feat: hardcode canvas vision to qwen3-vl and downscale pick-object input

app/api/canvas/auto-title/route.ts

@@ -1,6 +1,6 @@
 import { NextRequest, NextResponse } from "next/server";
 import { auth } from "@/app/(auth)/auth";
-import { callQwen25Vl } from "@/lib/ai/qwen25-vl";
+import { callQwen3Vl } from "@/lib/ai/qwen3-vl";
 
 export async function POST(request: NextRequest) {
   console.log("[Auto-title] API called");
@@ -19,12 +19,12 @@ export async function POST(request: NextRequest) {
       return NextResponse.json({ error: "Image URL is required" }, { status: 400 });
     }
 
-    console.log("[Auto-title] 使用硬编码模型: qwen2.5-vl");
+    console.log("[Auto-title] 使用硬编码模型: qwen3-vl");
 
     // 构造 Prompt
     const prompt = "请用2-4个字概括这张图片的主体内容，作为文件标题。直接返回标题，不要包含任何标点符号或其他文字。";
 
-    const { text, model } = await callQwen25Vl({
+    const { text, model } = await callQwen3Vl({
       imageUrl,
       prompt,
       temperature: 0.2,

app/api/canvas/text-detect/route.ts

@@ -1,9 +1,7 @@
 import { NextResponse } from "next/server";
 import { z } from "zod";
-import { generateText } from "ai";
 import { auth } from "@/app/(auth)/auth";
-import { getEnabledAiModelsWithProvider } from "@/lib/db/queries";
-import { getLanguageModelById } from "@/lib/ai/dynamic-provider";
+import { callQwen3Vl } from "@/lib/ai/qwen3-vl";
 
 export const runtime = "nodejs";
 export const maxDuration = 60;
@@ -18,26 +16,6 @@ interface OcrItem {
   confidence?: number;
 }
 
-function pickVisionModel(models: Awaited<ReturnType<typeof getEnabledAiModelsWithProvider>>) {
-  const normalized = models
-    .filter((item) => item.model.supportsVision)
-    .sort((a, b) => {
-      const aScore =
-        (a.model.modelId?.toLowerCase().includes("qwen3-vl") ? 100 : 0) +
-        (a.model.displayName?.toLowerCase().includes("qwen3-vl") ? 100 : 0) +
-        (a.provider.type === "alibaba" ? 20 : 0) +
-        (a.model.isDefault ? 5 : 0);
-      const bScore =
-        (b.model.modelId?.toLowerCase().includes("qwen3-vl") ? 100 : 0) +
-        (b.model.displayName?.toLowerCase().includes("qwen3-vl") ? 100 : 0) +
-        (b.provider.type === "alibaba" ? 20 : 0) +
-        (b.model.isDefault ? 5 : 0);
-      return bScore - aScore;
-    });
-
-  return normalized[0];
-}
-
 function parseJsonText(raw: string): any | null {
   if (!raw) return null;
 
@@ -124,18 +102,7 @@ export async function POST(request: Request) {
       return NextResponse.json({ success: false, error: "仅支持 URL 图片，不允许 base64 传输" }, { status: 400 });
     }
 
-    const models = await getEnabledAiModelsWithProvider();
-    const selected = pickVisionModel(models);
-
-    if (!selected) {
-      return NextResponse.json({ success: false, error: "未配置可用的视觉模型" }, { status: 500 });
-    }
-
-    const model = await getLanguageModelById(selected.model.id);
-    if (!model) {
-      return NextResponse.json({ success: false, error: "视觉模型加载失败" }, { status: 500 });
-    }
-
+    console.log("[Canvas Text Detect] 使用硬编码模型: qwen3-vl");
     const prompt = [
       "你是图片文字识别助手。",
       "请识别图片中可见文字并按阅读顺序输出 JSON。",
@@ -143,19 +110,11 @@ export async function POST(request: Request) {
       "格式：{\"items\":[{\"id\":\"1\",\"text\":\"示例\",\"confidence\":0.95}]}",
       "如果没有识别到文字，返回 {\"items\":[]}。",
     ].join("\n");
-
-    const { text } = await generateText({
-      model,
-      messages: [
-        {
-          role: "user",
-          content: [
-            { type: "text", text: prompt },
-            { type: "image", image: imageUrl },
-          ],
-        },
-      ],
+    const { text, model } = await callQwen3Vl({
+      imageUrl,
+      prompt,
       temperature: 0.1,
+      maxTokens: 800,
     });
 
