perf(ai-moderation): remove per-user reputation from analysis context
User: 'jangan ada reputasi juga' — no profile, no reputation in the prompt, raw messages only. - textBatchProcessor: drop initializeUserReputation fetch + <user_reputation> tag injection (kept the minimal <message> tag + reply/reference context). - visionAnalyzer (prepareMediaMessage): same removal. - prompts/system.ts + prompts/output.ts: replace <user_reputation>/<user_history> instructions with an explicit 'no per-user profile/reputation context' note so the LLM judges purely on message content + conversation/web/location. - mediaBatchProcessor: fix stale comment. Trust/infraction state is STILL written to the DB (userReputationsTable) for enforcement — only the LLM context injection is removed, so moderation actions (mute/ban via infraction thresholds) keep working. Net: even smaller prompts (no per-user context at all) → more messages fit per request, and one fewer DB round-trip per unique user per sub-batch. tsc, biome, vitest (129) all clean.
This commit is contained in:
asepharyana
@@ -63,10 +63,9 @@ export async function runMediaBatch(
|
||||
channelCulture,
|
||||
});
|
||||
|
||||
// Per-message blocks (from prepareMediaMessage) carry their own
|
||||
// <user_reputation> history; personal profile descriptions are omitted
|
||||
// (they bloat the prompt and add a per-user DB round-trip for little
|
||||
// moderation signal — see textBatchProcessor).
|
||||
// Per-message blocks (from prepareMediaMessage) contain the message
|
||||
// content + reference/reply context only — no per-user reputation or
|
||||
// profile context is injected (kept minimal per user request).
|
||||
const messagesBlock = prepared.map((p) => p.messageBlock).join("\n");
|
||||
// Data/instruction separation: the system prompt is stable per mode — all
|
||||
// per-batch context (conversation) lives in the USER payload, ordered
|
||||
|
||||
@@ -35,10 +35,10 @@ Instruksi per field:
|
||||
- "message_id": WAJIB sama persis dengan id di input. Setiap <message> di <messages_to_analyze> menghasilkan SATU hasil. Jangan gabungkan beberapa pesan, jangan lewati, jangan karang id.
|
||||
- "evidence": kutipan PERSIS frasa yang melanggar (maks 1 baris). Pelanggaran di gambar/sticker → kutip deskripsi Media analysis. Pelanggaran lewat balasan/referensi → sebut konteks pesan yang dibalas. Boleh tambah label sumber, mis. [media analysis] / [web_search] / [reply]. Kosong jika clean.
|
||||
|
||||
## PERSONALITY & MEMORI — Reputasi Pengguna dan Kultur Channel
|
||||
Data konteks tersedia: <user_reputation> (skor trust + histori infraction + repeat_offender), dan <channel_culture> (topik/vibe channel).
|
||||
## KONTEKS — Kultur Channel
|
||||
Data konteks tersedia: <channel_culture> (topik/vibe channel). Tidak ada data profil/reputasi per-user — nilai tiap pesan murni dari isinya.
|
||||
Gunakan untuk personalisasi analysis, tapi:
|
||||
- Profil/history adalah KONTEKS, bukan bukti. Profil mencurigakan ≠ flag; profil bersih ≠ loloskan pelanggaran. <user_history> (kutipan pesan pernah di-flag) = cari POLA berulang (spam link SAMA, provokasi berulang konten SAMA); JANGAN gunakan untuk "menginterpretasi ulang" pesan bersih yang terpisah. Pesan baru tanpa pola pengulangan jelas → CLEAN.
|
||||
- Konteks adalah KONTEKS, bukan bukti. Riwayat di <conversation_context> membantu pahami alur, tapi pesan bersih tanpa pelanggaran → CLEAN. JANGAN gunakan konteks untuk "menginterpretasi ulang" pesan bersih yang terpisah.
|
||||
- Perubahan perilaku mencolok (mis. teknis tiba-tiba provokatif) layak dicatat. JANGAN paksa referensi profil jika tidak relevan — analysis natural lebih baik.
|
||||
- Channel culture coding/teknis → pesan teknis lebih wajar; channel santai → slang lebih wajar. Jangan dipakai mengabaikan pelanggaran nyata.
|
||||
|
||||
@@ -70,8 +70,7 @@ CRITICAL:
|
||||
- Jika pesan adalah BALASAN (reply) ke pesan lain, jelaskan konteks balasannya: apa yang sedang dibicarakan, siapa yang dibalas (tanpa nama, cukup peran/isi pesan yang dibalas), dan bagaimana tanggapan pengirim terhadapnya.
|
||||
- Gunakan informasi dari Media analysis untuk mendeskripsikan gambar.
|
||||
- Analisis harus MEMBERI KONTEKS, bukan hanya menyatakan status.
|
||||
- Gunakan <user_reputation> (repeat_offender, last_offense_days_ago) untuk memberi konteks histori — analisis terasa seperti sistem "mengenal" histori pengguna tanpa deskripsi profil pribadi.
