PDFVisionFallback

src/agent.ts

@@ -338,18 +338,28 @@ export class AgentExecutor {
             const reqMessages = this.buildRequestHistory(this.chatHistory);
 
             // Handle Vision Content Injection
-            // Merge text prompt with file content instead of replacing, so the user's message is never lost
+            // visionContent 배열에서 이미지 base64 데이터를 추출하여 엔진에 맞는 형식으로 주입
             if (hasVisionContent && reqMessages.length > 0) {
                 const lastUserIdx = reqMessages.map(m => m.role).lastIndexOf('user');
                 if (lastUserIdx >= 0) {
                     const existingContent = reqMessages[lastUserIdx].content;
-                    const textParts: any[] = (typeof existingContent === 'string' && existingContent.trim())
-                        ? [{ type: 'text', text: existingContent }]
-                        : [];
+                    const textContent = (typeof existingContent === 'string' && existingContent.trim()) ? existingContent : '';
+                    
+                    // base64 이미지 데이터 추출
+                    const imageBase64List: string[] = [];
+                    for (const vc of (visionContent || [])) {
+                        if (vc && vc.data) {
+                            imageBase64List.push(vc.data);
+                        }
+                    }
+
+                    // Ollama 호환: images 배열 필드에 base64 데이터 직접 주입
+                    // LM Studio 호환: content 배열에 image_url 객체 주입
                     reqMessages[lastUserIdx] = {
                         role: 'user',
-                        content: JSON.stringify([...textParts, ...(visionContent || [])])
-                    };
+                        content: textContent,
+                        images: imageBase64List // Ollama native format
+                    } as any;
                 }
             }
 
@@ -1925,10 +1935,15 @@ export class AgentExecutor {
                 ? message.content
                 : JSON.stringify(message.content);
 
-            return {
+            const result: any = {
                 role: message.role,
                 content: normalizedContent
             };
+            // Ollama Vision: images 필드 보존
+            if ((message as any).images) {
+                result.images = (message as any).images;
+            }
+            return result;
         });
     }