     const parsedJson = parseJsonText(text || "");
@@ -173,7 +132,7 @@ export async function POST(request: Request) {
       success: true,
       data: {
         items,
-        model: selected.model.displayName,
+        model,
       },
     });
   } catch (error) {

app/api/pick-object/route.ts

@@ -1,6 +1,7 @@
 import { NextResponse } from "next/server";
 import { uploadFile } from "@/lib/storage/service";
-import { callQwen25Vl } from "@/lib/ai/qwen25-vl";
+import { callQwen3Vl } from "@/lib/ai/qwen3-vl";
+import sharp from "sharp";
 
 export const runtime = "nodejs";
 export const maxDuration = 30;
@@ -30,11 +31,53 @@ interface PickObjectResult {
   };
 }
 
-function inferFileExtFromMime(mimeType: string) {
-  if (mimeType.includes("png")) return "png";
-  if (mimeType.includes("webp")) return "webp";
-  if (mimeType.includes("gif")) return "gif";
-  return "jpg";
+const MAX_VISION_SIDE = 500;
+const PICK_OBJECT_FETCH_TIMEOUT_MS = 15000;
+
+async function downscaleImageForVision(buffer: Buffer) {
+  const source = sharp(buffer, { failOn: "none" });
+  const meta = await source.metadata();
+  const hasAlpha = Boolean(meta.hasAlpha);
+  const width = meta.width || 0;
+  const height = meta.height || 0;
+
+  const resized = source
+    .rotate()
+    .resize({
+      width: MAX_VISION_SIDE,
+      height: MAX_VISION_SIDE,
+      fit: "inside",
+      withoutEnlargement: true,
+    });
+
+  if (hasAlpha) {
+    return {
+      buffer: await resized.png({ compressionLevel: 9 }).toBuffer(),
+      ext: "png",
+      width,
+      height,
+    };
+  }
+
+  return {
+    buffer: await resized.jpeg({ quality: 84, mozjpeg: true }).toBuffer(),
+    ext: "jpg",
+    width,
+    height,
+  };
+}
+
+async function uploadVisionAnalysisImage(buffer: Buffer, filenamePrefix: string) {
+  const { buffer: processedBuffer, ext, width, height } = await downscaleImageForVision(buffer);
+  const filename = `${filenamePrefix}-${Date.now()}.${ext}`;
+  const uploaded = await uploadFile(filename, processedBuffer as any);
+  console.log("[PickObject] 视觉分析图片已缩放:", {
+    originalWidth: width,
+    originalHeight: height,
+    maxSide: MAX_VISION_SIDE,
+    outputExt: ext,
+  });
+  return uploaded.url;
 }
 
 async function normalizeImageInput(imageData: string): Promise<string> {
@@ -42,7 +85,21 @@ async function normalizeImageInput(imageData: string): Promise<string> {
   if (!raw) throw new Error("缺少图片数据");
 
   if (/^https?:\/\//i.test(raw)) {
-    return raw;
+    const controller = new AbortController();
+    const timeout = setTimeout(() => controller.abort(), PICK_OBJECT_FETCH_TIMEOUT_MS);
+    try {
+      const response = await fetch(raw, {
+        method: "GET",
+        signal: controller.signal,
+      });
+      if (!response.ok) {
+        throw new Error(`拉取图片失败: HTTP ${response.status}`);
+      }
+      const arrayBuffer = await response.arrayBuffer();
+      return uploadVisionAnalysisImage(Buffer.from(arrayBuffer), "pick-object-url");
+    } finally {
+      clearTimeout(timeout);
+    }
   }
 
   const base64Match = raw.match(/^data:(image\/[a-zA-Z0-9.+-]+);base64,(.+)$/);
@@ -50,13 +107,9 @@ async function normalizeImageInput(imageData: string): Promise<string> {
     throw new Error("图片格式不支持");
   }
 
-  const mimeType = base64Match[1];
   const base64Body = base64Match[2];
   const buffer = Buffer.from(base64Body, "base64");
-  const ext = inferFileExtFromMime(mimeType);
-  const filename = `pick-object-${Date.now()}.${ext}`;
-  const uploaded = await uploadFile(filename, buffer as any);
-  return uploaded.url;
+  return uploadVisionAnalysisImage(buffer, "pick-object-base64");
 }
 
 export async function POST(request: Request) {
@@ -83,7 +136,7 @@ export async function POST(request: Request) {
     
     const normalizedImage = await normalizeImageInput(imageData);
 
