@danceiny/gotry 0.0.1-rc.7 → 0.0.1-rc.8

package/dist/scripts/time-eval-tests.js

@@ -0,0 +1,318 @@
+import assert from 'node:assert/strict';
+import { readFileSync } from 'node:fs';
+import { join } from 'node:path';
+import { buildTimeAnchor, parseAbsoluteDate } from '../src/time-anchor.js';
+import { detectLanguage, flagExpiredSlots, scoreExtraction } from '../src/travel-slots.js';
+import { resolveSlotDate, resolveSlots, specDateMismatches } from '../src/slot-spec.js';
+import { createMockLlm } from '../src/mock-llm.js';
+import { createOpenAICompatLlm } from '../src/dsh-llm.js';
+const dataset = JSON.parse(readFileSync(join('..', 'data', 'time-slot-eval.json'), 'utf-8'));
+const [ay, am, ad] = dataset.anchor_date.split('-').map(Number);
+const ANCHOR_NOW = new Date(ay, am - 1, ad, 12);
+{
+    const a = buildTimeAnchor(ANCHOR_NOW);
+    assert.equal(a.today, '2026-08-27', 'today');
+    assert.equal(a.todayWeekdayZh, '周四', '2026-08-27 是周四');
+    assert.match(a.tzLabel, /^UTC[+-]\d/, '时区标注格式');
+    const c = a.card;
+    assert.ok(c.includes('明天 2026-08-28 周五'), '明天');
+    assert.ok(c.includes('大后天 2026-08-30 周日'), '大后天');
+    assert.ok(c.includes('下周:周一 2026-08-31'), '下周一=08-31');
+    assert.ok(c.includes('周日 2026-09-06'), '下周日=09-06');
+    assert.ok(c.includes('周四 2026-09-10'), '下下周四=09-10');
+    assert.ok(c.includes('下个月 2026-09:初 09-01~09-10|中旬 09-11~09-20|下旬 09-21~09-30'), '下个月分段');
+    assert.ok(c.includes('下个季度初:2026-10-01 起'), '下个季度初');
+    assert.ok(c.includes('国庆 2026-10-01'), '国庆锚点');
+    assert.ok(c.includes('元旦 2027-01-01'), '元旦取下一次发生');
+    assert.ok(c.includes('春节 2027-02-06'), '春节锚点(2026 春节已过,取 2027)');
+    console.log('1. 锚点卡断言 OK(固定 2026-08-27:下周一/大后天/下下周四/月分段/季度/节日)');
+    const a2030 = buildTimeAnchor(new Date(2030, 5, 1, 12));
+    assert.ok(a2030.card.includes('春节 2031-01-23'), 'D-9 扩表:2030 年视视角春节取 2031-01-23');
+    console.log('1b. D-9 扩表断言 OK(2030 锚点 → 春节 2031-01-23,表覆盖至 2031)');
+}{
+    const anchor = buildTimeAnchor(ANCHOR_NOW);
+    const base = {
+        schema_version: 'travel_slot_extraction.v1',
+        language: 'zh',
+        domains: [
+            'requisition'
+        ],
+        slots: {
+            requisition: {
+                start_date: '8.1',
+                end_date: '8.10'
+            }
+        },
+        missing_slots: []
+    };
+    const flagged = flagExpiredSlots(base, anchor);
+    assert.deepEqual(flagged.missing_slots, [
+        'requisition.start_date (date is expired)'
+    ], '8.1 在 08-27 锚点下过期');
+    assert.equal(flagged.slots.requisition.start_date, '8.1', '槽位原值逐字保留');
+    assert.equal(flagged.missing_slots.length, 1, 'end_date 不重复判(只判主日期字段,对齐 golden)');
+    const again = flagExpiredSlots(flagged, anchor);
+    assert.equal(again.missing_slots.length, 1, '幂等不重复加');
+    const future = flagExpiredSlots({
+        ...base,
+        slots: {
+            requisition: {
+                start_date: '9.1'
+            }
+        }
+    }, anchor);
+    assert.equal(future.missing_slots.length, 0, '9.1 不过期');
+    const relative = flagExpiredSlots({
+        ...base,
+        slots: {
+            requisition: {
+                start_date: '本周三'
+            }
+        }
+    }, anchor);
