From 9ae26b8ec92185db37a26e6d66dfeb6d8e434e31 Mon Sep 17 00:00:00 2001
From: asepharyana <superaseph@gmail.com>
Date: Sat, 15 Aug 2026 21:05:06 +0700
Subject: [PATCH] refactor(llm): unify vision routing with text moderation and
 remove dedicated endpoint

---
 .../src/modules/ai-moderation/llmClient.ts    | 82 ++-----------------
 .../src/shared/config/index.ts                |  9 +-
 2 files changed, 8 insertions(+), 83 deletions(-)

diff --git a/services/discord-gateway/src/modules/ai-moderation/llmClient.ts b/services/discord-gateway/src/modules/ai-moderation/llmClient.ts
index 3eb064e..e07b5e2 100644
--- a/services/discord-gateway/src/modules/ai-moderation/llmClient.ts
+++ b/services/discord-gateway/src/modules/ai-moderation/llmClient.ts
@@ -325,14 +325,11 @@ export async function llmChat(
  * Convenience for vision (image/sticker/emoji) analysis.
  * Returns the raw completion content (trimmed) or null.
  *
- * When AI_LLM_VISION_BASE_URL + AI_LLM_VISION_API_KEY are configured, vision is
- * sent DIRECTLY to a dedicated multimodal endpoint (e.g. NVIDIA direct API),
- * separate from the text/moderation router (omniroute/9router). This keeps
- * image analysis on a vision-capable model while text moderation stays on the
- * router's text combo.
- *
- * Otherwise it falls back to the shared AI_LLM_BASE_URL (router combo) using
- * AI_LLM_VISION_MODEL.
+ * Vision routes through the SAME router/base URL as text moderation
+ * (AI_LLM_BASE_URL) — the dedicated NVIDIA multimodal endpoint was removed.
+ * It uses AI_LLM_VISION_MODEL (a different model alias from the text combo)
+ * so image analysis stays on a vision-capable model. Thinking-disable from
+ * config.AI_LLM_DISABLE_THINKING applies automatically via buildLlmParams.
  *
  * NOTE: retries are disabled here on purpose — visionAnalyzer.ts already
  * wraps this call in its own 3-attempt loop with exponential backoff.
@@ -342,12 +339,6 @@ export async function llmVision(
   promptText: string,
   imageUrl: { url: string },
 ): Promise<string | null> {
-  // ── Dedicated vision endpoint (NVIDIA direct, etc.) ──────────────────────
-  if (config.AI_LLM_VISION_BASE_URL && config.AI_LLM_VISION_API_KEY) {
-    return llmVisionDirect(promptText, imageUrl);
-  }
-
-  // ── Fallback: shared router combo ─────────────────────────────────────────
   const completion = await llmChat({
     messages: [
       {
@@ -369,66 +360,3 @@ export async function llmVision(
   if (!completion) return null;
   return completion.choices[0]?.message?.content?.trim() ?? null;
 }
-
-import axios from "axios";
-
-/**
- * Direct vision call to a dedicated multimodal endpoint (NVIDIA integrate API).
- * Model is fixed to the vision-capable one configured via AI_LLM_VISION_MODEL
- * (default nvidia/nemotron-3-nano-omni-30b-a3b-reasoning). Uses reasoning_budget
- * + non-streaming (axios JSON) — NVIDIA direct does not need the SSE streaming
- * gymnastics the composite routers require.
- */
-async function llmVisionDirect(
-  promptText: string,
-  imageUrl: { url: string },
-): Promise<string | null> {
-  const visionModel =
-    config.AI_LLM_VISION_MODEL ||
-    "nvidia/nemotron-3-nano-omni-30b-a3b-reasoning";
-  try {
-    const response = await axios.post(
-      `${config.AI_LLM_VISION_BASE_URL}/chat/completions`,
-      {
-        messages: [
-          {
-            role: "user",
-            content: [
-              { type: "text", text: promptText },
-              { type: "image_url", image_url: imageUrl },
-            ],
-          },
-        ],
-        model: visionModel,
-        max_tokens: 65536,
-        reasoning_budget: config.AI_LLM_DISABLE_THINKING ? 0 : 16384,
-        stream: false,
-        temperature: 0.6,
-        top_p: 0.95,
-      },
-      {
-        headers: {
-          Authorization: `Bearer ${config.AI_LLM_VISION_API_KEY}`,
-          Accept: "application/json",
-          "Content-Type": "application/json",
-        },
-        timeout: 120_000,
-      },
-    );
-
-    const content: string | undefined =
-      response.data?.choices?.[0]?.message?.content;
-    if (!content) return null;
-    return content.trim();
-  } catch (err: any) {
-    const status = err?.response?.status ?? "n/a";
-    const detail = err?.response?.data
-      ? JSON.stringify(err.response.data).slice(0, 300)
-      : err?.message;
-    log.error(
-      { status, detail, model: visionModel },
-      "Direct vision API call failed",
-    );
-    return null;
-  }
-}
diff --git a/services/discord-gateway/src/shared/config/index.ts b/services/discord-gateway/src/shared/config/index.ts
index 19c1bac..b871e01 100644
--- a/services/discord-gateway/src/shared/config/index.ts
+++ b/services/discord-gateway/src/shared/config/index.ts
@@ -144,13 +144,10 @@ export const configSchema = z
       .url()
       .default("https://9router.asepharyana.my.id/v1"),
     AI_LLM_MODEL: z.string().default("text"),
+    // Vision uses the SAME router/base URL as text moderation
+    // (AI_LLM_BASE_URL) but a different model alias. The dedicated NVIDIA
+    // multimodal endpoint was removed.
     AI_LLM_VISION_MODEL: z.string().default("multimodal"),
-    // Vision can be routed to a dedicated endpoint (e.g. NVIDIA direct) that is
-    // separate from the text/moderation router. When both are set, llmVision()
-    // calls the dedicated vision endpoint directly; otherwise it falls back to
-    // the shared AI_LLM_BASE_URL with AI_LLM_VISION_MODEL.
-    AI_LLM_VISION_BASE_URL: z.string().url().optional(),
-    AI_LLM_VISION_API_KEY: z.string().optional(),
     AI_LLM_DISABLE_THINKING: z
       .string()
       .default("true")