feat(ai-moderation): profile learner ai_status filter + channel-aware + simple fallback memory
- userProfileLearner.ts: filter query to only clean messages (eq ai_status='clean') to prevent profile contamination from flagged content. Also select channel_id to group messages by channel in prompt, enabling channel-aware personality summarization (user may behave differently across channels). - llmModerationClient.ts (runSimpleTextFallback): inject user profile into both the classify prompt and the reason prompt, so even the last-resort fallback path has personality/memory context instead of being blind.
This commit is contained in:
@@ -1516,6 +1516,17 @@ export async function runSimpleTextFallback(
|
||||
? content.slice(0, MAX_CONTENT_CHARS) + "..."
|
||||
: content;
|
||||
|
||||
// ── Inject user profile for personality-aware fallback ──
|
||||
let userProfileCtx = "";
|
||||
try {
|
||||
const profile = await getUserProfile(message.user_id);
|
||||
if (profile?.profile_summary) {
|
||||
userProfileCtx = `\n\nProfil pengirim pesan:\n${profile.profile_summary}\n`;
|
||||
}
|
||||
} catch {
|
||||
// Profile fetch failure is non-fatal — proceed without context
|
||||
}
|
||||
|
||||
// ── Step 1: Single-word classification ──
|
||||
const classifyPrompt = `Pesan berikut perlu diklasifikasikan sebagai: clean, warn, atau flagged.
|
||||
|
||||
@@ -1529,7 +1540,7 @@ PENTING (False Positive Prevention):
|
||||
- Konten coding/programming (kode, log error, SQL, command line, error message, stack trace, nama library) = clean. JANGAN flag hanya karena ada kata "error" atau "crash" dalam konteks teknis.
|
||||
- Nama proyek, tools, framework (IMPHNEN, Bete, Cursor, Claude, React, Discord) = clean.
|
||||
- Percakapan multilingual (campuran Indonesia-Inggris) = clean.
|
||||
|
||||
${userProfileCtx}
|
||||
Pesan: "${truncatedContent}"
|
||||
|
||||
Jawab HANYA dengan satu kata: clean, warn, atau flagged`;
|
||||
@@ -1582,7 +1593,7 @@ Jawab HANYA dengan satu kata: clean, warn, atau flagged`;
|
||||
const categoryOptions =
|
||||
status === "flagged" ? "harassment, gambling, atau sara" : "spam";
|
||||
const reasonPrompt = `Pesan berikut telah diklasifikasikan sebagai "${status}".
|
||||
|
||||
${userProfileCtx}
|
||||
Pesan: "${truncatedContent}"
|
||||
|
||||
Jelaskan dalam 1-2 kalimat Bahasa Indonesia: APA yang melanggar dan KENAPA. Jangan gunakan kata "mungkin" atau "sepertinya". Jangan tulis ulang pesan. Langsung ke alasan.
|
||||
|
||||
@@ -18,29 +18,48 @@ async function learnUserProfile(
|
||||
): Promise<void> {
|
||||
const db = getDatabase();
|
||||
|
||||
// Get recent messages for this user
|
||||
// Get recent CLEAN messages for this user (avoid profiling from flagged content)
|
||||
const recentMessages = await db
|
||||
.select({
|
||||
content: messagesTable.content,
|
||||
channelId: messagesTable.channel_id,
|
||||
})
|
||||
.from(messagesTable)
|
||||
.where(
|
||||
and(
|
||||
eq(messagesTable.user_id, userId),
|
||||
eq(messagesTable.guild_id, guildId),
|
||||
eq(messagesTable.ai_status, "clean"),
|
||||
),
|
||||
)
|
||||
.orderBy(desc(messagesTable.created_at))
|
||||
.limit(100);
|
||||
|
||||
if (recentMessages.length < 10) {
|
||||
log.debug({ userId }, "Not enough messages to learn user profile");
|
||||
log.debug({ userId }, "Not enough clean messages to learn user profile");
|
||||
return;
|
||||
}
|
||||
|
||||
// Group messages by channel for channel-aware profiling
|
||||
const channelGroups = new Map<string, { content: string; channelId: string }[]>();
|
||||
for (const msg of recentMessages) {
|
||||
const ch = msg.channelId ?? "unknown";
|
||||
if (!channelGroups.has(ch)) channelGroups.set(ch, []);
|
||||
channelGroups.get(ch)!.push(msg);
|
||||
}
|
||||
|
||||
// Build messages text with channel context
|
||||
const messagesText = recentMessages
|
||||
.reverse()
|
||||
.map((m) => m.content)
|
||||
.map((m) => {
|
||||
const chLabel = m.channelId ? `[#channel:${m.channelId}]` : "";
|
||||
return `${chLabel} ${m.content}`;
|
||||
})
|
||||
.join("\n");
|
||||
|
||||
// Build channel activity summary
|
||||
const channelSummary = [...channelGroups.entries()]
|
||||
.map(([ch, msgs]) => ` - #channel ${ch}: ${msgs.length} pesan`)
|
||||
.join("\n");
|
||||
|
||||
const prompt = `Anda adalah AI ahli psikologi, analisis perilaku online, dan pembaca karakter.
|
||||
@@ -48,11 +67,14 @@ Tugas Anda adalah merangkum profil kepribadian SEORANG PRIBADI — bukan sekadar
|
||||
gaya bicara — berdasarkan riwayat pesan-pesan mereka di server Discord.
|
||||
Buatlah ringkasan yang KAYA AKAN PERSONALITAS sehingga pembaca merasa "mengenal" orang ini.
|
||||
|
||||
Pesan-pesan terakhir dari user "${userId}":
|
||||
Pesan-pesan terakhir dari user "${userId}" (hanya pesan bersih/clean):
|
||||
<messages>
|
||||
${messagesText}
|
||||
</messages>
|
||||
|
||||
Distribusi aktivitas user per channel:
|
||||
${channelSummary}
|
||||
|
||||
Berdasarkan pesan-pesan di atas, buatlah ringkasan singkat (maksimal 4 paragraf)
|
||||
mengenai:
|
||||
|
||||
@@ -60,8 +82,9 @@ mengenai:
|
||||
Apakah orang ini suka pake singkatan, emot, reaksi berlebihan ("WKWKWK"), atau nada datar?
|
||||
Bagaimana mereka memulai dan mengakhiri pembicaraan?
|
||||
|
||||
2. **Topik-topik yang sering dibahas** — apa PASSION mereka? Coding, gaming, musik, debat?
|
||||
Apakah mereka inisiator topik atau lebih suka merespon?
|
||||
2. **Topik-topik yang sering dibahas & channel favorit** — apa PASSION mereka? Coding, gaming, musik, debat?
|
||||
Apakah mereka inisiator topik atau lebih suka merespon? Di channel mana mereka paling aktif?
|
||||
Apakah perilaku mereka berbeda tergantung channel (misal: profesional di #coding vs santai di #general)?
|
||||
|
||||
3. **Kepribadian dan karakter yang terpancar** — Apakah mereka ramah dan hangat? Kritis dan analitis?
|
||||
Easy going? Gampang marah? Humoris? Supportif? Suka memprovokasi? Suka membantu?
|
||||
|
||||
Reference in New Issue
Block a user