Update app.py

app.py

@@ -50,7 +50,6 @@ HF_TOKEN = os.getenv("HUGGINGFACE_HUB_TOKEN") or os.getenv("HF_TOKEN")
 
 COHERE_API_KEY = os.getenv("COHERE_API_KEY")
 USE_HOSTED_COHERE = bool(COHERE_API_KEY and _HAS_COHERE)
-COHERE_TIMEOUT_SEC = float(os.getenv("COHERE_TIMEOUT_SEC", "30"))
 
 MAX_NEW_TOKENS = int(os.getenv("MAX_NEW_TOKENS", "512"))
 
@@ -104,7 +103,8 @@ def _history_to_prompt(message, history):
 # ---------- Cohere path (default first; fallback to local on failure) ----------
 _co_client = None
 if USE_HOSTED_COHERE:
-    _co_client = cohere.Client(api_key=COHERE_API_KEY, timeout=COHERE_TIMEOUT_SEC)
+    # Avoid passing unsupported args; some SDK builds don't accept timeout=
+    _co_client = cohere.Client(api_key=COHERE_API_KEY)
 
 def cohere_chat(message, history):
     """
@@ -305,7 +305,6 @@ def clarityops_reply(user_msg, history, tz, uploaded_files_paths):
         # Surface the error in-chat so the websocket doesn’t die silently
         err = f"Error: {e}"
         try:
-            # Optionally log stack for ops
             traceback.print_exc()
         except Exception:
             pass
@@ -349,15 +348,14 @@ with gr.Blocks(theme=theme, css=custom_css) as demo:
 
     # --- Background warmup so first message doesn't time out ---
     def _warmup():
+        # IMPORTANT: no return value, because we register with outputs=None
         def _bg():
             try:
-                # Preload local fallback model quietly (download may be large)
-                load_local_model()
+                load_local_model()  # Preload local fallback quietly
             except Exception:
                 pass
         threading.Thread(target=_bg, daemon=True).start()
-        return ""
-    demo.load(_warmup, None, None)  # fire-and-forget
+    demo.load(_warmup)  # no inputs, no outputs
 
     gr.Markdown("# ClarityOps Augmented Decision AI")
 
@@ -402,7 +400,6 @@ with gr.Blocks(theme=theme, css=custom_css) as demo:
             new_history = clarityops_reply(user_msg.strip(), history or [], tz, up_paths or [])
             return new_history, "", new_history
         except Exception as e:
-            # Return the error in the chat so UI doesn't show a websocket failure
             err = f"Error: {e}"
             try:
                 traceback.print_exc()
@@ -430,8 +427,8 @@ with gr.Blocks(theme=theme, css=custom_css) as demo:
 
 if __name__ == "__main__":
     port = int(os.environ.get("PORT", "7860"))
-    # max_threads gives the worker pool; leave as-is unless you need more parallelism
     demo.launch(server_name="0.0.0.0", server_port=port, show_api=False, max_threads=8)
 
 
 
+