iqbal/learning-lab-express
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Make the Brain agent ask instead of guess, and answer how schedules run

Browse Source 
Generic requests ("10 konten", "10 pdf") now become a rule whose kind the
agent asks about, and a rule with no frequency is asked about instead of
silently becoming weekly. Times like 09.00 no longer split sentences, and
requests about drafts, the hour, scheduling, sources and duplicates get a
truthful note rather than an "I did not understand" loop.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Iqbal
2026-10-28 14:09:27 +07:00
parent d018c6826f
commit b63e9c2ca8
5 changed files with 293 additions and 33 deletions
Download Patch File Download Diff File Expand all files Collapse all files

109
src/express/parse.ts
View File

@@ -30,6 +30,20 @@ const KIND_WORDS: [ContentKind, RegExp][] = [
['summary', /\b(summary|summaries|ringkasan|rangkuman|artikel|articles?|teks|texts?)\b/],
]
/** Every content kind a text names, in the order it names them: "ringkasan dan kuis" → summary, quiz. */
export function kindsIn(text: string): ContentKind[] {
return KIND_WORDS
.map(([kind, re]) => [kind, text.search(re)] as const)
.filter(([, at]) => at >= 0)
.sort((a, b) => a[1] - b[1])
.map(([kind]) => kind)
}
// "10 konten", "5 materi pembelajaran", "10 pdf": content of no particular kind, so the agent asks which.
// Only with a count or a create verb before it, so "gunakan seluruh materi" is not a request.
const GENERIC = /\b(konten|content|contents|materi|materials?|bahan ajar|pdf|pdfs|dokumen|documents?|modul|modules?|items?)\b/
const CREATE = /\b(bikin|bikinin|buat|buatkan|membuat|hasilkan|produksi|make|create|generate|produce|write|tulis)\b/
const NUMBER_WORDS: Record<string, number> = {
satu: 1, se: 1, sebuah: 1, dua: 2, tiga: 3, empat: 4, lima: 5, enam: 6, tujuh: 7, delapan: 8, sembilan: 9, sepuluh: 10,
a: 1, an: 1, one: 1, two: 2, three: 3, four: 4, five: 5, six: 6, seven: 7, eight: 8, nine: 9, ten: 10,
@@ -77,6 +91,10 @@ export interface ContentRule {
language?: OutputLanguage
/** Made from material chosen in the Brain panel rather than the whole Brain. */
chosen?: boolean
/** The instruction said "content" without a kind, so the kind is only a placeholder the agent asks about. */
kindGuessed?: boolean
/** The instruction named no frequency, so per week is only a placeholder the agent asks about. */
perGuessed?: boolean
}
export interface ResearchRule {
@@ -91,37 +109,81 @@ export interface ResearchRule {
export type Rule = ContentRule | ResearchRule
/** Something the instruction asked for that is not a rule: the agent answers it truthfully instead of ignoring it. */
export interface Note { key: Key; vars?: Record<string, string | number> }
export interface ParseResult {
rules: Rule[]
/** Clauses that said something but matched no rule, so the user can rephrase them. */
unknown: string[]
/** The output language the instruction named. */
language?: OutputLanguage
notes: Note[]
}
// Requests about how schedules work rather than what they make. Each is answered with what Express
// actually does, so the agent neither ignores them nor promises something it cannot do.
const META: [RegExp, Key | null][] = [
[/\b(?:seluruh|semua)\s+(?:materi|bahan|dokumen|otak|brain|isi)|\bwhole brain\b|\ball (?:the )?(?:materials?|documents?|brain)\b/, 'note.wholeBrain'],
[/duplika|duplicat|\b(?:tidak|jangan|tanpa|bukan)\s+(?:meng)?ulang|\b(?:without|not|don'?t|no)\s+repeat|\b(?:topik|topics?)\s+(?:yang\s+)?(?:berbeda|beda|different|unik|unique)|\bdifferent topics?\b/, 'note.novelty'],
[/\b(?:draf|draft|drafts|ditinjau|tinjau|peninjauan|reviews?|reviewed|approval|persetujuan|disetujui|dipublikasikan|publikasi|published?)\b/, 'note.draft'],
[/\b(?:scheduler|penjadwal|background|latar belakang|cron|otomatis|automatic(?:ally)?|terjadwal)\b/, 'note.automatic'],
// Wishes every schedule already aims for; nothing to add, but not a part the agent failed to understand.
