Update app.py

app.py

@@ -1,4 +1,4 @@
-# app.py - UPDATED to integrate behavior_model
+# app.py - UPDATED to inject emoji + min-length guidance into system prompt
 import re
 import json
 import asyncio
@@ -111,52 +111,80 @@ async def generate_response_stream(messages, max_tokens=600, temperature=0.85):
             logger.exception("Flow analysis failed: %s", e)
             flow_context = {}
 
-        # Use flow context to drive behavior
-        flow_desc = ""
-        if flow_context:
-            label = flow_context.get("flow_label","unknown")
-            conf = round(float(flow_context.get("confidence", 0.0)), 2)
-            expl = flow_context.get("explanation","")
-            flow_desc = f"\n[FLOW] Detected: {label} (confidence {conf}). {expl}\n"
-
-        if intent == "internet_search":
-            yield f"data: {json.dumps({'status': 'Searching the web...'})}\n\n"
-            await asyncio.sleep(0)
+        # Get vibe/emotion context
+        vibe_data = get_smart_context(last_user_msg)
+
+        # Determine min words & emoji range by parsing vibe_data and flow label
+        # Defaults
+        min_words = 30
+        emoji_min, emoji_max = 0, 1
+
+        # Heuristics based on vibe_data text
+        if "Deep Dive Mode" in vibe_data:
+            min_words = 70
+        elif "Standard Chat Mode" in vibe_data:
+            min_words = 30
+        elif "Ping-Pong Mode" in vibe_data:
+            min_words = 12
+
+        # Try to extract suggested emoji range from vibe_data string if present
+        import re as _re
+        m = _re.search(r"Use\s+(\d+)–(\d+)\s+emoji", vibe_data)
+        if m:
             try:
-                tool_data_struct = perform_web_search(last_user_msg)
-            except Exception as e:
-                logger.exception("Web search failed: %s", e)
-                tool_data_struct = {"query": last_user_msg, "results": []}
-            vibe_data = get_smart_context(last_user_msg)
-            strategy_data = get_thinking_strategy(is_complex=True, detail=want_detailed)
-
-        elif intent == "coding_request":
-            yield f"data: {json.dumps({'status': 'Analyzing Logic...'})}\n\n"
-            vibe_data = get_smart_context(last_user_msg)
-            strategy_data = get_thinking_strategy(is_complex=True, detail=want_detailed)
+                emoji_min = int(m.group(1))
+                emoji_max = int(m.group(2))
+            except:
+                emoji_min, emoji_max = emoji_min, emoji_max
+
+        # Adjust based on flow context
+        flow_label = flow_context.get("flow_label", "")
+        flow_conf = float(flow_context.get("confidence", 0.0) or 0.0)
+        flow_expl = flow_context.get("explanation", "")
+
+        if flow_label == "escalation":
+            # De-escalate: increase min words, restrict flashy emojis
+            min_words = max(min_words, 40)
+            emoji_min, emoji_max = 0, min(emoji_max, 1)
+            deescalation_note = "Tone: de-escalate and apologize briefly when appropriate."
+        elif flow_label == "small_talk":
+            # Friendly chat can use more emojis
+            emoji_min, emoji_max = max(emoji_min, 1), max(emoji_max, 2)
+            min_words = max(8, min_words//3)
+            deescalation_note = ""
+        else:
+            deescalation_note = ""
 
-        elif intent == "checking_time":
-            yield f"data: {json.dumps({'status': 'Checking Clock...'})}\n\n"
-            time_data = get_time_context()
-            vibe_data = get_smart_context(last_user_msg)
-            strategy_data = get_thinking_strategy(is_complex=False, detail=want_detailed)
+        # If user explicitly asked for short, respect it (override)
+        if re.search(r"\b(short|brief|quick|short and simple)\b", last_user_msg, re.IGNORECASE):
+            min_words = 6
 
