Update app.py

app.py

@@ -32,7 +32,7 @@ from gradio_client import Client
 # ============================================================
 
 APP_NAME = "X-RUDRA"
-VERSION = "3.7.3"   # bumped
+VERSION = "3.8.1"
 
 M1_REPO = os.getenv("M1_REPO", "Shrijanagain/M1")
 M2_REPO = os.getenv("M2_REPO", "Shrijanagain/M2")
@@ -74,29 +74,74 @@ def get_m2_client():
 
 
 # ============================================================
-# CASUAL QUERY DETECTOR
+# MODEL-BASED INTENT CLASSIFIER
+# ──────────────────────────────────────────────────────────────
+# Uses M1 (or M2) to decide ACTION or CASUAL.
+# No hardcoded keywords – the model decides.
+# If both fail, defaults to ACTION.
 # ============================================================
 
-def is_casual_query(text: str) -> bool:
-    text = text.lower().strip()
-    words = text.split()
-    if len(words) <= 2:
-        return True
-    question_words = {"what", "how", "why", "when", "where", "who", "which",
-                      "can", "could", "would", "will", "is", "are", "do", "does"}
-    if words[0] in question_words and len(words) >= 3:
-        return False
-    casual_patterns = [
-        "hey", "hi", "hello", "yo", "what's up", "how are you",
-        "good morning", "good evening", "good night",
-        "lol", "haha", "just testing", "timepass", "nothing",
-        "tell me a joke", "sing a song", "what's your name",
-        "who are you", "what can you do", "help", "thanks"
-    ]
-    for pattern in casual_patterns:
-        if text == pattern or text.startswith(pattern + " "):
-            return True
-    return False
+CLASSIFIER_SYSTEM_PROMPT = """
+You are an intelligent assistant that classifies user messages into two categories:
+- ACTION: The user asks for information that requires research, fact‑checking, retrieval of current data, or external knowledge. This includes questions about news, comparisons, statistics, history, technology, science, politics, etc.
+- CASUAL: The user is just chatting, greeting, making small talk, joking, or asking a simple question that can be answered from general knowledge without a search.
+
+Respond with ONLY ONE WORD: ACTION or CASUAL.
+Do NOT add any extra text, punctuation, or explanation.
+"""
+
+def classify_intent(question: str) -> str:
+    prompt = f"{CLASSIFIER_SYSTEM_PROMPT}\n\nUser message: \"{question}\"\n\nClassification:"
+    
+    # Try M1 first
+    m1_client = get_m1_client()
+    if m1_client is not None:
+        try:
+            result = m1_client.predict(
+                prompt=prompt,
+                max_tokens=10,
+                temperature=0.0,
+                api_name="/generate"
+            )
+            if result:
+                result = result.strip().upper()
+                if "ACTION" in result:
+                    print(f"[Classifier] M1 → ACTION")
+                    return "ACTION"
+                elif "CASUAL" in result:
+                    print(f"[Classifier] M1 → CASUAL")
+                    return "CASUAL"
+                else:
+                    print(f"[Classifier] M1 ambiguous: {result}")
+        except Exception as e:
+            print(f"[Classifier] M1 call failed: {e}")
+
+    # Try M2 as fallback
+    m2_client = get_m2_client()
+    if m2_client is not None:
+        try:
+            result = m2_client.predict(
+                prompt=prompt,
+                max_tokens=10,
+                temperature=0.0,
+                api_name="/generate"
+            )
+            if result:
+                result = result.strip().upper()
+                if "ACTION" in result:
+                    print(f"[Classifier] M2 → ACTION")
+                    return "ACTION"
+                elif "CASUAL" in result:
+                    print(f"[Classifier] M2 → CASUAL")
+                    return "CASUAL"
+                else:
+                    print(f"[Classifier] M2 ambiguous: {result}")
+        except Exception as e:
+            print(f"[Classifier] M2 call failed: {e}")
+
+    # Default to ACTION if classification fails
+    print("[Classifier] Defaulting to ACTION")
+    return "ACTION"
 
 
 # ============================================================
@@ -182,7 +227,7 @@ Answer:"""
         else:
             return "I couldn't find specific information on that topic. Could you rephrase?", ""
 
-    # If only one draft exists, use that as final (extract thinking if any)
+    # If only one draft exists, use that as final
     if not draft_m1:
         thinking = ""
         clean = draft_m2
@@ -214,17 +259,14 @@ Combine these two drafts into a single, comprehensive, accurate, and natural ans
 
 Final answer:"""
 
-    # Use a higher token limit for the merge step
-    merge_max_tokens = max(1024, max_tokens * 2)  # at least 1024, double the original
+    merge_max_tokens = max(1024, max_tokens * 2)
     merged = call_model(m2_client, merge_prompt, merge_max_tokens, temperature)
 
-    # If merged is too short (incomplete), retry with even more tokens
     if merged and len(merged) < 100:
         print(f"Merged answer too short ({len(merged)} chars), retrying with 2048 tokens...")
         merged = call_model(m2_client, merge_prompt, 2048, temperature)
 
     if not merged:
-        # Fallback to draft_m1 if merge fails completely
         print("Merge failed, falling back to draft_m1")
         merged = draft_m1
 
@@ -249,7 +291,7 @@ def get_casual_model_response(query: str) -> str:
         try:
             result = client.predict(
                 prompt=f"User: {query}\nAssistant:",
-                max_tokens=64,
+                max_tokens=128,
                 temperature=0.7,
                 api_name="/generate"
             )
@@ -264,7 +306,7 @@ def get_casual_model_response(query: str) -> str:
         try:
             result = client.predict(
                 prompt=f"User: {query}\nAssistant:",
-                max_tokens=64,
+                max_tokens=128,
                 temperature=0.7,
                 api_name="/generate"
             )
@@ -409,7 +451,7 @@ def safe_dict(value):
 
 
 # ============================================================
-# MAIN RESEARCH FUNCTION (returns 6 outputs)
+# MAIN RESEARCH FUNCTION
 # ============================================================
 
 async def do_research(question, max_results, max_rounds, use_models, freshness):
@@ -425,8 +467,12 @@ async def do_research(question, max_results, max_rounds, use_models, freshness):
 
     question = str(question).strip()
 
-    # ---- Casual query ----
-    if is_casual_query(question):
+    # ---- Step 1: Classify intent using M1 (or M2) ----
+    intent = classify_intent(question)
+    print(f"[Intent] {intent} for: {question}")
+
+    # ---- Step 2: If casual, reply directly ----
+    if intent == "CASUAL":
         answer = get_casual_model_response(question)
         history = [
             {"role": "user", "content": question},
@@ -434,7 +480,7 @@ async def do_research(question, max_results, max_rounds, use_models, freshness):
         ]
         return history, "⚡ Casual chat (model reply, no search).", "", "", "", ""
 
-    # ---- Serious query ----
+    # ---- Step 3: ACTION – run research pipeline ----
     started = time.perf_counter()
     try:
         engine = get_engine()
@@ -464,7 +510,7 @@ async def do_research(question, max_results, max_rounds, use_models, freshness):
                     })
             data["sources"] = sources
 
-        # Generate final answer and thinking content (tags already stripped)
+        # Generate final answer and thinking
         final_answer, thinking_content = get_combined_model_answer(question, sources)
 
         sources_md = format_sources(sources)
@@ -518,7 +564,7 @@ def health_check():
 
 
 # ============================================================
-# CSS – includes spinner animation for thinking
+# CSS
 # ============================================================
 
 CSS = """