Update app.py

app.py

@@ -2,7 +2,7 @@ import os
 import re
 import json
 import gradio as gr
-from typing import List, Dict, Any
+from typing import List, Dict, Any, Generator
 import requests
 from datetime import datetime
 import ast
@@ -133,42 +133,76 @@ def download_and_load_model(progress=gr.Progress()):
 def get_tool_descriptions() -> str:
     return "\n".join([f"- {tool.name}: {tool.description}" for tool in TOOLS])
 
-THINK_ONLY_PROMPT = """You are a helpful AI assistant. Solve problems step-by-step.
+THINK_ONLY_PROMPT = """You are an expert problem solver. Use your knowledge and reasoning to answer questions.
 
-Format:
-Thought: your reasoning
-Answer: your final answer
+You must show your complete reasoning process using this format:
+Thought: [Explain what you're thinking and why]
+Thought: [Continue your reasoning, breaking down the problem]
+Thought: [Build toward the solution step by step]
+Answer: [Your final, complete answer]
+
+Important: 
+- Show multiple thought steps
+- Break down complex problems
+- Explain your reasoning clearly
+- Only provide the Answer when you're certain
 
 Question: {question}
 
-Think step by step:"""
+Let me think through this step by step:
+
+Thought:"""
 
-ACT_ONLY_PROMPT = """You are a helpful AI assistant with tools.
+ACT_ONLY_PROMPT = """You are an AI agent with access to external tools. You MUST use tools to find information.
 
 Available tools:
 {tools}
 
+You MUST respond ONLY with actions - no thinking out loud:
+Action: [exact tool name]
+Action Input: [specific input for the tool]
+
+After receiving the Observation, you can:
+- Call another tool if you need more information
+- Provide the final Answer when you have enough information
+
 Format:
 Action: tool_name
-Action Input: input
+Action Input: input_string
+
+Then after observation:
+Action: another_tool
+Action Input: another_input
+
+OR
+Answer: [final answer based on observations]
 
 Question: {question}
 
 Action:"""
 
-REACT_PROMPT = """You are a helpful AI assistant with tools.
+REACT_PROMPT = """You are an expert AI agent that combines reasoning with tool usage (ReAct paradigm).
 
 Available tools:
 {tools}
 
-Pattern:
-Thought: what to do next
-Action: tool_name
-Action Input: input
-Observation: [result]
-... repeat as needed
-Thought: I know the answer
-Answer: final answer
+You MUST alternate between thinking and acting:
+
+1. Thought: [Reason about what information you need and which tool to use]
+2. Action: [exact tool name]
+3. Action Input: [specific input]
+4. Observation: [tool result - will be provided to you]
+5. Thought: [Analyze the observation and decide next steps]
+6. Repeat 2-5 until you have enough information
+7. Thought: [Final reasoning with all gathered information]
+8. Answer: [Complete final answer]
+
+Rules:
+- ALWAYS start with a Thought explaining your strategy
+- After each Observation, think about what you learned
+- Use multiple tools if needed
+- Only give Answer when you have sufficient information
+- Be specific in your Action Inputs
 
 Question: {question}
 
@@ -211,90 +245,115 @@ def call_llm(prompt: str, max_tokens: int = 500) -> str:
     except Exception as e:
         return f"Error during generation: {str(e)}"
 
-def think_only_mode(question: str) -> str:
+def think_only_mode(question: str) -> Generator[str, None, None]:
     if not model_loaded:
-        return "Error: Model not loaded."
+        yield "❌ **Error: Model not loaded. Click 'Download & Load Model' first.**\n\n"
+        return
+    
+    yield "🧠 **Mode: Think-Only (Chain-of-Thought)**\n\n"
+    yield "💭 Generating reasoning steps...\n\n"
     
-    output = "**Mode: Think-Only**\n\n"
     response = call_llm(THINK_ONLY_PROMPT.format(question=question), max_tokens=800)
     
     if response.startswith("Error"):
-        return output + response
+        yield f"❌ {response}\n\n"
+        return
     
     for line in response.split('\n'):
         if line.strip():
-            output += f"**{line.strip()}**\n\n" if line.strip().startswith(('Thought:', 'Answer:')) else f"{line}\n\n"
+            if line.strip().startswith('Thought:'):
+                yield f"💭 **{line.strip()}**\n\n"
+            elif line.strip().startswith('Answer:'):
+                yield f"✅ **{line.strip()}**\n\n"
+            else:
+                yield f"{line}\n\n"
     
-    return output + "\n---\n"
+    yield "\n---\n✓ **Completed**\n"
 
-def act_only_mode(question: str, max_iterations: int = 5) -> str:
+def act_only_mode(question: str, max_iterations: int = 5) -> Generator[str, None, None]:
     if not model_loaded:
-        return "Error: Model not loaded."
+        yield "❌ **Error: Model not loaded. Click 'Download & Load Model' first.**\n\n"
+        return
     