-        else:
-            vibe_data = get_smart_context(last_user_msg)
-            strategy_data = get_thinking_strategy(is_complex=False, detail=want_detailed)
+        strategy_data = get_thinking_strategy(is_complex=(intent=="coding_request" or want_detailed), detail=want_detailed, min_words_hint=min_words)
 
         base_system_instruction = (
             "### SYSTEM IDENTITY ###\n"
             "You are Nexari G1, an expressive and helpful AI created by Piyush.\n"
             "### RULES ###\n"
-            "1) If WEB_DATA (search results) is provided, you MUST use it and prioritize it over model-internal knowledge. Cite sources (numbered) when possible.\n"
+            "1) If WEB_DATA (search results) is provided, prioritize it over model-internal knowledge and cite sources.\n"
             "2) Do NOT invent facts when WEB_DATA contradicts model memory.\n"
-            "3) If user asked for detailed/line-by-line output, produce a numbered step-by-step response; aim for thorough coverage.\n"
-            "4) Avoid chain-of-thought; produce a short '🧠 Plan:' (max 2 lines) only for complex tasks, then '💡 Answer:' with final content.\n"
-            "5) Keep emojis to 0-2 per message. After answering, offer a concise follow-up question.\n"
+            "3) Avoid chain-of-thought; provide a short '🧠 Plan:' (max 2 lines) only when complex, then '💡 Answer:' with final content.\n"
+            "4) Use natural phrasing, contractions where appropriate, and follow emoji & verbosity guidance below.\n"
         )
 
-        final_system_prompt = f"{base_system_instruction}\n{flow_desc}\n{vibe_data}\n{time_data}\n{strategy_data}"
+        # Build flow description and emoji/verbosity instructions
+        flow_desc = ""
+        if flow_context:
+            label = flow_context.get("flow_label","unknown")
+            conf = round(float(flow_context.get("confidence", 0.0)), 2)
+            expl = flow_context.get("explanation","")
+            flow_desc = f"\n[FLOW] Detected: {label} (confidence {conf}). {expl}\n"
+
+        emoji_instruction = f"EMOJI GUIDELINE: Use between {emoji_min} and {emoji_max} emoji(s) appropriate to tone; prefer subtle emoji (no more than 2 total)."
+        verbosity_instruction = f"VERBOSITY GUIDELINE: Aim for ~{min_words} words unless user explicitly asked for 'short' or 'brief'. Avoid robotic one-line replies."
+
+        time_data = get_time_context()
+
+        final_system_prompt = f"{base_system_instruction}\n{flow_desc}\n{vibe_data}\n{emoji_instruction}\n{verbosity_instruction}\n{time_data}\n{strategy_data}\n{deescalation_note}"
 
         # Ensure system message present
         if messages[0].get("role") != "system":
@@ -164,6 +192,15 @@ async def generate_response_stream(messages, max_tokens=600, temperature=0.85):
         else:
             messages[0]["content"] = final_system_prompt
 
+        if intent == "internet_search":
+            yield f"data: {json.dumps({'status': 'Searching the web...'})}\n\n"
+            await asyncio.sleep(0)
+            try:
+                tool_data_struct = perform_web_search(last_user_msg)
+            except Exception as e:
+                logger.exception("Web search failed: %s", e)
+                tool_data_struct = {"query": last_user_msg, "results": []}
+
         if tool_data_struct:
             web_block = "### WEB_DATA (from live search) ###\n"
             items = tool_data_struct.get("results", [])
@@ -187,7 +224,6 @@ async def generate_response_stream(messages, max_tokens=600, temperature=0.85):
             yield "data: [DONE]\n\n"
             return
 
-        # Build prompt with tokenizer helper if available, else fallback
         try:
             if hasattr(tokenizer, "apply_chat_template"):
                 text_prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
@@ -230,11 +266,13 @@ async def generate_response_stream(messages, max_tokens=600, temperature=0.85):
             if fc.lower() in cleaned.lower():
                 cleaned = re.sub(re.escape(fc), "I am an AI — expressive and interactive.", cleaned, flags=re.IGNORECASE)
 
-        if want_detailed:
-            if not re.search(r"^\s*\d+[\.\)]\s+", cleaned, re.M):
-                cleaned = "1) " + cleaned.replace("\n", "\n2) ")
+        # If user requested detailed but model didn't produce numbered list, try to format lightly
+        if want_detailed and not re.search(r"^\s*\d+[\.\)]\s+", cleaned, re.M):
+            cleaned = "🧠 Plan: I'll provide step-by-step guidance.\n\n" + cleaned
 
-        cleaned = cleaned.replace("💡 **Answer:**", "\n\n---\n💡 **Answer:**")
+        # If user complained earlier about robotic short answers, encourage small natural touches (this is a gentle nudge in content)
+        # (We rely on system prompt; here only minor post-processing)
+        cleaned = cleaned.replace("I can help with that.", "I can help with that — let me explain. 🙂")
 
         payload = json.dumps({"choices":[{"delta":{"content": cleaned}}]})
         yield f"data: {payload}\n\n"