[/\b(?:relevan|relevant|informatif|informative|berkualitas|bermanfaat|menarik|useful|engaging|quality|referensi|reference)\b/, null],
]
const HOUR = /\b(?:pukul|pkl|jam|at)\s*(\d{1,2})(?:[.:](\d{2}))?\s*(am|pm)?/
const SUB_DAY = /\b\d+\s*(?:menit|jam|detik|minutes?|mins?|hours?|seconds?)\b|\b(?:tiap|setiap|per|every|each)\s+(?:menit|jam|minute|hour)\b|\bhourly\b/
function notesOf(text: string): Note[] {
const notes: Note[] = META.flatMap(([re, key]) => (key && re.test(text) ? [{ key }] : []))
const hour = text.match(HOUR)
if (hour) {
let h = Number(hour[1])
if (hour[3] === 'pm' && h < 12) h += 12
if (hour[3] === 'am' && h === 12) h = 0
const at = `${String(h).padStart(2, '0')}.${hour[2] ?? '00'}`
notes.push({ key: at === '09.00' ? 'note.hourMatches' : 'note.hourFixed', vars: { at } })
}
if (SUB_DAY.test(text)) notes.push({ key: 'note.subDay' })
return notes
}
const isMeta = (clause: string) => META.some(([re]) => re.test(clause)) || HOUR.test(clause) || SUB_DAY.test(clause)
// "dari file LLM" / "from the document onboarding" names a Brain file, which works as the topic.
const TOPIC = /\b(?:soal|tentang|mengenai|seputar|terkait|topik|bertema|perihal|about|regarding|covering|related to|on the topic of|topic|(?:dari|from)(?: the)? (?:file|berkas|dokumen|documents?|materi|material))\s+(?!latihan\b)(.+)$/
const FREQUENCY_TAIL = /\b(tiap|setiap|per|sehari|seminggu|every|each|daily|weekly|monthly|a day|a week|a month)\b.*$/
const TOPIC = /\b(?:soal|tentang|mengenai|seputar|terkait|topik|bertema|perihal|about|regarding|covering|related to|on the topic of|topic|(?:dari|from)(?: the)? (?:file|berkas|dokumen|documents?|materi|material))\s+(?!latihan\b|yang\b)(.+)$/
const FREQUENCY_TAIL = /\b(tiap|setiap|per|sehari|seminggu|every|each|daily|weekly|monthly|a day|a week|a month|pada pukul|pada jam|pukul|pkl|secara|otomatis|automatically|dalam \d|in \d|at \d)\b.*$/
function topicOf(clause: string): string {
return clause.match(TOPIC)?.[1].replace(FREQUENCY_TAIL, '').replace(/[\s,]+$/, '').trim() ?? ''
}
function countOf(clause: string, kindAt: number): number {
/** The count right before a kind word, or null when the instruction named none. */
function countBefore(clause: string, kindAt: number): number | null {
const before = clause.slice(0, kindAt)
const digits = before.match(/(\d+)\s*(?:buah|x|kali|new)?\s*$/)
// "0" schedules nothing; the preview shows 1, which the user can still change.
if (digits) return Math.max(1, Number(digits[1]))
const word = before.match(/(\p{L}+)\s*(?:buah|new)?\s*$/u)?.[1]
return (word && NUMBER_WORDS[word]) || 1
return (word && NUMBER_WORDS[word]) || null
}
const countOf = (clause: string, kindAt: number) => countBefore(clause, kindAt) ?? 1
const DAY_WORDS = /\b(per ?hari|tiap hari|setiap hari|sehari|harian|per day|a day|every day|each day|daily)\b/
const WEEK_WORDS = /\b(per ?minggu|tiap minggu|setiap minggu|seminggu|mingguan|per week|a week|every week|each week|weekly)\b/
const MONTH_WORDS = /\b(per ?bulan|tiap bulan|setiap bulan|sebulan|bulanan|per month|a month|every month|each month|monthly)\b/
export function perOf(clause: string): Per | 'month' {
if (/\b(per ?hari|tiap hari|setiap hari|sehari|harian|per day|a day|every day|each day|daily)\b/.test(clause)) return 'day'
if (/\b(per ?bulan|tiap bulan|setiap bulan|sebulan|bulanan|per month|a month|every month|each month|monthly)\b/.test(clause)) return 'month'
if (DAY_WORDS.test(clause)) return 'day'
if (MONTH_WORDS.test(clause)) return 'month'
return 'week'
}
/** Whether a clause says how often, rather than leaving perOf to fall back to per week. */
export const namesFrequency = (clause: string) =>
DAY_WORDS.test(clause) || WEEK_WORDS.test(clause) || MONTH_WORDS.test(clause) || weekdayOf(clause) !== null
export function weekdayOf(clause: string): number | null {
for (const [re, day] of WEEKDAYS) if (re.test(clause)) return day
return null
@@ -130,7 +192,8 @@ export function weekdayOf(clause: string): number | null {
const CONNECTORS = 'dan|lalu|serta|terus|kemudian|juga|and|then|also|plus'
// A connector or comma starts a new rule only when a new request follows; "about sales and marketing" stays one topic.