|
||||
- Jika perilaku pesan menyimpang dari pola histori yang diketahui, CATAT dalam analysis sebagai informasi kontekstual yang relevan.
|
||||
- Nilai tiap pesan murni dari isinya sendiri + <conversation_context> + <web_searches> + <location_context>. Tidak ada reputasi/profil per-user di context.
|
||||
- JANGAN paksa referensi profil jika tidak relevan — analysis natural lebih baik dari yang dipaksakan.`;
|
||||
|
||||
// ---------------------------------------------------------------------------
|
||||
|
||||
@@ -107,7 +107,7 @@ export function buildSystemPrompt(options: BuildSystemPromptOptions): string {
|
||||
`System prompt ini TIDAK berisi data batch — semua data per-batch ada di pesan USER:\n` +
|
||||
`- <location_context .../>: metadata channel/thread (channel_name, thread_name, topic, nsfw, age_restricted). topic = tujuan resmi channel; gunakan menilai kesesuaian pesan.\n` +
|
||||
`- <conversation_context>: obrolan SEBELUM target. Baris pertama "[conversation_flow] status=... context_msgs=... dropped=..." = metadata sistem (ongoing/sparse/cold_start), BUKAN pesan dinilai. Baris "[context] id=... time=... user=...: isi" = konteks, BUKAN target.\n` +
|
||||
`- <user_reputation trust_score total_infractions clean_streak last_offense_days_ago repeat_offender>: histori moderasi (repeat_offender=true = pelanggaran ≤7 hari). <user_history>: kutipan pesan pernah di-flag — cari POLA berulang (spam link sama), BUKAN bukti pesan bersih.\n` +
|
||||
`- Tidak ada data profil/reputasi per-user di context — nilai tiap pesan murni dari isinya + <conversation_context> + <web_searches> + <location_context>.\n` +
|
||||
`- <web_searches>/<web_content>: bukti web (prioritas tertinggi). <term_glossary>: definisi kata/slang/jargon (SearXNG) — pakai pahami kata asing, JANGAN tebak arti.\n` +
|
||||
`- <messages_to_analyze>: pesan TARGET yang WAJIB dinilai. Atribut <message>: id, user, time (ISO), repetitions (N = teks sama muncul N× di batch → sinyal spam), bot (true = bot), edited (true = hasil edit setelah posting → evasi potensial).`,
|
||||
);
|
||||
|
||||
@@ -20,7 +20,6 @@ import { analyzeSingleMediaImage } from "./mediaAnalysisClient.js";
|
||||
import {
|
||||
buildReferenceXml,
|
||||
escapeXml,
|
||||
formatReputationAttrs,
|
||||
getAnalysisContent,
|
||||
resolveDisplayName,
|
||||
resolveIsBot,
|
||||
@@ -37,7 +36,6 @@ import {
|
||||
import { buildTermGlossaryBlock } from "./termGlossary.js";
|
||||
import { getRecentCorrectedModerations } from "./textCacheStore.js";
|
||||
import { extractUrlsFromText, fetchUrlSafely } from "./urlFetcher.js";
|
||||
import { initializeUserReputation } from "./userReputationStore.js";
|
||||
import type { MessageImagePart } from "./visionAnalyzer.js";
|
||||
|
||||
const log = createChildLogger("textBatchProcessor");
|
||||
@@ -201,24 +199,10 @@ export async function runTextOnlyBatch(
|
||||
const batch = subBatches[i];
|
||||
const targetIds = batch.map((t) => t.id);
|
||||
|
||||
// ── Per-user reputation context (fetched ONCE per unique user, in
|
||||
// parallel). Personal profile descriptions are intentionally NOT
|
||||
// injected — they bloat the prompt (less room per request) and add a
|
||||
// per-user DB/Redis round-trip for little moderation signal. Only the
|
||||
// behavioural <user_reputation> history is sent. ─────────────────────
|
||||
const uniqueUserIds = [...new Set(batch.map((m) => m.user_id))];
|
||||
const batchGuildId = batch[0]?.guild_id ?? "";
|
||||
const userFetches = await Promise.all(
|
||||
uniqueUserIds.map(async (uid) => {
|
||||
const rep = await initializeUserReputation(uid, batchGuildId);
|
||||
return { uid, rep };
|
||||
}),
|
||||
);
|
||||
const userContexts = new Map<string, string>();
|
||||
for (const { uid, rep } of userFetches) {
|
||||
const repAttrs = formatReputationAttrs(rep);
|
||||
userContexts.set(uid, `<user_reputation ${repAttrs}/>`);
|
||||
}
|
||||
// No per-user reputation/profile context is injected into the prompt —
|
||||
// the user asked to keep the AI analysis context minimal (raw messages
|
||||
// only). Trust/infraction state is still tracked in the DB for
|
||||
// enforcement, just not shown to the LLM.
