trail3 app.py

app.py

@@ -7,6 +7,21 @@ memory_db = {}
 # Load the model
 interpreter = pipeline("text2text-generation", model="google/flan-t5-base")
 
+# Few-shot examples for priming the model
+few_shot_examples = """
+Example 1:
+Message: "sure. whatever you want"
+→ Interpretation: Passive-aggressive, likely annoyed but masking it with false agreement.
+
+Example 2:
+Message: "we’ll see"
+→ Interpretation: Avoidant, doesn’t want to commit or say no directly.
+
+Example 3:
+Message: "it's fine, don't worry about it"
+→ Interpretation: Likely masking frustration or disappointment, doesn’t want to express upset directly.
+"""
+
 # Interpret a message with optional memory and return AI explanation + history
 def interpret_message(msg, name, closeness, traits, relation, past_events, feelings, use_memory):
     if not msg.strip() or not name.strip():
@@ -30,7 +45,7 @@ def interpret_message(msg, name, closeness, traits, relation, past_events, feeli
             }
 
     # Construct rich prompt
-    prompt = f"""New message from {name}: \"{msg}\"\n\nKnown context:\n- Closeness: {closeness}\n- Personality: {traits}\n- Relationship: {relation}\n- Similar Past Events: {past_events}\n- Your Feelings: {feelings}\n\nBased on the message and context, explain what the person is *really thinking* or feeling but not saying directly.\nInfer their hidden intention, emotional state, and subtext. Then suggest how I should respond."""
+    prompt = f"""{few_shot_examples}\n\nNew message from {name}: \"{msg}\"\n\nKnown context:\n- Closeness: {closeness}\n- Personality: {traits}\n- Relationship: {relation}\n- Similar Past Events: {past_events}\n- Your Feelings: {feelings}\n\nBased on the message and context, explain what the person is *really thinking* or feeling but not saying directly.\nInfer their hidden intention, emotional state, and subtext. Then suggest how I should respond."""
 
     # Save to memory if applicable
     if use_memory:
@@ -68,7 +83,19 @@ def memory_chat(user_input, person_name):
     history = memory_db[name]["history"]
     past_text = "\n\n".join([f"- {h['message']}\n  -> {h['interpretation']}" for h in history])
 
-    chat_prompt = f"""You are a relationship and social communication expert.\n\nYou remember all past messages and responses exchanged with {name}.\nThis is the known message history:\n{past_text}\n\nNow the user wants to understand more about what {name} might be thinking or feeling deep down,\nbased on what they've said before and how they usually act.\n\nUser’s question: {user_input}\n\nUse what you know from the history to infer hidden meaning, emotional tone, and behavior patterns.\nGive a thoughtful and emotionally intelligent answer. Avoid repeating the user's feelings."""
+    chat_prompt = f"""You are a relationship and social communication expert trained in human behavior.
+
+Here is what you know about {name}'s recent messages and past behavior:
+{past_text}
+
+The user now asks: {user_input}
+
+Respond with a supportive and emotionally intelligent answer based on past messages.
+Avoid repeating the user's input. Be specific. If {name} is likely upset, say it.
+If the user should give space or check in, say so.
+
+Only output your advice in 2–4 sentences. No explanation of what you’re doing.
+"""
 
     response = interpreter(chat_prompt, max_new_tokens=200)[0]["generated_text"]
     return response