perf(ai-moderation): drop personal user-profile descriptions from context

User insight: personal profile summaries 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 kept.

- textBatchProcessor: stop fetching getUserProfile; remove <user_profiles>
  block + <user_profile_ref> from message tags. Keep <user_reputation>.
- mediaBatchProcessor + visionAnalyzer: same removal (profile fetch + ref).
- prompts/system.ts + prompts/output.ts: drop stale <user_profiles>/
  <user_profile_ref> instructions; point LLM at <user_reputation> instead.
- aiAnalyzer: gate userProfileLearner behind AI_USER_PROFILE_LEARNING_ENABLED
  (default false) — generates profiles nobody reads, pure LLM/DB waste.
- Add AI_USER_PROFILE_LEARNING_ENABLED config knob.

Net: smaller prompts (more messages fit per request), fewer DB round-trips
per sub-batch, and no background LLM calls learning unused profiles.

tsc, biome, vitest (129) all clean.

services/discord-gateway/src/modules/ai-moderation/aiAnalyzer.ts

@@ -136,9 +136,11 @@ export function startPendingAIAnalysisWorker(
   import("./cultureLearner.js")
     .then((m) => m.startCultureLearnerWorker())
     .catch(console.error);
+  if (config.AI_USER_PROFILE_LEARNING_ENABLED) {
     import("./userProfileLearner.js")
       .then((m) => m.startUserProfileLearnerWorker())
       .catch(console.error);
+  }
 
   setInterval(() => {
     // [D] Periodic cache hygiene: purge expired moderation verdicts from

services/discord-gateway/src/modules/ai-moderation/mediaBatchProcessor.ts

@@ -16,10 +16,8 @@ import { getChannelCulture } from "./channelCultureStore.js";
 import type { RetryState } from "./llmCaller.js";
 import { callModerationLLM } from "./llmCaller.js";
 import { prepareMediaMessage } from "./mediaAnalysisClient.js";
-import { buildUserProfilesBlock } from "./moderationBuilders.js";
 import { buildSystemPrompt as buildSystemPromptModular } from "./moderationPrompt.js";
 import { buildCorrectedFewShotExamples } from "./textBatchProcessor.js";
-import { getUserProfile } from "./userProfileStore.js";
 
 const log = createChildLogger("mediaBatchProcessor");
 
@@ -65,32 +63,15 @@ export async function runMediaBatch(
     channelCulture,
   });
 
-  // Gather user profiles ONCE for the whole batch and emit a deduplicated
+  // Per-message blocks (from prepareMediaMessage) carry their own
-  // <user_profiles> map (with last-generated timestamp); per-message blocks
+  // <user_reputation> history; personal profile descriptions are omitted
-  // (from prepareMediaMessage) reference it via <user_profile_ref>.
+  // (they bloat the prompt and add a per-user DB round-trip for little
-  const profileByUser = new Map<
+  // moderation signal — see textBatchProcessor).
-    string,
-    {
-      text: string;
-      asOf?: number | null;
-    }
-  >();
-  for (const t of targets) {
-    if (profileByUser.has(t.user_id)) continue;
-    const profile = await getUserProfile(t.user_id);
-    profileByUser.set(t.user_id, {
-      text: profile?.profile_summary ?? "",
-      asOf: profile?.last_analyzed_at ?? null,
-    });
-  }
-  const userProfilesBlock = buildUserProfilesBlock(profileByUser);
-
   const messagesBlock = prepared.map((p) => p.messageBlock).join("\n");
   // Data/instruction separation: the system prompt is stable per mode — all
-  // per-batch context (profiles, conversation) lives in the USER payload,
+  // per-batch context (conversation) lives in the USER payload, ordered
-  // ordered oldest-first so targets come last.
+  // oldest-first so targets come last.
   const userBlocks = [
-    userProfilesBlock?.trimEnd() ?? "",
     contextBlock?.trimEnd() ?? "",
     `<messages_to_analyze>\n${messagesBlock}\n</messages_to_analyze>`,
   ].filter((b) => b.trim().length > 0);

