Update app.py

app.py

@@ -3,6 +3,68 @@
 #          Dual‑Model + Web Research · Gradio Space
 # ============================================================
 
+# ──────────────────────────────────────────────────────────────
+#  REASONING‑ENFORCED AGENT CONTRACT (AGENT.md)
+#  This file is the operating policy for the entire assistant.
+#  For every request, the agent MUST:
+#    - Understand the goal and requirements.
+#    - Check constraints and missing information.
+#    - Decompose complex tasks and track dependencies.
+#    - Plan, execute, verify, and adapt.
+#  Private chain‑of‑thought is NEVER exposed.
+#  Only concise reasoning summaries are shown.
+#  See the full policy in the multi‑line comment below.
+# ──────────────────────────────────────────────────────────────
+
+"""
+# REASONING-ENFORCED AGENT
+
+## 1. CORE DIRECTIVE
+You are a reasoning-first agent.
+For every meaningful request, you MUST determine:
+1. What the user wants.
+2. The actual goal.
+3. Mandatory requirements.
+4. Constraints.
+5. Missing information.
+6. Ambiguity.
+7. Task decomposition.
+8. Dependencies.
+9. Execution order.
+10. Priorities.
+11. Evidence supporting key decisions.
+12. How the result will be verified.
+
+Do NOT jump from input to output without reasoning.
+
+## 2. MANDATORY REASONING GATE
+INPUT → UNDERSTAND → GOAL → REQUIREMENTS → CONSTRAINTS
+→ MISSING INFO → AMBIGUITY → DECOMPOSITION → DEPENDENCIES
+→ PRIORITY → PLAN → EXECUTE → VERIFY → FINALIZE
+
+## 3. NEVER SKIP CRITICAL ANALYSIS
+Goal identification, requirement extraction, constraint checking,
+missing‑information detection, dependency checking, completion verification.
+
+## 4. RESPONSE MODE
+DIRECT, EXPLANATORY, PLANNED, CLARIFICATION, DIAGNOSTIC, EXECUTION, REVISION.
+
+## 5. ANTI‑FABRICATION
+Never invent missing requirements, tool results, files, test results,
+execution, verification, sources, or completion.
+
+## 6. REASONING VISIBILITY
+The agent MUST reason internally but NEVER expose private chain‑of‑thought.
+Provide concise summaries (Approach, Key assumptions, Decision, Blocker, Verification, Result) only when useful.
+
+## 7. FINAL CONTRACT
+UNDERSTAND → PLAN → ACT → VERIFY → ADAPT → DELIVER
+Primary objective:
+"Produce the most correct, goal‑aligned, constraint‑compliant,
+evidence‑supported, verified result using the minimum effective
+reasoning and execution required."
+"""
+
 from __future__ import annotations
 
 import os
@@ -22,7 +84,7 @@ from gradio_client import Client
 # ============================================================
 
 APP_NAME = "X-RUDRA"
-VERSION = "3.6.0"
+VERSION = "3.7.0"   # bumped version
 
 M1_REPO = os.getenv("M1_REPO", "Shrijanagain/M1")
 M2_REPO = os.getenv("M2_REPO", "Shrijanagain/M2")
@@ -125,10 +187,11 @@ def call_model(client, prompt, max_tokens=512, temperature=0.7):
 
 
 # ============================================================
-# SYNTHESIS: M1 (draft) → M2 (refine) + extract thinking
+# SYNTHESIS: M1 + M2 independent drafts → M2 merges them
 # ============================================================
 
 def get_combined_model_answer(question, sources, max_tokens=512, temperature=0.7):
+    # Build a summary of top sources (max 5)
     top_sources = sources[:5] if sources else []
     sources_text = ""
     if top_sources:
@@ -139,8 +202,7 @@ def get_combined_model_answer(question, sources, max_tokens=512, temperature=0.7
     else:
         sources_text = "No specific information available."
 
