4.1 Prompting

 

4.1 Prompting 1

Sat, 16 May 26

Course Introduction: Building LLM Applications

• New 8-lesson unit focused on practical LLM application building

• Today’s lesson: Advanced prompt engineering techniques (part 1)

• Previous coverage: LangChain, Python/JavaScript compatibility, workflows

AI Tutor Case Study Framework

• Scenario: Interactive AI tutor for high school students

• Key challenge: Adapting to individual learning styles and preferences

• Focus on personalized education for students struggling with traditional methods

• Will revisit personalization concepts throughout unit

Prompt Engineering Fundamentals

• Definition: Iterative design, testing, and evaluation of LLM instructions

• Goal: Optimize output quality without adding context to model

• Leverages existing model capabilities through clear instructions

• Process: Create → Test → Refine → Repeat until desired output achieved

Basic vs Advanced Prompting

• Basic prompt example: “The ocean is”

• Produces verbose, unfocused responses

• Model uncertainty leads to token waste and higher costs

• Improved prompt: “Complete the sentence: The ocean is”

• More specific instruction yields better results

• Business impact: Unclear prompts = higher costs + poor user experience

Prompt Elements Structure

• Four key components for effective prompts:

• Instruction: Clear directive/task for the model

• Context: Additional grounding information

• Input Data: Specific data points (dates, preferences, constraints)

• Output Indicator: Format and structure requirements

• Japan itinerary example progression:

• Basic: “Create Japan itinerary for cherry blossom season”

• Enhanced: Added solo traveler context, April 1-14 dates, cultural focus, flexible exploration periods

• Result: More detailed, structured output with practical tips

LLM Parameter Settings

• Temperature (0-1 scale):

• Lower (0.2): More deterministic, constrained responses

• Higher (0.8): More creative, exploratory outputs

• Industry default: 0.7 (sweet spot between creativity and constraint)

• Medical/financial use cases: Lower temperatures for compliance

• Top P: Controls response breadth and diversity

• Max Length: Token limit to control costs and verbosity

• Stop Sequence: Defines where model stops generating (prevents endless output)

• Frequency Penalty: Reduces repetition to save tokens and costs

Enterprise Considerations

• Multi-tenancy token management

• Token capping by subscription tier (Free/Pro/Enterprise)

• Cost control through parameter optimization

• Preventing resource over-utilization by single tenants