services/discord-gateway/src/modules/ai-moderation/prompts/output.ts

@@ -35,8 +35,8 @@ 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 — Profil Pengguna dan Kultur Channel
+## PERSONALITY & MEMORI — Reputasi Pengguna dan Kultur Channel
-Data konteks tersedia: <user_profiles> (peta ringkasan kepribadian, di pesan USER), <user_reputation> (skor trust), dan <channel_culture> (topik/vibe channel). Setiap <message> dapat memuat <user_profile_ref user_id=".../> yang menunjuk ke entri di peta <user_profiles>.
+Data konteks tersedia: <user_reputation> (skor trust + histori infraction + repeat_offender), dan <channel_culture> (topik/vibe channel).
 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.
 - Perubahan perilaku mencolok (mis. teknis tiba-tiba provokatif) layak dicatat. JANGAN paksa referensi profil jika tidak relevan — analysis natural lebih baik.
@@ -70,8 +70,8 @@ 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_profile_ref>/<user_profiles> untuk personalisasi analysis — jadikan analysis terasa seperti sistem "mengenal" pengguna.
+- 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 profil yang diketahui, CATAT dalam analysis sebagai informasi kontekstual yang relevan.
+- Jika perilaku pesan menyimpang dari pola histori yang diketahui, CATAT dalam analysis sebagai informasi kontekstual yang relevan.
 - JANGAN paksa referensi profil jika tidak relevan — analysis natural lebih baik dari yang dipaksakan.`;
 
 // ---------------------------------------------------------------------------

services/discord-gateway/src/modules/ai-moderation/prompts/system.ts

@@ -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_profiles>/<user_profile_ref>: peta ringkasan kepribadian per user_id (as_of = kapan dibuat; profil lama mungkin usang). <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` +
+      `- <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` +
       `- <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).`,
   );

services/discord-gateway/src/modules/ai-moderation/textBatchProcessor.ts

@@ -19,8 +19,6 @@ import { callModerationLLM } from "./llmCaller.js";
 import { analyzeSingleMediaImage } from "./mediaAnalysisClient.js";
 import {
   buildReferenceXml,
-  buildUserProfileRef,
-  buildUserProfilesBlock,
   escapeXml,
   formatReputationAttrs,
   getAnalysisContent,
@@ -39,7 +37,6 @@ import {
 import { buildTermGlossaryBlock } from "./termGlossary.js";
 import { getRecentCorrectedModerations } from "./textCacheStore.js";
 import { extractUrlsFromText, fetchUrlSafely } from "./urlFetcher.js";
-import { getUserProfile } from "./userProfileStore.js";
 import { initializeUserReputation } from "./userReputationStore.js";
 import type { MessageImagePart } from "./visionAnalyzer.js";
 
@@ -204,40 +201,24 @@ export async function runTextOnlyBatch(
     const batch = subBatches[i];
     const targetIds = batch.map((t) => t.id);
 
-    // User reputation + profiles (raw summary text — deduplicated into a
+    // ── Per-user reputation context (fetched ONCE per unique user, in
-    // single <user_profiles> map per batch; messages only reference it).
+    //    parallel). Personal profile descriptions are intentionally NOT
-    const userContexts = new Map<string, string>();
+    //    injected — they bloat the prompt (less room per request) and add a
-    const userProfiles = new Map<
+    //    per-user DB/Redis round-trip for little moderation signal. Only the
-      string,
+    //    behavioural <user_reputation> history is sent. ─────────────────────
-      {
-        text: string;
-        asOf?: number | null;
-      }
-    >();
-    // ── Per-user reputation + profile context (fetched ONCE per unique user,
-    //    in parallel — was a serial per-message loop that cost ~2N sequential
-    //    DB/Redis round-trips per sub-batch and dominated latency on small
-    //    batches). ─────────────────────────────────────────────────────────
     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, profile] = await Promise.all([
+        const rep = await initializeUserReputation(uid, batchGuildId);
-          initializeUserReputation(uid, batchGuildId),
+        return { uid, rep };
-          getUserProfile(uid),
-        ]);
-        return { uid, rep, profile };
       }),
     );
-    for (const { uid, rep, profile } of userFetches) {
+    const userContexts = new Map<string, string>();
+    for (const { uid, rep } of userFetches) {
       const repAttrs = formatReputationAttrs(rep);
       userContexts.set(uid, `<user_reputation ${repAttrs}/>`);
-      userProfiles.set(uid, {
-        text: profile?.profile_summary ?? "",
-        asOf: profile?.last_analyzed_at ?? null,
-      });
     }
-    const userProfilesBlock = buildUserProfilesBlock(userProfiles);
 