-    console.log("[PickObject] 使用硬编码模型: qwen2.5-vl");
+    console.log("[PickObject] 使用硬编码模型: qwen3-vl");
     
     // 构建 prompt - AI 通过红色标记来识别用户选择的物品
     const prompt = `
@@ -113,7 +166,7 @@ ${POSITION_OPTIONS.join('、')}
     console.log('[PickObject] 调用视觉模型...');
 
     // 调用视觉模型，发送带红色标记的图片
-    const { text: responseText, model } = await callQwen25Vl({
+    const { text: responseText, model } = await callQwen3Vl({
       imageUrl: normalizedImage,
       prompt,
       temperature: 0.1,

lib/ai/qwen3-vl.ts

@@ -0,0 +1,116 @@
+import "server-only";
+
+import { getEnabledAiProviders } from "@/lib/db/queries";
+
+const QWEN3_VL_MODELS = [
+  "qwen3-vl-plus",
+  "qwen-vl-plus",
+] as const;
+
+const QWEN_COMPAT_BASE_URL = "https://dashscope.aliyuncs.com/compatible-mode/v1";
+
+function parseAssistantText(payload: any): string {
+  const choices = Array.isArray(payload?.choices) ? payload.choices : [];
+  const content = choices[0]?.message?.content;
+
+  if (typeof content === "string") {
+    return content.trim();
+  }
+
+  if (Array.isArray(content)) {
+    const chunks = content
+      .map((part: any) => (typeof part?.text === "string" ? part.text.trim() : ""))
+      .filter(Boolean);
+    return chunks.join("\n").trim();
+  }
+
+  return "";
+}
+
+async function getQwenImageProviderApiKey() {
+  const providers = await getEnabledAiProviders();
+  const provider = providers.find(
+    (item) => item.type === "alibaba-image" && typeof item.apiKey === "string" && item.apiKey.trim(),
+  );
+
+  if (!provider?.apiKey) {
+    throw new Error("未配置可用的千问生图 Provider API Key");
+  }
+
+  return provider.apiKey.trim();
+}
+
+export async function callQwen3Vl(params: {
+  imageUrl: string;
+  prompt: string;
+  temperature?: number;
+  maxTokens?: number;
+}) {
+  const apiKey = await getQwenImageProviderApiKey();
+  const endpoint = `${QWEN_COMPAT_BASE_URL}/chat/completions`;
+
+  let lastError: Error | null = null;
+
+  for (const model of QWEN3_VL_MODELS) {
+    try {
+      const controller = new AbortController();
+      const timeout = setTimeout(() => controller.abort(), 30000);
+
+      try {
+        const response = await fetch(endpoint, {
+          method: "POST",
+          signal: controller.signal,
+          headers: {
+            "Content-Type": "application/json",
+            Authorization: `Bearer ${apiKey}`,
+          },
+          body: JSON.stringify({
+            model,
+            temperature: params.temperature ?? 0.1,
+            max_tokens: params.maxTokens ?? 300,
+            messages: [
+              {
+                role: "user",
+                content: [
+                  { type: "text", text: params.prompt },
+                  {
+                    type: "image_url",
+                    image_url: {
+                      url: params.imageUrl,
+                    },
+                  },
+                ],
+              },
+            ],
+          }),
+        });
+
+        const payload = await response.json().catch(() => ({}));
+        if (!response.ok) {
+          const message =
+            payload?.error?.message || payload?.message || `HTTP ${response.status}`;
+          throw new Error(message);
+        }
+
+        const text = parseAssistantText(payload);
+        if (!text) {
+          throw new Error("Qwen3-VL 未返回文本内容");
+        }
+
+        return { text, model };
+      } finally {
+        clearTimeout(timeout);
+      }
+    } catch (error) {
+      lastError = error instanceof Error ? error : new Error(String(error));
+      const errorText = String(lastError.message || "");
+      const isModelNotFound =
+        /model.*not.*found|invalid model|not support/i.test(errorText);
+      if (!isModelNotFound) {
+        break;
+      }
+    }
+  }
+
+  throw lastError || new Error("Qwen3-VL 调用失败");
+}