From aad62bc66befb5eb7a334c54b8651d8af4e7ac55 Mon Sep 17 00:00:00 2001
From: MythEclipse <asepharyanalm001@gmail.com>
Date: Wed, 3 Jun 2026 20:59:10 +0700
Subject: [PATCH] fix(ai-moderation): strip fallback text from LLM analysis
 content

---
 .../ai-moderation/llmModerationClient.ts      | 23 +++++++++++++++++--
 1 file changed, 21 insertions(+), 2 deletions(-)

diff --git a/services/discord-gateway/src/modules/ai-moderation/llmModerationClient.ts b/services/discord-gateway/src/modules/ai-moderation/llmModerationClient.ts
index 795233f..54fc7e5 100644
--- a/services/discord-gateway/src/modules/ai-moderation/llmModerationClient.ts
+++ b/services/discord-gateway/src/modules/ai-moderation/llmModerationClient.ts
@@ -476,6 +476,25 @@ type MessageImagePart = {
   customEmojiName?: string;
 };
 
+// ---------------------------------------------------------------------------
+// Content helpers
+// ---------------------------------------------------------------------------
+
+/**
+ * Returns the real text content for AI analysis, stripping fallback text
+ * that getDisplayContent() synthesized ("[Attachment: ...]", "[Sticker: ...]",
+ * "[Embed]").  These filenames alone are meaningless to the LLM and can
+ * falsely inflate a "clean" verdict when the actual image failed to download.
+ */
+function getAnalysisContent(message: MessageRecord): string {
+  const raw = message.edited_content ?? message.content;
+  const stripped = raw.replace(
+    /\[(?:Attachment|Sticker):[^\]]*\]|\[Embed\]/g,
+    "",
+  );
+  return stripped.trim();
+}
+
 // ---------------------------------------------------------------------------
 // Media detection helper
 // ---------------------------------------------------------------------------
@@ -976,7 +995,7 @@ async function runTextOnlyBatch(
 
       const messagesBlock = batch
         .map((msg) => {
-          const content = msg.edited_content ?? msg.content;
+          const content = getAnalysisContent(msg);
           const textEvidence = textEvidenceMap.get(msg.id) ?? "";
           const textContext = textEvidence ? `\n${textEvidence}` : "";
           // XML delimiters wrap each message for prompt safety (R1)
@@ -1129,7 +1148,7 @@ async function _runSingleMediaAnalysis(
     att.uploaded_url ?? null;
 
   const maxDimension = config.AI_LLM_IMAGE_MAX_DIMENSION ?? 1024;
-  const content = target.edited_content ?? target.content;
+  const content = getAnalysisContent(target);
 
   // ── 1-3. Parallel download of ALL media sources ──
   // Build all download promises upfront and execute them in one Promise.all.