Table of Contents

Enum PromptOptimizationStrategy

Namespace
AiDotNet.Enums
Assembly
AiDotNet.dll

Represents strategies for optimizing prompts to improve language model performance.

public enum PromptOptimizationStrategy

Fields

DiscreteSearch = 1

Discrete optimization that tests variations of prompt components.

For Beginners: Discrete optimization tries different versions of prompt parts and picks what works best.

Think of it like A/B testing:

  • Create multiple variations of different parts
  • Test each combination
  • Measure which performs best
  • Use the winning combination

Example - Optimizing a classification prompt:

Component 1 - Instruction phrase: A: "Classify the following" B: "Categorize this text as" C: "Determine if this is"

Component 2 - Output format: A: "Respond with just the category" B: "Output format: [CATEGORY]" C: "Answer in one word"

Component 3 - Examples: A: 2 examples B: 5 examples C: 10 examples

The system tests combinations:

  • Instruction A + Format A + Examples A
  • Instruction A + Format B + Examples A
  • ...and so on

After testing, it finds: Best combination: Instruction B + Format C + Examples B = "Categorize this text as... Answer in one word... [5 examples]"

Advantages:

  • Systematic exploration of prompt variations
  • Finds good combinations of components
  • Interpretable results (you know what changed)
  • Works well for structured prompts

Disadvantages:

  • Can be slow with many components
  • Only tests pre-defined variations
  • May miss optimal wordings not in the variation set

Use this when:

  • You have ideas for prompt variations to test
  • Prompt has clear, separable components
  • You want to understand what makes prompts work
  • You have a good evaluation metric
Ensemble = 3

Ensemble multiple prompts and combine their outputs for better performance.

For Beginners: Ensemble optimization uses multiple different prompts and combines their results.

Think of it like asking multiple experts:

  • Expert 1 gives their opinion
  • Expert 2 gives their opinion
  • Expert 3 gives their opinion
  • You combine all opinions to reach a conclusion

Often, the combined opinion is better than any single expert.

How it works:

  1. Create multiple diverse prompts (or use the best from optimization)
  2. Send the query to the model with each prompt
  3. Get multiple responses
  4. Combine responses (voting, averaging, or consensus)

Example - Sentiment classification:

Prompt 1 (formal): "Analyze the sentiment of the following text. Classify as positive, negative, or neutral." → Output: "Positive"

Prompt 2 (instructive): "You are a sentiment analysis expert. Categorize this review." → Output: "Positive"

Prompt 3 (with examples): [Few-shot examples] "Now classify: [text]" → Output: "Positive"

Prompt 4 (chain-of-thought): "Think step-by-step about the sentiment of this text." → Output: "The text mentions 'loved it' and 'amazing', which are positive. Negative words are absent. Therefore: Positive"

Combined result: Positive (4/4 agree) → High confidence

For more complex tasks:

  • If prompts disagree, use majority voting
  • Weight prompts by their historical accuracy
  • Use the most confident prediction

Advantages:

  • More robust than single prompts
  • Reduces impact of prompt brittleness
  • Can provide confidence estimates
  • Often better accuracy

Disadvantages:

  • More expensive (multiple API calls)
  • Slower response time
  • Complex to combine results for generative tasks

Use this when:

  • Accuracy is critical
  • Cost and speed are acceptable trade-offs
  • Single prompts are too brittle
  • You need confidence estimates
Evolutionary = 5

Evolutionary optimization using genetic algorithms to evolve better prompts.

For Beginners: Evolutionary optimization mimics biological evolution to improve prompts.

How evolution works in nature:

  1. Start with a population of individuals (genetic variation)
  2. Individuals with better traits survive (natural selection)
  3. Survivors reproduce, creating new individuals (crossover)
  4. Random mutations add new variations
  5. Repeat for many generations
  6. Eventually, you get highly adapted individuals

For prompts:

  1. Create a "population" of different prompts
  2. Test each prompt's performance (fitness)
  3. Keep the best-performing prompts
  4. Create new prompts by combining parts of good prompts (crossover)
  5. Add random changes to some prompts (mutation)
  6. Repeat for many generations

Example: Generation 1 - Random population:

  • Prompt A: "Classify this text" → 75% accuracy
  • Prompt B: "Categorize the sentiment" → 82% accuracy
  • Prompt C: "Determine the category" → 70% accuracy
  • Prompt D: "Analyze and classify this text" → 85% accuracy

Selection: Keep B and D (best performers)

Generation 2 - Crossover (combine parts of B and D):

  • New E: "Categorize and analyze this text" (from B + D)
  • New F: "Analyze the sentiment" (from D + B)

Mutation (random changes):

  • New G: "Carefully analyze the sentiment" (added "carefully")
  • New H: "Categorize the emotional tone" (changed "sentiment" to "emotional tone")

Test Generation 2:

  • Prompt E: 87% accuracy
  • Prompt F: 83% accuracy
  • Prompt G: 84% accuracy
  • Prompt H: 90% accuracy ← Best so far!

