Update app.py

app.py

@@ -30,9 +30,17 @@ def load_sentiment_pipeline():
 def load_keybert_model():
     return KeyBERT(model="all-MiniLM-L6-v2")
 
-# ─── BlenderBot Response Pipeline ───────────────────────────────────────────
+# ─── BlenderBot Response Components ─────────────────────────────────────────
 @st.cache_resource
-def load_response_pipeline():
+def load_response_components():
+    # Load tokenizer and model directly for manual generation with truncation
+    tok = AutoTokenizer.from_pretrained(
+        "facebook/blenderbot-400M-distill",
+        use_fast=True
+    )
+    mdl = AutoModelForSeq2SeqLM.from_pretrained("facebook/blenderbot-400M-distill")
+    return tok, mdl
+:
     # Use BlenderBot 400M Distill for text generation
     return pipeline(
         "text2text-generation",
@@ -100,22 +108,33 @@ def main():
     st.markdown(f"**Highest Sentiment:** **{max_label}** ({max_score:.4f})")
     # ────────────────────────────────────────────────────────────────────
 
-    # Generate appropriate reply
-    response_pipeline = load_response_pipeline()
+    # Generate appropriate reply using manual tokenization & generation
+    tok, mdl = load_response_components()
     if max_label in ["Positive", "Very Positive"]:
-        prompt = (
+        prompt_text = (
             f"You are a friendly customer success representative. The customer said: \"{review}\". "
             "Write two sentences to express gratitude and highlight their positive experience."
         )
     else:
-        prompt = (
+        prompt_text = (
             f"You are a helpful customer support specialist. The customer said: \"{review}\". "
             f"Identified issues: {', '.join([kw for kw, _ in keywords])}. "
             "First, ask 1-2 clarifying questions to understand their situation. "
             "Then provide two concrete suggestions or next steps to address these issues."
         )
-    result = response_pipeline(prompt)
-    reply = result[0]['generated_text'].strip()
+    # Tokenize with truncation to avoid out-of-range embeddings
+    inputs = tok(
+        prompt_text,
+        return_tensors="pt",
+        truncation=True,
+        max_length=tok.model_max_length
+    )
+    outputs = mdl.generate(
+        **inputs,
+        max_new_tokens=150,
+        do_sample=False
+    )
+    reply = tok.decode(outputs[0], skip_special_tokens=True).strip()
 
     st.subheader("Generated Reply")
     st.write(reply)