feat(vision): add Qwen3-VL vision model support with Dockerfile and tests; improve invoice OCR conversion and prompts; simplify extraction flow by removing consensus voting

test/helpers/docker.ts

@@ -311,9 +311,8 @@ export async function ensureOllamaModel(modelName: string): Promise<boolean> {
     if (response.ok) {
       const data = await response.json();
       const models = data.models || [];
-      const exists = models.some((m: { name: string }) =>
-        m.name === modelName || m.name.startsWith(modelName.split(':')[0])
-      );
+      // Exact match required - don't match on prefix
+      const exists = models.some((m: { name: string }) => m.name === modelName);
 
       if (exists) {
         console.log(`[Ollama] Model already available: ${modelName}`);
@@ -371,3 +370,16 @@ export async function ensureMinistral3(): Promise<boolean> {
   // Then ensure the Ministral 3 8B model is pulled
   return ensureOllamaModel('ministral-3:8b');
 }
+
+/**
+ * Ensure Qwen3-VL 8B model is available (vision-language model)
+ * Q4_K_M quantization (~5GB) - fits in 15GB VRAM with room to spare
+ */
+export async function ensureQwen3Vl(): Promise<boolean> {
+  // First ensure the Ollama service is running
+  const ollamaOk = await ensureMiniCpm();
+  if (!ollamaOk) return false;
+
+  // Then ensure Qwen3-VL 8B is pulled
+  return ensureOllamaModel('qwen3-vl:8b');
+}