Update ai-routes.js

ai-routes.js

@@ -94,6 +94,49 @@ function convertGeminiToOpenAI(baseParams) {
     return messages;
 }
 
+// Helper to adapt params for Gemma (Remove systemInstruction and prepend to prompt)
+function prepareGemmaParams(baseParams, modelName) {
+    // Deep copy to avoid mutating original for other providers
+    const newParams = JSON.parse(JSON.stringify(baseParams));
+    newParams.model = modelName;
+
+    // Gemma models do not support systemInstruction in config, move it to text
+    if (newParams.config && newParams.config.systemInstruction) {
+        const sysInst = newParams.config.systemInstruction;
+        delete newParams.config.systemInstruction;
+
+        const systemPrefix = `[System Instruction: ${sysInst}]\n\n`;
+
+        if (Array.isArray(newParams.contents) && newParams.contents.length > 0) {
+            // Chat mode: Prepend to first user message
+            const firstMsg = newParams.contents[0];
+            if (firstMsg.role === 'user' && firstMsg.parts && firstMsg.parts.length > 0) {
+                const textPart = firstMsg.parts.find(p => p.text);
+                if (textPart) {
+                    textPart.text = systemPrefix + textPart.text;
+                } else {
+                    // Only has image/audio, add text part
+                    firstMsg.parts.unshift({ text: systemPrefix });
+                }
+            } else {
+                // If history starts with model (rare), prepend a new user message
+                newParams.contents.unshift({ role: 'user', parts: [{ text: systemPrefix }] });
+            }
+        } else if (newParams.contents && newParams.contents.parts) {
+            // Generate content mode (single object)
+            if (Array.isArray(newParams.contents.parts)) {
+                const textPart = newParams.contents.parts.find(p => p.text);
+                if (textPart) {
+                    textPart.text = systemPrefix + textPart.text;
+                } else {
+                    newParams.contents.parts.unshift({ text: systemPrefix });
+                }
+            }
+        }
+    }
+    return newParams;
+}
+
 // --- Dynamic Provider Management ---
 
 const PROVIDERS = {
@@ -179,13 +222,14 @@ async function callOpenRouterProvider(baseParams) {
 
 async function callGemmaProvider(baseParams) {
     const client = await getGeminiClient(); // AWAIT client
-    // Gemma via Gemini SDK usually shares quota, but using different models might help
-    const fallbackModels = ['gemma-3-27b', 'gemma-3-9b', 'gemma-3-4b']; 
+    // Updated to Gemma 3 series as per user quota (including smaller ones)
+    const fallbackModels = ['gemma-3-27b-it', 'gemma-3-9b-it', 'gemma-3-4b-it']; 
     let lastError = null;
     for (const modelName of fallbackModels) {
         try {
             console.log(`🛡️ [AI Debug] Switching to Backup (Gemma): ${modelName}`);
-            const currentParams = { ...baseParams, model: modelName };
+            // Use adapter to remove systemInstruction
+            const currentParams = prepareGemmaParams(baseParams, modelName);
             return await callAIWithRetry(() => client.models.generateContent(currentParams), 1);
         } catch (e) {
             lastError = e;
@@ -298,12 +342,14 @@ async function streamOpenRouter(baseParams, res) {
 
 async function streamGemma(baseParams, res) {
     const client = await getGeminiClient(); // AWAIT client
-    const models = ['gemma-3-27b', 'gemma-3-9b', 'gemma-3-4b'];
+    // Updated to Gemma 3 series as per user quota
+    const models = ['gemma-3-27b-it', 'gemma-3-9b-it', 'gemma-3-4b-it'];
     let lastError = null;
     for (const modelName of models) {
         try {
             console.log(`🛡️ [AI Debug] STREAMING Gemma: ${modelName}`);
-            const currentParams = { ...baseParams, model: modelName };
+            // Use adapter to remove systemInstruction
+            const currentParams = prepareGemmaParams(baseParams, modelName);
             const result = await client.models.generateContentStream(currentParams);
             let fullText = "";
             // FIX: Iterate result directly
@@ -318,6 +364,7 @@ async function streamGemma(baseParams, res) {
             return fullText;
         } catch (e) {
             lastError = e;
+            console.warn(`Streaming GEMMA ${modelName} failed:`, e.message);
         }
     }
     throw lastError || new Error("Gemma streaming failed");