const NEXT_REQUEST = String.raw`(?=(?:\d|satu|dua|tiga|empat|lima|enam|tujuh|delapan|sembilan|sepuluh|se(?:buah)?\s|one|two|three|four|five|six|seven|eight|nine|ten|an?\s|riset|research|bikin|buat|cari|telusuri|pantau|make|create|generate|search|monitor|tiap|setiap|every|each))`
const SPLIT = new RegExp(String.raw`[\n.;]+|\s*,\s*(?:(?:${CONNECTORS})\s+)?${NEXT_REQUEST}|\s+(?:${CONNECTORS})\s+${NEXT_REQUEST}`)
// A full stop ends a sentence, but not inside a time or number such as "09.00".
const SPLIT = new RegExp(String.raw`(?:[\n;]|\.(?!\d))+|\s*,\s*(?:(?:${CONNECTORS})\s+)?${NEXT_REQUEST}|\s+(?:${CONNECTORS})\s+${NEXT_REQUEST}`)
function parseClause(clause: string): Rule | null {
if (/\b(riset|research|telusuri|cari(?:kan)?|pantau|look up|search|monitor)\b/.test(clause)) {
@@ -143,17 +206,25 @@ function parseClause(clause: string): Rule | null {
}
for (const [kind, re] of KIND_WORDS) {
const m = re.exec(clause)
if (!m) continue
const count = countOf(clause, m.index)
const per = perOf(clause)
if (per === 'month') {
return { type: 'content', kind, count: Math.max(1, Math.round(count / 4)), per: 'week', topic: topicOf(clause), monthly: count }
}
return { type: 'content', kind, count, per, topic: topicOf(clause) }
if (m) return contentRule(clause, kind, countOf(clause, m.index))
}
const g = GENERIC.exec(clause)
if (g) {
const count = countBefore(clause, g.index)
if (count !== null || CREATE.test(clause.slice(0, g.index))) return { ...contentRule(clause, 'summary', count ?? 1), kindGuessed: true }
}
return null
}
function contentRule(clause: string, kind: ContentKind, count: number): ContentRule {
const per = perOf(clause)
const guessed = namesFrequency(clause) ? {} : { perGuessed: true }
if (per === 'month') {
return { type: 'content', kind, count: Math.max(1, Math.round(count / 4)), per: 'week', topic: topicOf(clause), monthly: count }
}
return { type: 'content', kind, count, per, topic: topicOf(clause), ...guessed }
}
export function parseInstruction(text: string): ParseResult {
const language = languageOf(text)
const clauses = text
@@ -170,6 +241,8 @@ export function parseInstruction(text: string): ParseResult {
for (const clause of clauses) {
const rule = parseClause(clause)
if (rule) rules.push(rule)
// "simpan sebagai draft", "jalankan otomatis": answered as notes, not asked about.
else if (isMeta(clause)) continue
// A clause that only names a topic ("about leadership") applies to the rules around it.
else if (topicOf(clause) || /^(soal|tentang|topik|about|topic)\b/.test(clause)) shared = topicOf(clause) || clause.replace(/^\S+\s*/, '')
else if (!/^(tolong|bikin|buat|buatkan|ya|please|make|create|ok)$/.test(clause)) unknown.push(clause)
@@ -187,7 +260,8 @@ export function parseInstruction(text: string): ParseResult {
if (at >= 0) r.topic = text.slice(at, at + r.topic.length)
}
}
return language ? { rules, unknown, language } : { rules, unknown }
const notes = notesOf(text.toLowerCase())
return language ? { rules, unknown, language, notes } : { rules, unknown, notes }
}
export function weeklyTarget(rule: ContentRule): number {
@@ -209,10 +283,11 @@ export function describeRule(rule: Rule, from?: string): string {
const when = rule.cadence === 'daily' ? t('rule.everyDay') : t('rule.everyWeekday', { day: dayName(rule.weekday) })
return t('rule.research', { when, about })
}
// A placeholder kind or frequency is shown as still open, not as if the user had chosen it.
return t('rule.content', {
kind: kindLabel(rule.kind),
kind: rule.kindGuessed ? t('rule.anyKind') : kindLabel(rule.kind),
count: rule.count,
per: t(rule.per === 'day' ? 'ins.perDay' : 'ins.perWeek'),
per: rule.perGuessed ? t('rule.perUnset') : t(rule.per === 'day' ? 'ins.perDay' : 'ins.perWeek'),
about: about || (from ?? (rule.chosen ? t('rule.chosenMaterial') : null)
? t('rule.from', { source: from ?? t('rule.chosenMaterial') })
: t('rule.wholeBrain')),