+    assert.equal(relative.missing_slots.length, 0, '相对表达(本周三)永不判过期');
+    const iso = flagExpiredSlots({
+        ...base,
+        slots: {
+            requisition: {
+                start_date: '2026-08-01'
+            }
+        }
+    }, anchor);
+    assert.equal(iso.missing_slots.length, 1, 'ISO 绝对日期判过期');
+    const zhDate = flagExpiredSlots({
+        ...base,
+        slots: {
+            requisition: {
+                start_date: '8月5日'
+            }
+        }
+    }, anchor);
+    assert.equal(zhDate.missing_slots.length, 1, '中文月日判过期');
+    const enDate = flagExpiredSlots({
+        ...base,
+        slots: {
+            requisition: {
+                start_date: 'Aug-5'
+            }
+        }
+    }, anchor);
+    assert.equal(enDate.missing_slots.length, 0, '英文月名不判过期(逐字传递,对齐 golden time_english_date_range)');
+    const flight = flagExpiredSlots({
+        ...base,
+        domains: [
+            'flight'
+        ],
+        slots: {
+            flight: {
+                departure_date: '8.1'
+            }
+        }
+    }, anchor);
+    assert.deepEqual(flight.missing_slots, [
+        'flight.departure_date (date is expired)'
+    ], 'flight 主日期字段');
+    assert.equal(parseAbsoluteDate('下周一', 2026), null, 'parseAbsoluteDate 不碰相对表达');
+    for (const c of dataset.cases){
+        assert.equal(detectLanguage(c.utterance), c.expected.language, `detectLanguage ${c.id}`);
+    }
+    console.log('2. 过期校验断言 OK(8.1 过期/9.1 不/本周三不/ISO+中文月日判/英文不判/幂等)+ detectLanguage 全题对齐');
+}{
+    const exp = dataset.cases[0].expected;
+    const identical = structuredClone(exp);
+    assert.equal(scoreExtraction(exp, identical).pass, true, '完全一致应 pass');
+    const wrongLang = {
+        ...structuredClone(exp),
+        language: 'en'
+    };
+    assert.equal(scoreExtraction(exp, wrongLang).pass, false, 'language 不一致应 fail');
+    const wrongDomain = {
+        ...structuredClone(exp),
+        domains: [
+            'flight'
+        ]
+    };
+    assert.equal(scoreExtraction(exp, wrongDomain).pass, false, 'domains 不一致应 fail');
+    const wrongSlot = structuredClone(exp);
+    wrongSlot.slots.requisition.destination = '北京';
+    const r = scoreExtraction(exp, wrongSlot);
+    assert.equal(r.pass, false, '槽位值不一致应 fail');
+    assert.ok(r.diffs.some((d)=>d.includes('slots.requisition.destination')), 'diff 带路径');
+    const wrongMissing = {
+        ...structuredClone(exp),
+        missing_slots: [
+            'requisition.start_date (date is expired)'
+        ]
+    };
+    assert.equal(scoreExtraction(exp, wrongMissing).pass, false, 'missing_slots 不一致应 fail');
+    const extraKey = structuredClone(exp);
+    extraKey.slots.requisition['adults'] = 2;
+    const rk = scoreExtraction(exp, extraKey);
+    assert.equal(rk.pass, true, '多出键不 fail(非致命)');
+    assert.equal(rk.warnings.length, 1, '多出键记 warning');
+    console.log('3. 评分器自测 OK(pass/language/domains/槽位值/missing_slots/extra-key warning)');
+}{
+    const script = dataset.cases.map((c)=>({
+            when: c.utterance,
+            extraction: c.expected
+        }));
+    const llm = createMockLlm(undefined, script);
+    let pass = 0;
+    for (const c of dataset.cases){
+        const ext = await llm.extractSlots([
+            {
