From 172edf402af21fc2e6846a6bdfd27696da48ebe2 Mon Sep 17 00:00:00 2001
From: suguanYang <wangbinqi77@gmail.com>
Date: Tue, 25 Aug 2026 19:00:41 +0800
Subject: [PATCH] perf: skip VLM thinking on image inspection
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Inspect is OCR and provenance boxes, not multi-step reasoning. Qwen 3.5 Flash thinks by default and that was 20–40s; turning thinking off dropped it to ~8–11s without a clear quality loss.

Co-authored-by: Cursor <cursoragent@cursor.com>
---
 .../chat/image-inspection-model.test.ts       | 28 +++++++++++++++++++
 src/domains/chat/image-inspection-model.ts    | 19 +++++++++++++
 2 files changed, 47 insertions(+)

diff --git a/src/domains/chat/image-inspection-model.test.ts b/src/domains/chat/image-inspection-model.test.ts
index 6181c6d..d5815d3 100644
--- a/src/domains/chat/image-inspection-model.test.ts
+++ b/src/domains/chat/image-inspection-model.test.ts
@@ -112,6 +112,20 @@ describe("image inspection model", () => {
       regions: [{ x: 0.4, y: 0.5, w: 0.2, h: 0.1 }],
     })
     expect(result.source).toBe("structured")
+    expect(generateObject).toHaveBeenCalledWith(
+      expect.objectContaining({
+        providerOptions: {
+          google: {
+            thinkingConfig: {
+              thinkingLevel: "minimal",
+            },
+          },
+          alibaba: {
+            enableThinking: false,
+          },
+        },
+      }),
+    )
   })
 
   it("salvages valid structured JSON from failed generateObject text", async () => {
@@ -230,6 +244,20 @@ describe("image inspection model", () => {
       source: "structured_text",
     })
     expect(generateText).toHaveBeenCalledTimes(1)
+    expect(generateText).toHaveBeenCalledWith(
+      expect.objectContaining({
+        providerOptions: {
+          google: {
+            thinkingConfig: {
+              thinkingLevel: "minimal",
+            },
+          },
+          alibaba: {
+            enableThinking: false,
+          },
+        },
+      }),
+    )
   })
 
   it("falls back to analysis-only text when unsupported structured models cannot emit JSON", async () => {
diff --git a/src/domains/chat/image-inspection-model.ts b/src/domains/chat/image-inspection-model.ts
index 8a498af..237d108 100644
--- a/src/domains/chat/image-inspection-model.ts
+++ b/src/domains/chat/image-inspection-model.ts
@@ -14,6 +14,22 @@ import { logger } from "@/lib/logger"
 const VISION_MODEL = process.env.VISION_MODEL ?? CHAT_MODEL
 const IMAGE_INSPECTION_BATCH_SIZE = 6
 
+/**
+ * Inspect is OCR + provenance boxes, not multi-step reasoning.
+ * Gemini 3 Flash cannot fully disable thinking; `minimal` is the lowest.
+ * Qwen 3.5 Flash thinks by default on the gateway; turn that off.
+ */
+const IMAGE_INSPECTION_PROVIDER_OPTIONS = {
+  google: {
+    thinkingConfig: {
+      thinkingLevel: "minimal",
+    },
+  },
+  alibaba: {
+    enableThinking: false,
+  },
+}
+
 export const imageInspectionResultSchema = z.object({
   analysis: z.string(),
   pages: z
@@ -105,6 +121,7 @@ async function generateImageInspectionBatchResult(input: {
     const response = await generateObject({
       model: VISION_MODEL,
       schema: imageInspectionResultSchema,
+      providerOptions: IMAGE_INSPECTION_PROVIDER_OPTIONS,
       messages: [
         {
           role: "user",
@@ -170,6 +187,7 @@ async function generateImageInspectionBatchResult(input: {
   try {
     const jsonTextResponse = await generateText({
       model: VISION_MODEL,
+      providerOptions: IMAGE_INSPECTION_PROVIDER_OPTIONS,
       messages: [
         {
           role: "user",
@@ -223,6 +241,7 @@ async function generateImageInspectionBatchResult(input: {
   try {
     const response = await generateText({
       model: VISION_MODEL,
+      providerOptions: IMAGE_INSPECTION_PROVIDER_OPTIONS,
       messages: [
         {
           role: "user",