perf(discord-gateway): start AI moderation analysis immediately, running attachment uploads in parallel instead of blocking

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

@@ -239,6 +239,7 @@ function getAnalysisWorkerUrl(): URL {
 const workerPool = new Piscina({
   filename: fileURLToPath(getAnalysisWorkerUrl()),
   execArgv: process.execArgv,
+  maxThreads: config.PISCINA_MAX_THREADS ?? availableParallelism(),
 });
 
 interface AnalysisWorkerResponse {
@@ -275,8 +276,8 @@ export function pickBatchWithinBudget(
 
   for (const msg of messages) {
     const content = msg.edited_content ?? msg.content;
-    // Rough token estimate: ~3 chars per token + metadata overhead
-    const msgTokens = Math.ceil(content.length / 3) + tokensPerMessage;
+    // Accurate token count via tiktoken (+ overhead for JSON structure)
+    const msgTokens = estimateTokens(content) + tokensPerMessage;
 
     if (usedTokens + msgTokens <= maxTokens) {
       batch.push(msg);

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

@@ -26,6 +26,55 @@ interface BadwordCacheEntry {
 const badwordCache = new Map<string, BadwordCacheEntry>();
 const inFlightBadwordLookups = new Map<string, Promise<string[]>>();
 
+// ---------------------------------------------------------------------------
+// Local rule-based pre-filter — short-circuits definitively safe messages
+// without calling the LLM badword detection API.
+// Conservative: only flags messages that are 100% certainly clean.
+// ---------------------------------------------------------------------------
+
+const SAFE_PATTERNS: Array<{ test: (text: string) => boolean; reason: string }> = [
+  {
+    // Pure laughter patterns
+    test: (t) => /^(wkwk+|w+kw+k+|wkwkw+|haha+|hehe+|hihi+|huhu+|xixi+|wakak+|awkwa+)$/i.test(t),
+    reason: "laughter pattern",
+  },
+  {
+    // Single-word affirmatives (common in Indonesian Discord)
+    test: (t) => /^(ok|oke|okay|sip|siap|aman|mantap|gas|gass|gaskeun|santuy|gaskan|lah|wih|wah|eh|nah|loh|hmm|hm|heh)$/i.test(t),
+    reason: "single-word affirmative",
+  },
+  {
+    // Greetings / common short expressions
+    test: (t) => /^(hai|halo|hello|hi|oi|woy|woi|pagi|siang|sore|malam|mlm|p|w|L|F|gws|thx|thks|makasih|ty|thanks|yw|sama-sama|ok sip|ok bang|siap bang)$/i.test(t),
+    reason: "greeting/common expression",
+  },
+  {
+    // Very short messages (1-2 characters) — reactions, single letters
+    test: (t) => t.length <= 2,
+    reason: "very short message (1-2 chars)",
+  },
+  {
+    // Pure numeric or pure punctuation
+    test: (t) => /^[\d\s.,!?;:'"()\-_]+$/.test(t),
+    reason: "numeric/punctuation only",
+  },
+];
+
+/**
+ * Checks whether a text message is definitively safe and does not need
+ * LLM-based badword detection.  This is a conservative local pre-filter
+ * — it only returns true for patterns that CANNOT be violations.
+ */
+export function isDefinitivelySafe(text: string): boolean {
+  // Normalize: strip Discord custom emoji, trim whitespace
+  const { text: normalized } = normalizeDiscordCustomEmoji(text);
+  const trimmed = normalized.trim();
+
+  if (trimmed.length === 0) return true;
+
+  return SAFE_PATTERNS.some((p) => p.test(trimmed));
+}
+
 export interface ModerationTextEvidence {
   raw: string;
   normalized: string;
@@ -123,6 +172,11 @@ export async function detectIndonesianBadwords(
     return cached;
   }
 
+  // ── Tier 0: Local rule-based pre-filter (fastest — no API call) ──
+  if (isDefinitivelySafe(text)) {
+    return [];
+  }
+
   // De-duplicate concurrent lookups
   const inFlight = inFlightBadwordLookups.get(cacheKey);
   if (inFlight) {

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

@@ -1,5 +1,6 @@
 import { createChildLogger } from "@bete/shared/logger";
 import { retryWithBackoff } from "@bete/shared/utils";
+import { LRUCache } from "lru-cache";
 import type { ChatCompletion } from "openai/resources/chat/completions";
 import { AbortError } from "p-retry";
 import { z } from "zod";
@@ -28,11 +29,14 @@ import {
   buildStickerVisionPrompt,
 } from "./stickerPrompt.js";
 import {
+  computeImagePhash,
   getCachedMediaAnalysis,
+  getCachedMediaByPhash,
   makeCustomEmojiCacheKey,
   makeImageCacheKey,
   makeStickerCacheKey,
   upsertCachedMediaAnalysis,
+  upsertCachedMediaByPhash,
 } from "./textCacheStore.js";
 import { extractUrlsFromText, fetchUrlSafely } from "./urlFetcher.js";
 

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

@@ -184,9 +184,15 @@ CRITICAL:
 // Composer: assembles all sections with XML delimiters
 // ---------------------------------------------------------------------------
 
+/** Prompt mode — determines which few-shot example section is included. */
+export type PromptMode = "text" | "media" | "mixed";
+
 export interface BuildSystemPromptOptions {
   contextText: string;
-  includeMediaInstructions: boolean;
+  /** Prompt mode — determines which sections are included. */
+  mode: PromptMode;
+  /** @deprecated Use `mode` instead. */
+  includeMediaInstructions?: boolean;
   correction?: { error: string; preview: string };
 }