Files
lik 59022dc5dd fix(agent): 放宽出诊清单输出与历史工具结果保留长度
输出上限支持按端配置并放宽到 escort 3000 / admin 5000;医院与网页类工具结果入历史保留 3000 字,数据库类维持 300 字,避免出诊信息被截断或追问时丢失。
2026-11-19 19:47:58 +08:00

212 lines
7.4 KiB
JavaScript
Raw Permalink Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
import logger from "../../utils/logger.js";
/**
* 输出护栏(主 agent 与管理 agent 共用)
*
* 流式输出无法回撤:字符一旦发给客户端,就会被追加进气泡。
* 因此这里只做两件在流式下可做对的兜底:
* 1. 禁用词拦截:命中即整段剔除;尾部若可能是禁用词的前缀就先不发送,
* 避免「作为A」+「I」这种跨 chunk 的命中被漏过;
* 2. 超长截断:总输出超过上限时在最近句末/换行处截断并补省略号,之后不再输出。
* 篇幅与文风仍由提示词负责,这里不放宽口径。
*/
// 单条回复的字符上限:需完整容纳一次出诊/排班清单,日常对话的篇幅仍由提示词控制在 1-3 句。
// 这里只是防止模型输出异常长文的兜底,不应把正常的清单类回答截断。
export const OUTPUT_MAX_CHARS = { escort: 3000, admin: 5000 };
/**
* 解析单条回复上限:配置可写成数字(两端共用)或按端对象 { escort, admin },
* 非法值回落到默认上限。
* @param {number|Object|undefined} configured conf.json 中的 agent.maxOutputChars
* @param {"escort"|"admin"} role 当前 agent 角色
* @param {number} fallback 默认上限
* @returns {number}
*/
export function resolveOutputMaxChars(configured, role, fallback) {
if (typeof configured === "number" && Number.isFinite(configured) && configured > 0) {
return configured;
}
const value = configured?.[role];
return typeof value === "number" && Number.isFinite(value) && value > 0 ? value : fallback;
}
// 身份泄露与空泛客套:两个 agent 都不该出现
const IDENTITY = [
"作为AI",
"作为一个AI",
"作为人工智能",
"作为一个人工智能",
"作为语言模型",
"作为大模型",
"我是一个AI",
"我是AI",
"AI助手",
// 与上面重叠的完整自我描述:按长度优先匹配,避免只删掉前半截留下「助手」
"作为AI助手",
"作为一个AI助手",
"作为一个人工智能助手",
"我是一个AI助手",
"我是AI助手",
"我是您的AI助手",
];
const COURTESY = [
"很高兴为您服务",
"很高兴为您解答",
"希望对您有帮助",
"希望这对您有帮助",
"您还有什么问题吗",
"还有其他问题吗",
"祝您早日康复",
];
// 越界承诺:医疗结果与服务可得性都不能打包票
const OVERPROMISE = [
"一定能治好",
"一定可以治好",
"保证治好",
"保证治愈",
"包治百病",
"百分之百安全",
"绝对安全",
];
export const OUTPUT_FORBIDDEN = [...IDENTITY, ...COURTESY, ...OVERPROMISE];
// 句末候选:中英文标点 + 换行,列表项按换行断开更自然
const SENTENCE_ENDS = ["。", "!", "?", ";", "\n", "!", "?", ";"];
// 在限额内尽量切在句末;找不到时硬切,并去掉尾部悬挂的标点/空白
function pickCut(text, room) {
// 留一个字符给省略号
const window = text.slice(0, Math.max(room - 1, 1));
const minIndex = Math.floor(window.length * 0.4);
let cutAt = -1;
for (const ch of SENTENCE_ENDS) {
const index = window.lastIndexOf(ch);
if (index >= minIndex && index > cutAt) {
cutAt = index;
}
}
const cut = cutAt >= 0 ? window.slice(0, cutAt + 1) : window;
return cut.replace(/[\s,、,]+$/u, "");
}
export class OutputGuard {
constructor({ label = "agent", maxChars = 400, forbidden = OUTPUT_FORBIDDEN } = {}) {
this.label = label;
this.maxChars = maxChars;
// 长的先匹配:避免「作为AI助手」被「作为AI」先删成半截
this.forbidden = forbidden.filter(Boolean).sort((a, b) => b.length - a.length);
this.buffer = "";
this.emitted = 0;
this.done = false;
this.hits = [];