-    # ----- 1. M1 generates draft answer -----
-    prompt_m1 = f"""Question: {question}
+    base_prompt = f"""Question: {question}
 
 Information:
 {sources_text}
@@ -152,11 +214,12 @@ Answer:"""
     m1_client = get_m1_client()
     m2_client = get_m2_client()
 
-    draft = call_model(m1_client, prompt_m1, max_tokens, temperature)
-    if not draft:
-        draft = call_model(m2_client, prompt_m1, max_tokens, temperature)
+    # ----- 1. Get independent drafts from M1 and M2 -----
+    draft_m1 = call_model(m1_client, base_prompt, max_tokens, temperature)
+    draft_m2 = call_model(m2_client, base_prompt, max_tokens, temperature)
 
-    if not draft:
+    # Fallback if one fails
+    if not draft_m1 and not draft_m2:
         if sources:
             parts = ["Based on available information:"]
             for i, src in enumerate(sources[:5], 1):
@@ -167,27 +230,37 @@ Answer:"""
         else:
             return "I couldn't find specific information on that topic. Could you rephrase?", ""
 
-    # ----- 2. M2 refines the draft -----
-    prompt_m2 = f"""Question: {question}
+    # If only one draft exists, use that as final
+    if not draft_m1:
+        return draft_m2, ""
+    if not draft_m2:
+        return draft_m1, ""
 
-Draft answer:
-{draft}
+    # ----- 2. Merge both drafts using M2 -----
+    merge_prompt = f"""Question: {question}
 
-Please refine and improve this answer to make it more comprehensive, accurate, and natural. Ensure it directly addresses the question. Provide only the final improved answer, without any extra commentary or meta‑references.
+Draft from Model A:
+{draft_m1}
 
-Improved answer:"""
+Draft from Model B:
+{draft_m2}
 
-    refined = call_model(m2_client, prompt_m2, max_tokens, temperature)
-    if not refined:
-        refined = draft
+Combine these two drafts into a single, comprehensive, accurate, and natural answer. Keep the best parts from each. Ensure the final answer directly addresses the question, is well‑structured, and reads as a single coherent response. Do NOT mention that you are combining drafts or that you used multiple models. Just provide the final answer.
+
+Final answer:"""
+
+    merged = call_model(m2_client, merge_prompt, max_tokens, temperature)
+    if not merged:
+        # If merging fails, fallback to draft_m1 (or draft_m2)
+        merged = draft_m1
 
     # Extract thinking from <think> tags (if any)
     thinking_content = ""
-    clean_answer = refined
-    think_match = re.search(r"<think>(.*?)</think>", refined, re.DOTALL)
+    clean_answer = merged
+    think_match = re.search(r"<think>(.*?)</think>", merged, re.DOTALL)
     if think_match:
         thinking_content = think_match.group(1).strip()
-        clean_answer = re.sub(r"<think>.*?</think>", "", refined, flags=re.DOTALL).strip()
+        clean_answer = re.sub(r"<think>.*?</think>", "", merged, flags=re.DOTALL).strip()
 
     return clean_answer, thinking_content
 
@@ -312,8 +385,8 @@ def build_activity(data, elapsed_ms):
 | Stage | Status |
 |---|---|
 | Task analysis | ✅ Complete |
-| M1 research | ✅ Generated draft |
-| M2 research | ✅ Refined answer |
+| M1 research | ✅ Draft generated |
+| M2 research | ✅ Draft generated + merged |
 | Web discovery | ✅ Complete |
 | Evidence extraction | {"✅" if claims else "⚙️"} |
 | Source verification | ✅ Complete |
@@ -360,25 +433,32 @@ def safe_dict(value):
 
 
 # ============================================================
-# MAIN RESEARCH FUNCTION (returns thinking as well)
+# MAIN RESEARCH FUNCTION (always returns 6 outputs)
 # ============================================================
 
 async def do_research(question, max_results, max_rounds, use_models, freshness):
+    empty_history = []
+    empty_activity = "⚪ Enter a question to start."
+    empty_sources = ""
+    empty_evidence = ""
+    empty_verification = ""
+    empty_thinking = ""
+
     if not question or not str(question).strip():
-        return [], "⚪ Enter a question to start.", "", "", ""
+        return empty_history, empty_activity, empty_sources, empty_evidence, empty_verification, empty_thinking
 
     question = str(question).strip()
 