+                role: 'user',
+                text: c.utterance
+            }
+        ]);
+        assert.ok(ext, `mock 应命中 ${c.id}`);
+        const r = scoreExtraction(c.expected, ext);
+        assert.ok(r.pass, `mock 回放 ${c.id} 应 pass:${r.diffs.join(';')}`);
+        pass++;
+    }
+    console.log(`4. mock 回放管道 OK(${pass}/${dataset.cases.length}——评分管道自身正确)`);
+}{
+    const anchor = buildTimeAnchor(ANCHOR_NOW);
+    const r = (expr)=>resolveSlotDate(expr, anchor);
+    assert.equal(r('2026-09-05').date, '2026-09-05', 'ISO');
+    assert.equal(r('8.20').date, '2026-08-20', '点分');
+    assert.equal(r('9月5日').date, '2026-09-05', '中文月日');
+    assert.equal(r('下周一').date, '2026-08-31', '下周一');
+    assert.equal(r('本周六').date, '2026-08-29', '本周六');
+    assert.equal(r('下下周四').date, '2026-09-10', '下下周四');
+    assert.equal(r('明天').date, '2026-08-28', '明天');
+    assert.equal(r('下个月中旬').date, '2026-09-11', '下个月中旬取首日');
+    assert.equal(r('下周五+3').date, '2026-09-07', '下周五+3');
+    assert.equal(r('8.20+2天').date, '2026-08-22', '8.20+2天');
+    assert.equal(r('近期').kind, 'unresolved', '模糊词');
+    assert.equal(r('这阵子').date, null, '模糊词不猜');
+    assert.equal(r('next Monday').kind, 'unresolved', '英文相对不解析');
+    assert.equal(r('next Monday').raw, 'next Monday', 'unresolved 保留原话');
+    const ext = {
+        schema_version: 'travel_slot_extraction.v1',
+        language: 'zh',
+        domains: [
+            'hotel',
+            'flight'
+        ],
+        slots: {
+            hotel: {
+                action: 'search',
+                city: '上海',
+                check_in_date: '下周五',
+                check_out_date: '下周五+3'
+            },
+            flight: {
+                action: 'search',
+                origin: '上海',
+                destination: '杭州',
+                departure_date: '这阵子',
+                trip_type: 'one_way'
+            }
+        },
+        missing_slots: []
+    };
+    const resolved = resolveSlots(ext, anchor);
+    assert.equal(resolved.hotel?.check_in_date?.date, '2026-09-04', 'hotel 入住=下周五');
+    assert.equal(resolved.hotel?.check_out_date?.date, '2026-09-07', 'hotel 退房=下周五+3');
+    assert.deepEqual(resolved.unresolved, [
+        {
+            field: 'flight.departure_date',
+            raw: '这阵子'
+        }
+    ], 'unresolved 清单');
+    assert.equal(specDateMismatches({
+        segments: [
+            {
+                date: '2026-09-04'
+            }
+        ]
+    }, resolved).length, 0, '一致则零 mismatch');
+    const mm = specDateMismatches({
+        segments: [
+            {
+                date: '2026-09-05'
+            }
+        ]
+    }, resolved);
+    assert.equal(mm.length, 1, '单日期段分歧暴露');
+    assert.equal(mm[0]?.specDate, '2026-09-05');
+    assert.equal(mm[0]?.slotDate, '2026-09-04');
+    assert.equal(specDateMismatches({
+        segments: [
+            {
+                date: '2026-09-04'
+            },
+            {
+                date: '2026-09-05'
+            }
+        ]
+    }, resolved).length, 0, '多段无逐段真值不判(巡检修正)');
+    const unresolvedOnly = {
+        unresolved: [
+            {
+                field: 'flight.departure_date',
+                raw: '近期'
+            }
+        ]
+    };