// 禁用词被整段剔除后,下一段开头可能残留逗号/句号,需要一并吞掉
this._stripLeading = false;
}
/** 输入一段模型输出,返回可以发送给客户端的文本(可能为空串) */
feed(chunk) {
if (this.done || !chunk) return "";
if (this.emitted >= this.maxChars) {
this.done = true;
logger.warn(`[OutputGuard:${this.label}] 输出超过上限 ${this.maxChars} 字,后续内容已丢弃`);
return "";
}
const text = this.buffer + chunk;
this.buffer = "";
if (this.emitted + text.length > this.maxChars) {
const cut = pickCut(text, this.maxChars - this.emitted);
// 截断点若落在禁用词中间,退到词首,避免把半截词发出去
const hold = this._holdback(cut);
const kept = hold > 0 ? cut.slice(0, cut.length - hold) : cut;
// 已经输出过内容时,若剩余额度放不下一句完整的话,直接以省略号收尾,
// 避免把下一段的前几个字拼成半句(如「。后面还有…」)
const endsAtSentence = SENTENCE_ENDS.includes(kept.slice(-1));
const fragment = this.emitted > 0 && kept && !endsAtSentence ? "" : kept;
this.done = true;
const out = `${this._sanitize(fragment)}…`;
this.emitted += out.length;
logger.warn(`[OutputGuard:${this.label}] 输出超过上限 ${this.maxChars} 字,已在句末截断`, {
emitted: this.emitted,
dropped: text.length - fragment.length,
});
return out;
}
const hold = this._holdback(text);
const ready = hold > 0 ? text.slice(0, text.length - hold) : text;
if (hold > 0) {
this.buffer = text.slice(text.length - hold);
}
const out = this._sanitize(ready);
this.emitted += out.length;
return out;
}
/** 流结束时补发 holdback 里的残余文本 */
finish() {
if (this.done) return "";
this.done = true;
const out = this._sanitize(this.buffer);
this.buffer = "";
this.emitted += out.length;
return out;
}
// 尾部若可能是某个禁用词的前缀,先留在 buffer 等下一段拼起来再判断;
// 完整命中也要留(作为它自身的前缀),避免「作为」刚发出就在下一段补上「AI」
_holdback(text) {
let hold = 0;
for (const word of this.forbidden) {
const max = Math.min(word.length, text.length);
for (let len = max; len > hold; len--) {
if (word.startsWith(text.slice(text.length - len))) {
hold = len;
break;
}
}
}
return hold;
}
_sanitize(text) {
if (!text) return "";
let out = text;
if (this._stripLeading) {
out = out.replace(/^[,、,;;::。.!!??\s]+/u, "");
// 整段都是标点时继续保持标记,等下一段真正有正文再清除
this._stripLeading = out === "";
}
if (!this.forbidden.length) return out;
let hit = false;
for (const word of this.forbidden) {
if (!out.includes(word)) continue;
const removed = out.split(word).length - 1;
out = out.split(word).join("");
if (!this.hits.includes(word)) this.hits.push(word);
hit = true;
logger.warn(`[OutputGuard:${this.label}] 拦截禁用词「${word}」`, { removed });
}
if (!hit) return out;
// 剔除后留下的重复标点做最小修补;整段被剔除时空结果不宜直接发送
out = out
.replace(/([,、,])[,、,]+/gu, "$1")
.replace(/([。!?!?])[。!?!?]+/gu, "$1")
.replace(/[,、,]\s*([。!?!?])/gu, "$1")
.replace(/^[,、,;;::。.!!??\s]+/u, "");
this._stripLeading = out === "";
return out;
}
}
export default OutputGuard;