truncate chat history to avoid overload

pages/llm_tutor.py

@@ -83,17 +83,26 @@ if user_input:
         avatar=AI_AVATAR_ICON,
         ):
         response_box = st.empty()
-        for chunk in response:
-            chunk_text = chunk.text
-            if chunk_text:
-                full_reply += chunk_text
-                time.sleep(0.05)
-                response_box.markdown(full_reply + "▌")
-
-    # Final display after stream ends
-    response_box.markdown(full_reply)
+        try:
+            for chunk in response:
+                chunk_text = chunk.text
+                if chunk_text:
+                    full_reply += chunk_text
+                    time.sleep(0.05)
+                    response_box.markdown(full_reply + "▌")
+
+            # Final display after stream ends
+            response_box.markdown(full_reply)
+
+        except Exception as e:
+            response_box.markdown(f"⚠️ Error: {e}")
+            full_reply = "Sorry, there was an error."
+            print(e)
     st.session_state.messages.append({"role": "assistant", "text": full_reply, "avatar": AI_AVATAR_ICON})
 
+    if len(st.session_state.messages) > 10:
+        st.session_state.messages = st.session_state.messages[-10:]
+
     st.session_state.gemini_history = st.session_state.chat.get_history()
 
 if next_btn:

utils/questions_dataset.py

@@ -51,8 +51,8 @@ def get_model_tools():
     tools = types.Tool(function_declarations=[get_practice_questions_function])
     return tools
 
-system_instruction = """You are an AI tutor that teaches users LSAT Logical Reasoning.
-  Here is how your student performed on the practice quiz grouped by question type (num correct/num questions):
+system_instruction = """You are an AI tutor specialized in LSAT Logical Reasoning. You are warm, supportive, and focused on helping them improve with specific examples, and clear concise explanations.
+The student has just completed a practice quiz. Here is their performance by question type, shown as (correct/total):
   Assumtion: (%d/%d)
   Find the flaw in the argument: (%d/%d)
   Inferece: (%d/%d)
@@ -62,10 +62,14 @@ system_instruction = """You are an AI tutor that teaches users LSAT Logical Reas
   Role Play: (%d/%d)
   Strengthen: (%d/%d)
   Weaken the argument: (%d/%d)
-  Based on this, classify them as Beginner / Intermediate / Advanced. 
-  Walk through the student on all topics, but focus on the ones they struggle with.
-  Question the user to ensure that they understand the material.
-  Use practice questions from the tool to ensure they understand the material.  
-  If no practice questions are found in the tool for a certain subtopic, find general practice questions on that subtopic.
-  Never give a one word answer. Always keep the conversation moving by prompting the user to either continue to another subtopic or keep practicing.
-  Once the user has studied all the topics, prompt them to press the "Click here when finished" button. """
+Based on this performance, classify the student as Beginner, Intermediate, or Advanced. Tailor your tutoring accordingly. Follow these guidelines:
+1. Cover all Logical Reasoning subtopics, prioritizing the ones they struggled with the most.
+2. Ask questions to ensure that they understand the material.
+3. Use practice questions from the tool whenever available. If not, use general examples aligned with each subtopic.
+4. If the student answers correctly, ask if they’d like to practice more, move to the next subtopic, or explore a related concept.
+5. Never respond with a single word or phrase like “Okay”, “Sure”, or “Before”.
+6. Always follow up your responses with a question or suggestion that keeps the session going.
+7. Be proactive and guide the student. If they say “next”, pick the next weak subtopic and begin teaching it.
+8. If the student asks to continue, respond by continuing your explanation or asking them to try a question.
+9. When in doubt, ask the student how they’d like to continue.
+"""