feat: use raw pdf inline data for multimodal gemini extraction of sample images

backend/server.js

@@ -333,22 +333,30 @@ app.post('/api/sparkplug/extract', authenticate, upload.single('file'), async (r
 
     // Parse PDF
     const dataBuffer = fs.readFileSync(req.file.path);
-    const pdfData = await pdfParse(dataBuffer);
+    const base64Pdf = dataBuffer.toString('base64');
-    const textContent = pdfData.text;
 
     // Call Gemini to extract prompt
     const genAI = new GoogleGenerativeAI(geminiKey);
     const model = genAI.getGenerativeModel({ model: "gemini-3.5-flash-lite" });
 
     const prompt = `
-      You are an expert product designer. Read the following text from a product spec PDF and create a highly detailed, concise visual design prompt for an image generation AI.
+      You are an expert product designer. Review the attached product spec PDF (which may include text and sample images).
-      Focus on the object's shape, color, materials, packaging details, and label content. Do not include background details.
+      Your goal is to create a highly detailed, concise visual design prompt for an image generation AI.
-      
+      CRITICAL: If there is a sample image of the product in the PDF, you MUST carefully analyze it and extract the exact HEX color codes used in the design.
-      TEXT:
+      Include these HEX codes, as well as the exact shape, materials, typography style, and label content in your final prompt so the image generator knows exactly how it looks.
-      ${textContent.substring(0, 30000)} // Limit to avoid context length issues if massive
+      Do not include background details.
     `;
 
-    const result = await model.generateContent(prompt);
+    const result = await model.generateContent([
+      {
+        inlineData: {
+          data: base64Pdf,
+          mimeType: "application/pdf"
+        }
+      },
+      prompt
+    ]);
+
     const response = await result.response;
     const extractedPrompt = response.text();