4.2 Prompt Pt 2 - Course Notes

 

4.2 Prompt Pt 2 - Gen AI: Building LLM Apps

Sun, 17 May 26

Course Overview & Lab Setup Issues

• Advanced prompt engineering techniques (Part 2) for building LLM applications

• Multiple lab access problems across students

• API key failures in Colab

• Virtual environment setup difficulties

• Missing demo files (Demo 5 unavailable for some students)

• Instructor using alternative lab instance due to access issues

Chain of Thought (CoT) Prompting

• Breaks complex problems into sequential subtasks vs direct input-output

• Benefits over zero/few-shot approaches:

• Easier reasoning for lightweight models

• Better context window utilization

• Reduces hallucination risk

• Implementation signals:

• “Think through the problem step by step”

• “Break down into steps”

• Modern LLMs automatically implement CoT internally

• Can combine with zero-shot or few-shot prompting

Self-Consistency Prompting

• Built on CoT to improve reasoning consistency

• Process:

• Ask same question multiple times

• Model takes diverse reasoning paths

• Apply majority voting to final answers

• Addresses probabilistic nature of LLMs

• Benefits:

• Improved accuracy through multiple perspectives

• Reduced bias via diverse reasoning

• Enhanced critical thinking

Tree of Thought (ToT) Prompting

• Multiple chain-of-thought branches like tree structure

• Key features:

• Different strategies per branch

• Backtracking capability to previous steps

• Graph traversal approach to problem solving

• More robust than sequential CoT

• Example: Finding number 24 using digits 4, 9, 10, 13

• Multiple arithmetic operation combinations

• Different starting number pairs per branch

LangChain Prompt Templates

• Reusable, parameterized prompt structures

• Template types:

• Basic PromptTemplate - simple variable substitution

• ChatPromptTemplate - structured chat interactions

• Custom templates using string formatting

• Benefits:

• Consistency across applications

• Dynamic variable insertion

• Structured predefined formats

Jinja2 Template Integration

• Advanced templating engine for dynamic content generation

• Features:

• Conditional logic support

• Variable substitution

• Age-appropriate content adaptation (kids vs adult events)

• Use cases:

• Automated report generation

• Personalized invitations

• Flask/Django web applications

Lab Environment Setup Requirements

• Virtual environment creation: python3 -m venv myenv

• Activation: source myenv/bin/activate

• Install requirements.txt (contains all dependencies)

• Kernel selection: Must use “myenv” not default Python 3

• API key setup only needed for Colab users

• Simply Learn platform provides pre-configured API access

Next Steps & Resources

• No class next Saturday (May 24, 2026)

• Next session: May 30, 2026

• Students should practice demos independently

• Experiment with different:

• Temperature settings (0 for deterministic, 0.7 for creative)

• Model types (GPT-3.5-turbo vs GPT-4o-mini)

• Custom prompts and variables

• Use AI (Gemini) to learn AI - ask for parameter explanations and improvements