
if (!response.ok) {
|
||
let detail = '';
|
||
try {
|
||
detail = (await response.text()).slice(0, 200);
|
||
} catch {
|
||
// 读错误体失败不影响主流程
|
||
}
|
||
_logger.warn('ocr http error', {
|
||
turnId,
|
||
requestId,
|
||
status: response.status,
|
||
detail,
|
||
bytes: imageBytes,
|
||
encodeMs,
|
||
requestMs: Date.now() - requestStartedAt,
|
||
elapsedMs: Date.now() - startedAt
|
||
});
|
||
throw new Error(`ocr http ${response.status}${detail ? `: ${detail}` : ''}`);
|
||
}
|
||
|
||
const parseStartedAt = Date.now();
|
||
const payload = await response.json();
|
||
const content = payload?.choices?.[0]?.message?.content ?? '';
|
||
const raw = Array.isArray(content)
|
||
? content.map(part => part?.text || '').join('\n')
|
||
: String(content);
|
||
const usage = payload?.usage || {};
|
||
|
||
_logger.info('ocr finished', {
|
||
turnId,
|
||
requestId,
|
||
model: pickedModel,
|
||
kind,
|
||
profile,
|
||
bytes: imageBytes,
|
||
encodeMs,
|
||
requestMs: parseStartedAt - requestStartedAt,
|
||
parseMs: Date.now() - parseStartedAt,
|
||
elapsedMs: Date.now() - startedAt,
|
||
// token / 输出长度用于判断耗时主要花在图片 prefill 还是文本 decode
|
||
promptTokens: usage.prompt_tokens ?? usage.input_tokens ?? 0,
|
||
completionTokens: usage.completion_tokens ?? usage.output_tokens ?? 0,
|
||
totalTokens: usage.total_tokens ?? 0,
|
||
outputChars: raw.length,
|
||
finishReason: payload?.choices?.[0]?.finish_reason || ''
|
||
});
|
||
// 输出被截断意味着「原文转写」不完整:单独告警,便于排查长报告丢行
|
||
if (payload?.choices?.[0]?.finish_reason === 'length') {
|
||
_logger.warn('ocr output truncated', {
|
||
turnId,
|
||
requestId,
|
||
model: pickedModel,
|
||
outputChars: raw.length,
|
||
completionTokens: usage.completion_tokens ?? usage.output_tokens ?? 0
|
||
});
|
||
}
|
||
return parseVisionResult(raw, disclaimer).text;
|
||
},
|
||
|
||
/** 中止本轮识别:丢弃缓冲;已有请求一并取消,之后 finish 返回空串 */
|
||
abort() {
|
||
aborted = true;
|
||
chunks.length = 0;
|
||
activeController?.abort();
|
||
}
|
||
};
|
||
}
|
||
|
||
export default createQwenVisionSession;
|