Advanced Prompting Frameworks (ReAct, ToT)

# CHAPTER 19

Advanced Prompting Frameworks (ReAct, ToT) #

1. Introduction #

2. Learning Objectives #

• Understand the limitation of linear reasoning in LLMs.

• Implement the ReAct (Reason + Act) framework for Agentic AI.

• Explain the Tree of Thoughts (ToT) methodology.

• Identify when to use advanced frameworks vs. standard prompts.

3. Beginner-Friendly Explanation #

4. ReAct (Reasoning and Acting) #

The prompt provides the AI with "Tools" (like Google Search or a Calculator).

The ReAct Prompt Loop:

Task: What is the exact age of the actor who played Iron Man?

[AI GENERATES:]
Thought: I need to find out who played Iron Man.
Action: Search_Google("Who played Iron Man")
Observation: Robert Downey Jr.
Thought: Now I need to find his age.
Action: Search_Google("Robert Downey Jr age")
Observation: 59 years old.
Thought: I have the answer.
Final Answer: Robert Downey Jr. is 59 years old.

*Why it works:* The AI is fact-checking itself in real-time by interacting with the outside world, completely eliminating factual hallucinations.

5. Tree of Thoughts (ToT) #

The ToT Prompt Structure:

Task: Solve this complex logical puzzle.
Step 1: Brainstorm 3 distinct, different possible ways to solve this. (Generate Path A, Path B, Path C).
Step 2: Act as an expert critic. Evaluate the viability of Path A, Path B, and Path C. Score each out of 10.
Step 3: Discard the lowest scoring paths. Take the highest scoring path and expand it into a final, detailed solution.

*Why it works:* By forcing the AI to critique its own ideas *before* committing to an answer, you simulate high-level human problem-solving.

6. Meta-Prompting (Self-Correction) #

# Prompt 1
Write a Python script to scrape a website.

# Prompt 2
Review the script you just generated. Identify any security flaws, inefficiencies, or syntax errors. 
Output a revised, optimized version of the script.

Never accept the AI's first draft of code. Always pass it through a Meta-Prompting review.

7. Python Example: The ReAct Loop #

# Conceptual ReAct Loop
while True:
    # Get AI's next thought/action
    response = call_llm(prompt_history)
    
    if "Final Answer" in response:
        print(response)
        break
        
    elif "Action:" in response:
        # Extract the command (e.g., Search_Google)
        action = extract_action(response)
        
        # Execute the Python tool to search the internet
        observation = execute_tool(action)
        
        # Feed the real-world observation back to the AI
        prompt_history += f"Observation: {observation}\nThought: "

8. Mini Project #

9. Best Practices #

• Token Awareness: Advanced frameworks (especially ToT) burn through massive amounts of tokens because the AI generates paragraphs of brainstorming and critique before giving the final answer. Use these frameworks *only* for highly complex tasks, not simple requests.

10. Common Mistakes #

• Infinite Loops in ReAct: If an AI Agent is poorly prompted, it might get stuck in an infinite loop (e.g., searching the same Google phrase over and over and getting confused). Developers must implement "Max Iteration" limits (e.g., stop the loop after 5 actions) to prevent the AI from running forever and racking up API bills.

11. Exercises #

1. 1. Explain the difference between linear "Chain-of-Thought" prompting and the branching logic of "Tree of Thoughts."

12. MCQs with Answers #

13. Interview Questions #

• Q: Describe the ReAct framework. How does integrating real-time "Observations" from external tools solve the LLM hallucination problem?

• Q: In an enterprise environment, when would you justify the high token cost of using a Tree of Thoughts (ToT) prompt architecture over a simple Chain-of-Thought (CoT) prompt?

14. FAQs #

15. Summary #

16. Next Chapter Recommendation #