Update app.py

app.py

@@ -39,6 +39,7 @@ GROQ_API_KEY_2 = os.getenv("GROQ_API_KEY_2")  # Reserved for STT
 GROQ_API_KEY_3 = os.getenv("GROQ_API_KEY_3")  # Reserved for TTS
 GROQ_API_KEY_4 = os.getenv("GROQ_API_KEY_4")  # Additional Key (Fallback)
 SERPAPI_KEY = os.getenv("SERPAPI_KEY")        # Search
+COHERE_API_KEY = os.getenv("COHERE_KEY")
 
 # List of API Keys for the Chat function
 GROQ_CHAT_KEYS = [
@@ -722,6 +723,104 @@ def chat():
     prompt = data.get("prompt", "")
     history = data.get("history", [])
 
+    # ======================
+    # 🖼️ VISION MODE (AUTO DETECT - BASE64 ONLY)
+    # ======================
+    image_base64 = data.get("image_base64")
+    if image_base64:
+        print("[VISION] 🖼️ Image detected → using c4ai-aya-vision-32b (Cohere)")
+
+        # Validate base64 quickly
+        try:
+            # If user sends full data URI like "data:image/png;base64,....", strip prefix for decoding test
+            test_b64 = image_base64
+            if test_b64.startswith("data:"):
+                test_b64 = test_b64.split(",", 1)[1]
+            base64.b64decode(test_b64, validate=True)
+        except Exception as e:
+            print(f"[VISION] ❌ Invalid base64: {e}")
+            return jsonify({"error": "Invalid base64 image"}), 400
+
+        # Build payload for Cohere v2 chat (vision)
+        cohere_url = "https://api.cohere.ai/v2/chat"
+        payload = {
+            "model": "c4ai-aya-vision-32b",
+            "messages": [
+                {
+                    "role": "user",
+                    "content": [
+                        {
+                            "type": "input_text",
+                            "text": prompt if prompt else "Describe this image."
+                        },
+                        {
+                            # Cohere accepts data URI as image url - keep the data URI if present
+                            "type": "input_image",
+                            "image": image_base64
+                        }
+                    ]
+                }
+            ],
+            # optional tuning
+            "temperature": 0.6,
+            "max_tokens": 800
+        }
+
+        headers = {
+            "Authorization": f"Bearer {COHERE_API_KEY}",
+            "Content-Type": "application/json"
+        }
+
+        try:
+            res = requests.post(cohere_url, json=payload, headers=headers, timeout=60)
+            try:
+                res_json = res.json()
+            except Exception:
+                print("[VISION] ❌ Cohere returned non-json response")
+                return jsonify({"error": "Cohere returned non-json response", "status_code": res.status_code, "text": res.text}), 500
+
+            # Robust extraction of text reply (support several possible shapes)
+            ai_reply = None
+            # Try common v2 shape: res_json["messages"][0]["content"][0]["text"]
+            try:
+                ai_reply = res_json.get("messages", [])[0].get("content", [])[0].get("text")
+            except Exception:
+                ai_reply = None
+
+            # Fallback: sometimes Cohere returns 'output' or 'text' fields
+            if not ai_reply:
+                # Try to find any 'text' in nested dicts
+                def find_text(obj):
+                    if isinstance(obj, dict):
+                        for k, v in obj.items():
+                            if k == "text" and isinstance(v, str):
+                                return v
+                            else:
+                                found = find_text(v)
+                                if found:
+                                    return found
+                    elif isinstance(obj, list):
+                        for item in obj:
+                            found = find_text(item)
+                            if found:
+                                return found
+                    return None
+                ai_reply = find_text(res_json) or ""
+
+            return jsonify({
+                "mode": "vision",
+                "reply": ai_reply,
+                "raw": res_json
+            })
+
+        except requests.exceptions.RequestException as e:
+            print(f"[VISION] ❌ RequestException: {e}")
+            return jsonify({"error": "Failed to call Cohere", "details": str(e)}), 500
+
+    # =====================================================
+    # 🧩 🎚️ (VISION DONE) — LANJUTKAN MODE TEXT SEPERTI BIASA
+    # =====================================================
+
     # Flags
     user_timezone_str = data.get("user_timezone", "Asia/Jakarta")
     current_username = data.get("current_username")
@@ -738,10 +837,10 @@ def chat():
     # SUPER GTE FLAG
     super_gte_active = data.get("super_gte", False)
 
-    # Rate limit logic
+    # Rate limit logic (kept placeholder as in your original)
     
 
-    # LIMIT CHECK
+    # LIMIT CHECK (kept placeholder)
     
 
     print(f"[CHAT] 💬 User Prompt (Text Mode): {prompt}")
@@ -771,13 +870,9 @@ def chat():
             return Response("Deep research requires a question.", mimetype="text/plain")
 
         def gen_deep():
-            
-
             final_answer = deep_research_mode(deep_query, history, num_sources=15)
             yield final_answer
 
-        
-
         response = Response(gen_deep(), mimetype="text/plain")
         
         return response
@@ -808,12 +903,9 @@ def chat():
         # Note: If agent_active is True, the Agent logic is handled inside stream_chat
 
     # ======================
-    # 💬 4. STANDARD CHAT (FIXED)
+    # 💬 4. STANDARD STREAM CHAT (unchanged)
     # ======================
 
-    # --- FIX: decrement MUST be OUTSIDE generator ---
-    
-
     def generate():
         for chunk in stream_chat(
             prompt,
@@ -827,9 +919,6 @@ def chat():
         ):
             yield chunk
 
-    # Ambil sisa limit setelah decrement
-    
-
     response = Response(generate(), mimetype="text/plain")
     
     return response