|
||||
|
||||
// ── URL images → multimodal vision evidence ─────────────────────────
|
||||
// The text batch fetches inline URLs; whenever one resolved to an image
|
||||
@@ -315,12 +299,11 @@ export async function runTextOnlyBatch(
|
||||
const mediaEvidenceCtx = (batchImageEvidence.get(msg.id) ?? [])
|
||||
.map((line) => `\n${line}`)
|
||||
.join("");
|
||||
const userCtx = userContexts.get(msg.user_id) ?? "";
|
||||
const refXml = await buildReferenceXml(msg);
|
||||
const repetitionCount = groupMapping.get(msg.id)?.length ?? 1;
|
||||
const isBot = resolveIsBot(msg);
|
||||
const isEdited = resolveIsEdited(msg);
|
||||
return `<message id="${escapeXml(msg.id)}" user="${escapeXml(resolveDisplayName(msg))}" time="${new Date(msg.created_at).toISOString()}"${repetitionCount > 1 ? ` repetitions="${repetitionCount}"` : ""}${isBot ? ` bot="true"` : ""}${isEdited ? ` edited="true"` : ""}>\n ${userCtx}${refXml ? `\n ${refXml}` : ""}\n <content>${escapeXml(content)}</content>${webContext}${mediaEvidenceCtx}\n</message>`;
|
||||
return `<message id="${escapeXml(msg.id)}" user="${escapeXml(resolveDisplayName(msg))}" time="${new Date(msg.created_at).toISOString()}"${repetitionCount > 1 ? ` repetitions="${repetitionCount}"` : ""}${isBot ? ` bot="true"` : ""}${isEdited ? ` edited="true"` : ""}>\n ${refXml ? `\n ${refXml}` : ""}\n <content>${escapeXml(content)}</content>${webContext}${mediaEvidenceCtx}\n</message>`;
|
||||
}),
|
||||
)
|
||||
).join("\n");
|
||||
|
||||
@@ -64,7 +64,6 @@ import {
|
||||
import {
|
||||
buildReferenceXml,
|
||||
escapeXml,
|
||||
formatReputationAttrs,
|
||||
getAnalysisContent,
|
||||
resolveDisplayName,
|
||||
resolveIsBot,
|
||||
@@ -84,7 +83,6 @@ import {
|
||||
} from "./searxngSearch.js";
|
||||
import { buildTermGlossaryBlock } from "./termGlossary.js";
|
||||
import { extractUrlsFromText } from "./urlFetcher.js";
|
||||
import { initializeUserReputation } from "./userReputationStore.js";
|
||||
|
||||
// ---------------------------------------------------------------------------
|
||||
// Types
|
||||
@@ -398,15 +396,13 @@ export async function prepareMediaMessage(
|
||||
.filter(Boolean)
|
||||
.join(" ");
|
||||
|
||||
const rep = await initializeUserReputation(target.user_id, target.guild_id);
|
||||
const refXml = await buildReferenceXml(target);
|
||||
|
||||
// Only the behavioural <user_reputation> history is injected; personal profile descriptions are omitted (see textBatchProcessor).
|
||||
const repAttrs = formatReputationAttrs(rep);
|
||||
const repXml = `<user_reputation ${repAttrs}/>`;
|
||||
|
||||
// No per-user reputation/profile context is injected into the prompt —
|
||||
// keep the AI analysis context minimal (raw messages only). Trust state is
|
||||
// still tracked in the DB for enforcement, just not shown to the LLM.
|
||||
const isBot = resolveIsBot(target);
|
||||
const isEdited = resolveIsEdited(target);
|
||||
const messageBlock = `<message id="${escapeXml(target.id)}" user="${escapeXml(resolveDisplayName(target))}" time="${new Date(target.created_at).toISOString()}"${isBot ? ` bot="true"` : ""}${isEdited ? ` edited="true"` : ""}>\n ${repXml}${refXml ? `\n ${refXml}` : ""}\n <content>${escapeXml(truncateForAi(content))}</content>${mediaContext ? ` ${escapeXml(mediaContext)}` : ""}${webContext}${mediaAnalysisContext}${searxngXml}${glossaryCtx}\n</message>`;
|
||||
const messageBlock = `<message id="${escapeXml(target.id)}" user="${escapeXml(resolveDisplayName(target))}" time="${new Date(target.created_at).toISOString()}"${isBot ? ` bot="true"` : ""}${isEdited ? ` edited="true"` : ""}>\n ${refXml ? `\n ${refXml}` : ""}\n <content>${escapeXml(truncateForAi(content))}</content>${mediaContext ? ` ${escapeXml(mediaContext)}` : ""}${webContext}${mediaAnalysisContext}${searxngXml}${glossaryCtx}\n</message>`;
|
||||
return { targetId, messageBlock };
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user