Update app.py

app.py

@@ -46,8 +46,9 @@ SPACE_REPO_ID  = os.environ.get("SPACE_REPO_ID", "")
 # Generation / toggles
 ALLOW_WIKIPEDIA = False
 DEBUG = True
-MAX_NEW_TOKENS_GROUNDED = 512
-MAX_NEW_TOKENS_FALLBACK = 256
+MAX_NEW_TOKENS_GROUNDED = 384   # was 512
+MAX_NEW_TOKENS_FALLBACK = 192   # was 256
+
 MIN_USEFUL_CHARS = 260
 
 def dlog(tag, msg):
@@ -970,17 +971,18 @@ with gr.Blocks(theme="soft") as demo:
     )
 
     send_btn.click(
-        predict,
-        inputs=[user_in, chat, state],
-        outputs=[chat, user_in, feedback_grp, rating, comment, state],
-        concurrency_limit=8,  # heavy LLM call
+    predict,
+    inputs=[user_in, chat, state],
+    outputs=[chat, user_in, feedback_grp, rating, comment, state],
+    concurrency_limit=3,
     )
 
+
     user_in.submit(
-        predict,
-        inputs=[user_in, chat, state],
-        outputs=[chat, user_in, feedback_grp, rating, comment, state],
-        concurrency_limit=8,  # heavy LLM call
+    predict,
+    inputs=[user_in, chat, state],
+    outputs=[chat, user_in, feedback_grp, rating, comment, state],
+    concurrency_limit=3,
     )
 
     clear_btn.click(