+    assert.equal(specDateMismatches({
+        segments: [
+            {
+                date: '2026-09-05'
+            }
+        ]
+    }, unresolvedOnly).length, 0, '全 unresolved 不参与比对');
+    assert.equal(specDateMismatches({
+        segments: []
+    }, resolved).length, 0, 'spec 无日期不比对');
+    console.log('5. 槽位→日期解析 OK(绝对/词表/+N/unresolved 边界/整张解析/spec 一致性闸)');
+}console.log('\nTIME-EVAL TESTS: 5/5 OK(确定性部分,CI 口径)');
+if (process.argv.includes('--real')) {
+    try {
+        for (const line of readFileSync(join('..', '.env'), 'utf-8').split('\n')){
+            const m = line.match(/^([A-Z_]+)=(.*)$/);
+            if (m && !process.env[m[1]]) process.env[m[1]] = m[2].trim();
+        }
+    } catch  {}
+    const hasKey = Boolean(process.env['LLM_API_KEY'] ?? process.env['DEEPSEEK_API_KEY']);
+    if (!hasKey) {
+        console.log('\n--real:无 LLM_API_KEY,跳过真模型巡检(ADR-8 回退原则)');
+    } else {
+        const clock = ()=>ANCHOR_NOW;
+        const llm = createOpenAICompatLlm(undefined, clock);
+        console.log(`\n=== 真模型巡检(${process.env['LLM_MODEL'] ?? 'MiniMax-M2'},锚点 ${dataset.anchor_date})===`);
+        let pass = 0;
+        let p0Pass = 0;
+        let p0Total = 0;
+        for (const c of dataset.cases){
+            if (c.priority === 'P0') p0Total++;
+            try {
+                const ext = await llm.extractSlots([
+                    {
+                        role: 'user',
+                        text: c.utterance
+                    }
+                ]);
+                if (!ext) {
+                    console.log(`FAIL ${c.id}(${c.priority}):模型未产出合法 v1 JSON`);
+                    continue;
+                }
+                const r = scoreExtraction(c.expected, ext);
+                if (r.pass) {
+                    pass++;
+                    if (c.priority === 'P0') p0Pass++;
+                    console.log(`PASS ${c.id}${r.warnings.length ? `(warning: ${r.warnings.join(';')})` : ''}`);
+                } else {
+                    console.log(`FAIL ${c.id}(${c.priority}):\n  ${r.diffs.join('\n  ')}`);
+                }
+            } catch (e) {
+                console.log(`ERROR ${c.id}(${c.priority}):${e.message.slice(0, 120)}`);
+            }
+        }
+        console.log(`\nREAL ACCURACY: ${pass}/${dataset.cases.length}(P0: ${p0Pass}/${p0Total})——巡检报告口径,不设红线`);
+    }
+}
+
+
+//# sourceURL=/Users/bytedance/work/gotry/ts/scripts/time-eval-tests.ts

package/dist/src/dsh-llm.js

@@ -1,17 +1,19 @@
 import { parseFlightPackToSpec } from './unified.js';
-const MODEL = process.env['LLM_MODEL'] ?? process.env['DEEPSEEK_MODEL'] ?? 'MiniMax-M2';
-const BASE = (process.env['LLM_BASE_URL'] ?? process.env['DEEPSEEK_BASE_URL'] ?? 'https://api.minimax.io/v1').replace(/\/$/, '');
+import { buildTimeAnchor } from './time-anchor.js';
+import { buildSlotSystem, flagExpiredSlots, normalizeExtraction } from './travel-slots.js';
+const model = ()=>process.env['LLM_MODEL'] ?? process.env['DEEPSEEK_MODEL'] ?? 'MiniMax-M2';
+const base = ()=>(process.env['LLM_BASE_URL'] ?? process.env['DEEPSEEK_BASE_URL'] ?? 'https://api.minimax.io/v1').replace(/\/$/, '');
 async function chat(messages, json) {
     const key = process.env['LLM_API_KEY'] ?? process.env['DEEPSEEK_API_KEY'];