-    output = "**Mode: Act-Only**\n\n"
+    yield "🔧 **Mode: Act-Only (Tool Use Only)**\n\n"
     conversation = ACT_ONLY_PROMPT.format(question=question, tools=get_tool_descriptions())
     
     for iteration in range(max_iterations):
+        yield f"🔄 **Iteration {iteration + 1}**\n\n"
+        
         response = call_llm(conversation, max_tokens=300)
         
         if response.startswith("Error"):
-            return output + response
-        
-        output += f"**Iteration {iteration + 1}:**\n{response}\n\n"
+            yield f"❌ {response}\n\n"
+            return
         
         if 'Answer:' in response:
             match = re.search(r'Answer:\s*(.+)', response, re.IGNORECASE | re.DOTALL)
             if match:
-                output += f"**Final Answer:** {match.group(1).strip()}\n\n"
+                yield f"✅ **Answer:** {match.group(1).strip()}\n\n"
                 break
         
         action_name, action_input = parse_action(response)
         if action_name and action_input:
-            output += f"**Action:** {action_name}\n**Input:** {action_input}\n\n"
+            yield f"🔧 **Action:** `{action_name}`\n"
+            yield f"📝 **Action Input:** {action_input}\n\n"
+            
+            yield f"⏳ Executing tool...\n\n"
             observation = call_tool(action_name, action_input)
-            output += f"**Observation:** {observation}\n\n"
+            
+            yield f"👁️ **Observation:** {observation}\n\n"
             conversation += f"\n{response}\nObservation: {observation}\n\n"
         else:
-            output += f"No valid action found in iteration {iteration + 1}.\n\n"
+            yield f"⚠️ No valid action found. Response: {response}\n\n"
             break
     
-    return output + "\n---\n"
+    yield "\n---\n✓ **Completed**\n"
 
-def react_mode(question: str, max_iterations: int = 5) -> str:
+def react_mode(question: str, max_iterations: int = 5) -> Generator[str, None, None]:
     if not model_loaded:
-        return "Error: Model not loaded."
+        yield "❌ **Error: Model not loaded. Click 'Download & Load Model' first.**\n\n"
+        return
     
-    output = "**Mode: ReAct**\n\n"
+    yield "🤖 **Mode: ReAct (Reasoning + Acting)**\n\n"
     conversation = REACT_PROMPT.format(question=question, tools=get_tool_descriptions())
     
     for iteration in range(max_iterations):
+        yield f"🔄 **Step {iteration + 1}**\n\n"
+        
         response = call_llm(conversation, max_tokens=400)
         
         if response.startswith("Error"):
-            return output + response
-        
-        output += f"**Iteration {iteration + 1}:**\n"
+            yield f"❌ {response}\n\n"
+            return
         
+        # Extract and display thoughts
         for thought in re.findall(r'Thought:\s*(.+?)(?=\n(?:Action:|Answer:|$))', response, re.IGNORECASE | re.DOTALL):
-            output += f"**Thought:** {thought.strip()}\n\n"
+            yield f"💭 **Thought:** {thought.strip()}\n\n"
         
+        # Check for final answer
         if 'Answer:' in response:
             match = re.search(r'Answer:\s*(.+)', response, re.IGNORECASE | re.DOTALL)
             if match:
-                output += f"**Answer:** {match.group(1).strip()}\n\n"
+                yield f"✅ **Answer:** {match.group(1).strip()}\n\n"
                 break
         
+        # Parse and execute action
         action_name, action_input = parse_action(response)
         if action_name and action_input:
-            output += f"**Action:** {action_name}\n**Input:** {action_input}\n\n"
+            yield f"🔧 **Action:** `{action_name}`\n"
+            yield f"📝 **Action Input:** {action_input}\n\n"
+            
+            yield f"⏳ Executing tool...\n\n"
             observation = call_tool(action_name, action_input)
-            output += f"**Observation:** {observation}\n\n"
+            
+            yield f"👁️ **Observation:** {observation}\n\n"
             conversation += f"\n{response}\nObservation: {observation}\n\nThought:"
         else:
-            output += f"No action found in iteration {iteration + 1}.\n{response}\n\n"
+            if 'Answer:' not in response:
+                yield f"⚠️ No action found. Response: {response}\n\n"
             break
     
-    return output + "\n---\n"
+    yield "\n---\n✓ **Completed**\n"
 
 EXAMPLES = [
     "What is 25 * 47?",
@@ -304,15 +363,31 @@ EXAMPLES = [
 ]
 
 def run_comparison(question: str, mode: str):
+    """Run selected mode with real-time streaming."""
+    if not question.strip():
+        yield "Please enter a question.", "", ""
+        return
+    
     if mode == "Think-Only":
-        return think_only_mode(question), "", ""
+        think_out = ""
+        for chunk in think_only_mode(question):
+            think_out += chunk
+            yield think_out, "", ""
+    
     elif mode == "Act-Only":
-        return "", act_only_mode(question), ""
+        act_out = ""
+        for chunk in act_only_mode(question):
+            act_out += chunk
+            yield "", act_out, ""
+    
     elif mode == "ReAct":
-        return "", "", react_mode(question)
-    elif mode == "All (Compare)":
-        return think_only_mode(question), act_only_mode(question), react_mode(question)
-    return "Invalid mode.", "", ""
+        react_out = ""
+        for chunk in react_mode(question):
+            react_out += chunk
+            yield "", "", react_out
+    
+    else:
+        yield "Invalid mode selected.", "", ""
 
 with gr.Blocks(title="LLM Reasoning Modes") as demo:
     gr.Markdown("# LLM Reasoning Modes Comparison\n\n**Model:** openai/gpt-oss-20b\n\n**Tools:** DuckDuckGo | Wikipedia | Weather | Calculator | Python")
@@ -324,7 +399,7 @@ with gr.Blocks(title="LLM Reasoning Modes") as demo:
     with gr.Row():
         with gr.Column(scale=3):
             question_input = gr.Textbox(label="Question", lines=3)
-            mode_dropdown = gr.Dropdown(choices=["Think-Only", "Act-Only", "ReAct", "All (Compare)"], value="Think-Only", label="Mode")
+            mode_dropdown = gr.Dropdown(choices=["Think-Only", "Act-Only", "ReAct"], value="ReAct", label="Mode")
             submit_btn = gr.Button("Run", variant="primary", size="lg")
         with gr.Column(scale=1):
             gr.Markdown("**Examples**")