Keep best from Gen 2, create Gen 3, and continue...

After many generations: Final prompt: "Carefully categorize the specific emotional tone and sentiment of the following text" → 94% accuracy

Advantages:

  • Can find creative solutions
  • Balances exploration and exploitation
  • Doesn't require gradients
  • Often finds good solutions

Disadvantages:

  • Requires many evaluations
  • Can be slow
  • Complex to implement well
  • Results can be unpredictable

Use this when:

  • You want to explore creative prompt variations
  • You have evaluation budget
  • Other methods haven't worked well
  • You enjoy bio-inspired algorithms
GradientBased = 2

Gradient-based optimization using automatic differentiation (APE - Automatic Prompt Engineering).

For Beginners: Gradient-based optimization automatically finds better prompts using math, similar to how neural networks learn.

Think of it like a robot learning to navigate:

  • Instead of trying random directions
  • It calculates which direction leads "downhill" toward the goal
  • It takes steps in that direction
  • It repeats until it reaches the destination

For prompts:

  • Start with an initial prompt
  • Calculate how to change it to improve performance
  • Make small adjustments
  • Repeat until performance is good enough

This uses "soft prompts" or "continuous prompts" - mathematical representations that can be optimized with gradients, then converted back to text.

Example process: Initial: "Classify this text" → Calculate gradient (math indicating how to improve) → Adjust toward "Categorize the sentiment of this text as" → Evaluate performance → Continue adjusting... Final: "Analyze and categorize the emotional tone of the following text as"

Advantages:

  • Can find prompts you wouldn't think of manually
  • Efficient optimization process
  • Explores a continuous space of possibilities
  • State-of-the-art performance in research

Disadvantages:

  • Complex to implement
  • Requires differentiable models
  • May produce unnatural-sounding prompts
  • Harder to interpret why it works

Use this when:

  • You want cutting-edge performance
  • You have computational resources
  • You can use gradient-based optimization
  • Quality is more important than interpretability
MonteCarlo = 4

Monte Carlo optimization that randomly samples and tests prompt variations.

For Beginners: Monte Carlo optimization tries many random variations and keeps what works.

Think of it like:

  • Throwing darts at a dartboard
  • Some hit the bullseye, some miss
  • After many throws, you learn where to aim
  • You focus on the areas that work best

For prompts:

  1. Generate random variations of the prompt
  2. Test each variation's performance
  3. Keep track of what works
  4. Generate more variations around successful ones
  5. Repeat until performance is good enough

Example: Base prompt: "Classify this text"

Random variations generated:

  1. "Categorize the following text"
  2. "What is the category of this text?"
  3. "Classify the sentiment of this text"
  4. "Determine the classification of this text"
  5. "Please classify this text into a category" ...50 more variations...

Test all variations: Variation 3 ("Classify the sentiment of this text") → 92% accuracy Variation 5 ("Please classify this text into a category") → 89% accuracy Others: 75-85% accuracy

Generate more variations similar to #3:

  • "Analyze and classify the sentiment of this text"
  • "Classify the emotional tone of this text"
  • etc.

Continue until optimal performance is found.

Advantages:

  • Simple to implement
  • No gradient computation needed
  • Can escape local optima (bad solutions)
  • Good for exploring large spaces

Disadvantages:

  • Requires many evaluations
  • Can be slow and expensive
  • May miss optimal solutions
  • No guarantee of finding the best prompt

Use this when:

  • You can't use gradient-based methods
  • You have computational budget for many tests
  • The optimization space is complex
  • You want a simple, robust approach
None = 0

No optimization - use the prompt as provided.

For Beginners: This means using your prompt exactly as you wrote it, without any optimization.

Use this when:

  • You've already manually crafted a great prompt
  • Speed is critical and you can't afford optimization time
  • You want complete control over the exact prompt text
  • You're in early development and want to test specific prompt ideas

Remarks

For Beginners: Prompt optimization strategies automatically improve prompts to get better results.

Think of it like tuning a recipe:

  • You start with a basic recipe
  • Try different variations (more salt? less sugar? different temperature?)
  • Taste test each variation
  • Keep the version that tastes best

Prompt optimization does the same thing:

  • Start with a basic prompt
  • Generate variations
  • Test each variation's performance
  • Keep the best-performing version

Different strategies use different approaches to search for better prompts.

Edit this page