     // ── URL images → multimodal vision evidence ─────────────────────────
     // The text batch fetches inline URLs; whenever one resolved to an image
@@ -335,16 +316,11 @@ export async function runTextOnlyBatch(
               .map((line) => `\n${line}`)
               .join("");
             const userCtx = userContexts.get(msg.user_id) ?? "";
-            const userProfileRef = (
-              userProfiles.get(msg.user_id)?.text ?? ""
-            ).trim()
-              ? buildUserProfileRef(msg.user_id)
-              : "";
             const refXml = await buildReferenceXml(msg);
             const repetitionCount = groupMapping.get(msg.id)?.length ?? 1;
             const isBot = resolveIsBot(msg);
             const isEdited = resolveIsEdited(msg);
-            return `<message id="${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  ${userCtx}${userProfileRef ? `\n  ${userProfileRef}` : ""}${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  ${userCtx}${refXml ? `\n  ${refXml}` : ""}\n  <content>${escapeXml(content)}</content>${webContext}${mediaEvidenceCtx}\n</message>`;
           }),
         )
       ).join("\n");
@@ -359,10 +335,10 @@ export async function runTextOnlyBatch(
               .join("\n")}\n</web_searches>`
           : "";
       // Data/instruction separation: the system prompt is stable per mode —
-      // all per-batch context (profiles, conversation, web evidence) lives in
+      // all per-batch context (conversation, web evidence) lives in the USER
-      // the USER payload, ordered oldest-first so targets come last.
+      // payload, ordered oldest-first so targets come last. Personal user
+      // profile descriptions are intentionally omitted (see above).
       const userBlocks = [
-        userProfilesBlock?.trimEnd() ?? "",
         contextBlock?.trimEnd() ?? "",
         searxngBlock,
         glossaryBlock,

services/discord-gateway/src/modules/ai-moderation/visionAnalyzer.ts

@@ -63,7 +63,6 @@ import {
 } from "./mediaDownloader.js";
 import {
   buildReferenceXml,
-  buildUserProfileRef,
   escapeXml,
   formatReputationAttrs,
   getAnalysisContent,
@@ -85,7 +84,6 @@ import {
 } from "./searxngSearch.js";
 import { buildTermGlossaryBlock } from "./termGlossary.js";
 import { extractUrlsFromText } from "./urlFetcher.js";
-import { getUserProfile } from "./userProfileStore.js";
 import { initializeUserReputation } from "./userReputationStore.js";
 
 // ---------------------------------------------------------------------------
@@ -401,21 +399,14 @@ export async function prepareMediaMessage(
     .join(" ");
 
   const rep = await initializeUserReputation(target.user_id, target.guild_id);
-  const profile = await getUserProfile(target.user_id);
   const refXml = await buildReferenceXml(target);
-  // Profile is emitted ONCE per batch in a <user_profiles> map (see
-  // mediaBatchProcessor); here we only reference it to avoid repeating the
-  // full summary on every message of the same user.
-  const profileRef = profile?.profile_summary?.trim()
-    ? buildUserProfileRef(target.user_id)
-    : "";
 
-  // Rich reputation — attrs only, no user history injection (per channel context preference)
+  // Only the behavioural <user_reputation> history is injected; personal profile descriptions are omitted (see textBatchProcessor).
   const repAttrs = formatReputationAttrs(rep);
   const repXml = `<user_reputation ${repAttrs}/>`;
 
   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  ${repXml}${profileRef ? `\n  ${profileRef}` : ""}${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  ${repXml}${refXml ? `\n  ${refXml}` : ""}\n  <content>${escapeXml(truncateForAi(content))}</content>${mediaContext ? ` ${escapeXml(mediaContext)}` : ""}${webContext}${mediaAnalysisContext}${searxngXml}${glossaryCtx}\n</message>`;
   return { targetId, messageBlock };
 }

services/discord-gateway/src/shared/config/index.ts

@@ -217,6 +217,14 @@ export const configSchema = z
       .default(true),
     // Max glossary terms looked up per analysis batch (keeps latency bounded).
     AI_GLOSSARY_MAX_TERMS: z.coerce.number().int().min(1).max(20).default(6),
+    // Per-user personal profile summaries (userProfileLearner). Disabled by
+    // default: profiles bloat the analysis context and add LLM/DB cost for
+    // little moderation signal — only <user_reputation> history is injected.
+    AI_USER_PROFILE_LEARNING_ENABLED: z
+      .string()
+      .optional()
+      .transform((v) => v === "true")
+      .default(false),
     // Min word length for a term to be considered glossary-worthy.
     AI_GLOSSARY_MIN_WORD_LENGTH: z.coerce
       .number()