     if (!key) throw new Error('LLM_API_KEY 未设置(兼容 DEEPSEEK_API_KEY 别名)——真 LLM 路径不可用,请回退 mock(ADR-8)');
-    const res = await fetch(`${BASE}/chat/completions`, {
+    const res = await fetch(`${base()}/chat/completions`, {
         method: 'POST',
         headers: {
             'Content-Type': 'application/json',
             Authorization: `Bearer ${key}`
         },
         body: JSON.stringify({
-            model: MODEL,
+            model: model(),
             messages,
             ...json ? {
                 response_format: {
@@ -46,9 +48,10 @@ const SKELETON_SYSTEM = `你是行程骨架抽取器。从对话中抽取行程
 输出 JSON:{"scenario":"erhai|workation|yunnan|generic","segments":[{"id":"f1","role":"choice|fixed","route":"HKG->HKT","dateHint":"2026-07-18","anchors":{"arriveByMin":885}}]}
 规则:每个跨城移动一段;锚点只放用户明说或必然的(如"当天到"→arriveByMin 23:59=1439);时刻用当日分钟。
 scenario 判定:「洱海/大理/千岛湖/太湖+选目的地」→erhai(候选集);「普吉/workation/远程办公+多城链」→workation(五段链);「云南/大理丽江」→yunnan;不确定→generic。只输出 JSON。`;
-export function createOpenAICompatLlm(flightPackPath) {
+export function createOpenAICompatLlm(flightPackPath, clock = ()=>new Date()) {
     const pack = flightPackPath;
     const historyText = (h)=>h.map((t)=>`${t.role === 'user' ? '用户' : '助手'}: ${t.text}`).join('\n');
+    const anchorContext = ()=>`时间锚点卡:\n${buildTimeAnchor(clock()).card}`;
     return {
         async extractFacts (history) {
             const out = await chat([
@@ -58,7 +61,7 @@ export function createOpenAICompatLlm(flightPackPath) {
                 },
                 {
                     role: 'user',
-                    content: historyText(history)
+                    content: `${anchorContext()}\n\n${historyText(history)}`
                 }
             ], true);
             const obj = parseJsonBlock(out);
@@ -78,7 +81,7 @@ export function createOpenAICompatLlm(flightPackPath) {
             };
         },
         async extractSpec (history, state) {
-            const context = `已断言日历:${JSON.stringify(state.calendar.assertedWeekdays)}\nprofile:${JSON.stringify(state.profile)}`;
+            const context = `${anchorContext()}\n已断言日历:${JSON.stringify(state.calendar.assertedWeekdays)}\nprofile:${JSON.stringify(state.profile)}`;
             const out = await chat([
                 {
                     role: 'system',
@@ -134,6 +137,23 @@ export function createOpenAICompatLlm(flightPackPath) {
             packSpec.budgetCny = 9000;
             return packSpec;
         },
+        async extractSlots (history, now) {
+            const anchor = buildTimeAnchor(now ?? clock());
+            const out = await chat([
+                {
+                    role: 'system',
+                    content: buildSlotSystem(anchor)
+                },
+                {
+                    role: 'user',
+                    content: historyText(history)
+                }
+            ], true);
+            const obj = parseJsonBlock(out);
+            if (!obj) return null;
+            const ext = normalizeExtraction(obj, history.map((t)=>t.text).join('\n'));
+            return ext ? flagExpiredSlots(ext, anchor) : null;
+        },
         async polishQuestion (q) {
             const out = await chat([
                 {