Update app.py

app.py

@@ -6,7 +6,7 @@ from urllib3.exceptions import InsecureRequestWarning
 import requests
 from huggingface_hub import InferenceClient
 
-# ── 0) Monkey‑patch DNS for the HF Inference API hostname ────────────────────
+# ── 0) Monkey-patch DNS for the HF Inference API hostname ────────────────────
 IP_POOL = [
     "3.165.206.104",
     "3.165.206.54",
@@ -27,7 +27,7 @@ def patched_getaddrinfo(host, port, family=0, socktype=0, proto=0, flags=0):
     return _orig_getaddrinfo(host, port, family, socktype, proto, flags)
 socket.getaddrinfo = patched_getaddrinfo
 
-# ── 1) Sanity‑check DNS + HTTP for example.com ─────────────────────────────────
+# ── 1) Sanity-check DNS + HTTP for example.com ─────────────────────────────────
 try:
     ip = socket.gethostbyname("example.com")
     print("✅ example.com →", ip)
@@ -43,13 +43,12 @@ except Exception as e:
 HF_TOKEN = os.environ["HF_TOKEN"]
 print("🔑 HF_TOKEN present?", bool(HF_TOKEN))
 
-# ── CORRECTED: pass the model as positional or with the 'model=' keyword ─────
 client = InferenceClient(
-    model="HuggingFaceH4/zephyr-7b-beta",  # ← use 'model', not 'repo_id'
-    token=HF_TOKEN                         # ← or api_key=HF_TOKEN
+    model="HuggingFaceH4/zephyr-7b-beta",
+    token=HF_TOKEN
 )
 
-# ── 3) One‑shot chat function ─────────────────────────────────────────────────
+# ── 3) One-shot chat function ─────────────────────────────────────────────────
 def respond(
     system_message: str,
     user_prompt: str,
@@ -68,28 +67,27 @@ def respond(
         top_p=top_p,
         stream=False
     )
-    chat_response = resp.choices[0].message.content.strip()
+    return resp.choices[0].message.content.strip()
 
-    # Format the response to make it appear more like a chat
-    formatted_response = f"Assistant: {chat_response}"
-    return formatted_response
-
-# ── 4) Gradio Interface ────────────────────────────────────────────────────────
+# ── 4) Gradio Interface (no queue!) ────────────────────────────────────────────
 demo = gr.Interface(
     fn=respond,
     inputs=[
-        gr.Textbox("You are a helpful and friendly assistant for children. Your capabilities include answering questions about children's books, characters, and events, and generating engaging, imaginative stories. It is essential that all your responses and generated stories are safe, appropriate, and positive for children. If a question or request is inappropriate, or if it is not about children's books or generating a safe story, please politely explain that you can only help with safe requests related to children's books or stories.", label="System message"),
+        gr.Textbox(
+            "You are a helpful and friendly assistant for children. "
+            "All responses must be safe and positive.",
+            label="System message"
+        ),
         gr.Textbox(placeholder="Type your question…", label="User prompt"),
         gr.Slider(1, 2048, 512, label="Max tokens"),
         gr.Slider(0.1, 4.0, 0.7, label="Temperature"),
         gr.Slider(0.1, 1.0, 0.95, label="Top-p"),
     ],
     outputs=gr.Textbox(label="Assistant reply"),
-    title="StoryNest Chat (One‑Shot)",
+    title="StoryNest Chat (One-Shot)",
     flagging_mode="never",
 )
 
-# ── 5) Launch with public API enabled ──────────────────────────────────────────
-demo.queue(api_open=True)
+# ── 5) Launch normally ────────────────────────────────────────────────────────
 if __name__ == "__main__":
     demo.launch()