update api.py

phase/Student_view/chatbot.py

@@ -9,7 +9,7 @@ from huggingface_hub import InferenceClient
 # Hugging Face model setup
 # -------------------------------
 HF_TOKEN = os.getenv("HF_TOKEN")
-GEN_MODEL = os.getenv("GEN_MODEL", "HuggingFaceTB/SmolLM2-135M-Instruct")  # CPU-friendly
+GEN_MODEL = os.getenv("GEN_MODEL", "google/flan-t5-base")  # ✅ CPU-friendly seq2seq
 
 if not HF_TOKEN:
     st.error("⚠️ HF_TOKEN is not set. In your Space, add a Secret named HF_TOKEN.")
@@ -33,45 +33,53 @@ def add_message(text: str, sender: str):
     )
 
 
+TUTOR_PROMPT = (
+    "You are a kind Jamaican primary-school finance tutor. "
+    "Keep answers short, friendly, and age-appropriate. "
+    "Teach step-by-step with tiny examples. Avoid giving personal financial advice."
+)
+
+
 # -------------------------------
 # Hugging Face reply function
 # -------------------------------
+def _format_history_for_flan(messages: list[dict]) -> str:
+    # Map your session_state messages → "User:" / "Tutor:" lines
+    lines = []
+    for m in messages:
+        txt = (m.get("text") or "").strip()
+        if not txt:
+            continue
+        if m.get("sender") == "assistant":
+            lines.append(f"Tutor: {txt}")
+        else:
+            lines.append(f"User: {txt}")
+    return "\n".join(lines)
+
 def _reply_with_hf():
     if "client" not in globals():
         raise RuntimeError("HF client not initialized")
 
-    # System prompt
-    system = "You are a kind Jamaican primary-school finance tutor. Keep answers short, friendly, and age-appropriate."
+    # Build single instruction string FLAN-T5 prefers
+    convo = _format_history_for_flan(st.session_state.get("messages", []))
+    prompt = f"{TUTOR_PROMPT}\n\n{convo}\n\nTutor:"
 
-    # Build conversation history
-    convo = []
-    for m in st.session_state.messages:
-        role = "USER" if m["sender"] != "assistant" else "ASSISTANT"
-        convo.append(f"{role}: {m['text']}")
-    prompt = (system + "\n\n" + "\n".join(convo) + "\nASSISTANT:").strip()
-
-    # Call Hugging Face text-generation
     try:
-        response = client.text_generation(
+        response = client.text2text_generation(
             prompt,
             max_new_tokens=220,
-            temperature=0.2,
-            stream=False
+            temperature=0.2
         )
-
-        # Handle different response formats
         if isinstance(response, dict) and "generated_text" in response:
-            return response["generated_text"].strip()
+            return (response["generated_text"] or "").strip()
         elif isinstance(response, str):
             return response.strip()
         else:
-            raise ValueError(f"Unexpected response format: {response}")
-
+            raise ValueError(f"Unexpected response format: {type(response)}")
     except Exception as e:
         err_text = ''.join(traceback.format_exception_only(type(e), e)).strip()
         raise RuntimeError(f"Hugging Face API Error: {err_text}")
 
-
 # -------------------------------
 # Streamlit page
 # -------------------------------