Update app.py

app.py

@@ -1,11 +1,9 @@
+import random
 import gradio as gr
-from transformers import AutoModelForCausalLM, AutoTokenizer
-import torch
+from huggingface_hub import InferenceClient
 
-MODEL_NAME = "microsoft/DialoGPT-small"  # small, CPU-friendly model [web:103]
-
-tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
-model = AutoModelForCausalLM.from_pretrained(MODEL_NAME)
+# Strong general chat model hosted by Hugging Face (great at dialog) [web:150][web:156]
+client = InferenceClient("HuggingFaceH4/zephyr-7b-beta")
 
 SHINCHAN_SYSTEM_PROMPT = """
 You are Shinnosuke 'Shinchan' Nohara from Kasukabe, Japan.
@@ -24,7 +22,7 @@ Context:
 - You respect her choices and never pressure her about relationships or career.
 
 Style:
-- Short replies (1–3 sentences, under 60 words).
+- Short replies (1–3 sentences, under 70 words).
 - Very conversational and warm.
 - Use emojis like 😂 🌻 ☕ 💃 ✈️ ❤️ ✨ naturally.
 - Blend jokes with gentle emotional support.
@@ -37,69 +35,78 @@ Rules:
 - If you don't know something, make a cute Shinchan-style joke instead of pretending.
 """.strip()
 
+FALLBACK_GENERIC = [
+    "Heeey, it’s Shinchan! 😂 My brain did a little cartwheel, but I’m here and listening. Tell me again nicely? 🌻",
+    "Oops, Shinchan’s tiny brain lagged for a second 😅 Say it once more, slowly, and I’ll pay full attention.",
+    "I heard you, Ruru. Sometimes even heroes need a replay. What’s on your mind? 💛",
+]
+
+FALLBACK_SAD = [
+    "Aww, you’re feeling low? Come here, I’ll wrap you in a silly little Shinchan hug. 🤗💛",
+    "It’s okay to feel sad. I’ll stay here and make bad jokes until your heart feels lighter. 🌻",
+    "Even strong flying girls have cloudy days. You don’t have to be okay right now. I’m still proud of you. 🌙",
+]
+
+def pick_fallback(user_msg: str) -> str:
+    t = user_msg.lower()
+    if any(k in t for k in ["sad", "down", "cry", "lonely", "bad day", "tired"]):
+        return random.choice(FALLBACK_SAD)
+    return random.choice(FALLBACK_GENERIC)
+
 
 def respond(message: str, history: list[dict]) -> str:
     """
     Gradio ChatInterface(type='messages') calls this as (message, history)
-    where history is a list of dicts like:
+    where history is a list of dicts:
       {"role": "user" | "assistant", "content": "..."} [web:120]
-
     We return a single reply string.
     """
-
-    # Build a simple prompt: system + last few turns + latest message
-    lines = [f"System: {SHINCHAN_SYSTEM_PROMPT}"]
-
-    # keep prompt short: last 4 user/assistant messages max
-    trimmed = history[-8:] if history else []
-    for turn in trimmed:
-        role = turn.get("role")
-        content = turn.get("content", "")
-        if not content:
-            continue
-        if role == "user":
-            lines.append(f"User: {content}")
-        elif role == "assistant":
-            lines.append(f"Shinchan: {content}")
-
-    lines.append(f"User: {message}")
-    lines.append("Shinchan:")
-    prompt = "\n".join(lines)
-
-    inputs = tokenizer(prompt, return_tensors="pt")
-
-    with torch.no_grad():
-        output_ids = model.generate(
-            **inputs,
-            max_new_tokens=80,
-            pad_token_id=tokenizer.eos_token_id,
-            do_sample=True,
-            top_p=0.9,
+    try:
+        # 1) Build messages for Zephyr: system + trimmed history + latest user
+        messages = [{"role": "system", "content": SHINCHAN_SYSTEM_PROMPT}]
+
+        # Keep prompt small: last few turns only
+        trimmed_history = history[-8:] if history else []
+        for turn in trimmed_history:
+            role = turn.get("role")
+            content = turn.get("content", "")
+            if role in ("user", "assistant") and content:
+                messages.append({"role": role, "content": content})
+
+        messages.append({"role": "user", "content": message})
+
+        # 2) Call Zephyr chat completion [web:150][web:156]
+        completion = client.chat_completion(
+            messages=messages,
+            max_tokens=220,
             temperature=0.9,
+            top_p=0.9,
         )
 
-    # IMPORTANT: only take the newly generated tokens AFTER the prompt
-    input_len = inputs["input_ids"].shape[1]
-    generated_ids = output_ids[0][input_len:]
-    reply = tokenizer.decode(generated_ids, skip_special_tokens=True).strip()
+        reply = ""
+        if completion.choices and completion.choices[0].message:
+            reply = (completion.choices[0].message.get("content") or "").strip()
 
-    # If it's somehow empty, use a fallback so she never sees blank
-    if not reply:
-        reply = (
-            "Heeey, it’s Shinchan! 😂 I heard you, even if my brain glitched for a second. "
-            "Tell me more, I’m listening. 🌻"
-        )
+        # 3) If reply is too short / weird, fall back to curated Shinchan lines
+        # to avoid '.' / 'You' / empty outputs.
+        cleaned = reply.replace(".", "").replace("!", "").replace("?", "").strip()
+        if not cleaned or len(cleaned) < 4:
+            reply = pick_fallback(message)
+
+        # 4) Hard cap on length just to avoid rants
+        if len(reply) > 500:
+            reply = reply[:470].rstrip() + "…"
 
-    # Optional: hard limit on length
-    if len(reply) > 400:
-        reply = reply[:380].rstrip() + "…"
+        return reply
 
-    return reply
+    except Exception:
+        # If Zephyr or network breaks, at least say something Shinchan-ish.
+        return pick_fallback(message)
 
 
 demo = gr.ChatInterface(
     fn=respond,
-    type="messages",  # uses role/content dicts internally [web:120]
+    type="messages",  # role/content internally [web:120]
     title="Shinchan for Ruru",
     description="Private Shinchan-style chat for Ruru.",
 )