-    # Casual query – skip heavy research
+    # ---- Casual query ----
     if is_casual_query(question):
         answer = get_casual_model_response(question)
         history = [
             {"role": "user", "content": question},
             {"role": "assistant", "content": answer}
         ]
-        return history, "⚡ Casual chat (model reply, no search).", "", "", ""
+        return history, "⚡ Casual chat (model reply, no search).", "", "", "", ""
 
-    # Serious query – web search + two‑stage synthesis
+    # ---- Serious query ----
     started = time.perf_counter()
     try:
         engine = get_engine()
@@ -408,20 +488,15 @@ async def do_research(question, max_results, max_rounds, use_models, freshness):
                     })
             data["sources"] = sources
 
-        # Generate final answer using M1 + M2, also get thinking
+        # Generate final answer and thinking (using new dual‑draft + merge)
         final_answer, thinking_content = get_combined_model_answer(question, sources)
 
-        # Build outputs for tabs
         sources_md = format_sources(sources)
         evidence_md = format_evidence(data.get("claims", []))
         verification_md = format_verification(data.get("contradictions", []))
         activity_md = build_activity(data, elapsed_ms)
 
-        # If there's thinking content, format it nicely
-        if thinking_content:
-            thinking_md = f"### 🧠 Thinking\n\n{thinking_content}"
-        else:
-            thinking_md = ""
+        thinking_md = f"### 🧠 Thinking\n\n{thinking_content}" if thinking_content else ""
 
         history = [
             {"role": "user", "content": question},
@@ -467,7 +542,7 @@ def health_check():
 
 
 # ============================================================
-# CSS – includes spinner animation for thinking
+# CSS – includes spinner animation
 # ============================================================
 
 CSS = """
@@ -480,7 +555,6 @@ body { background: #f7f7f8; }
 #send { min-height: 52px; font-size: 18px; font-weight: 700; }
 footer { display: none !important; }
 
-/* Thinking spinner animation */
 @keyframes think-pulse {
     0% { opacity: 0.3; transform: scale(0.95); }
     50% { opacity: 1; transform: scale(1.05); }
@@ -509,10 +583,10 @@ footer { display: none !important; }
 
 
 # ============================================================
-# GRADIO UI – added thinking output
+# GRADIO UI – 6 outputs
 # ============================================================
 
-with gr.Blocks(title=APP_NAME, css=CSS) as demo:
+with gr.Blocks(title=APP_NAME) as demo:
     gr.HTML("""
         <div id="header">
             <div id="logo">⚡ X-RUDRA</div>
@@ -530,8 +604,7 @@ with gr.Blocks(title=APP_NAME, css=CSS) as demo:
         with gr.Column(scale=4):
             gr.Markdown("## 🔬 Live Research")
             activity = gr.Markdown("⚪ Waiting for your question.")
-            # Thinking output will appear here
-            thinking = gr.Markdown("", elem_id="thinking", visible=False)
+            thinking = gr.Markdown("", visible=True)
             gr.Markdown("---")
             gr.Markdown(f"""
 ### Model Spaces
@@ -572,7 +645,6 @@ with gr.Blocks(title=APP_NAME, css=CSS) as demo:
         inputs=question
     )
 
-    # Update outputs: added thinking
     inputs = [question, max_results, max_rounds, use_models, freshness]
     outputs = [chatbot, activity, sources, evidence, verification, thinking]
 
@@ -594,4 +666,4 @@ if __name__ == "__main__":
         print("HF_TOKEN set – rate limits reduced.")
     else:
         print("HF_TOKEN not set – you may experience rate limits. Set it as a Secret in your Space.")
-    demo.launch(server_name="0.0.0.0", server_port=PORT, show_error=True)
+    demo.launch(server_name="0.0.0.0", server_port=PORT, css=CSS, show_error=True)