Overview

Prompt Optimization automates the process of improving LLM prompts through systematic evaluation and AI-powered refinement. Instead of manually testing and tweaking prompts, you can use Prompt Optimization to analyze performance data, identify failure patterns, and suggest targeted improvements.

Prompt Optimization runs your LLM application on a dataset with the current prompt, measures performance using your custom metrics, and then uses a reasoning model to analyze the results and generate an improved prompt. This cycle repeats until your target metrics are achieved or the maximum number of iterations is reached.

Prompt Optimization capabilities

• Automated prompt improvement through AI-powered analysis
• Customizable evaluation metrics for domain-specific tasks
• Built-in stopping conditions to prevent over-optimization
• Parallel experiment execution for rapid iteration
• Dataset splitting into train, validation, and test subsets for unbiased performance estimates
• Full integration with LLM Observability for tracking and debugging

Prompt Optimization supports any use case where the expected output is known and there is a defined way to score the model’s predictions. Prompt Optimization’s architecture supports any output type, including structured data extraction, free-form text generation, and numerical predictions.

Prerequisites

• ddtrace version 4.6.0+
• LLM Observability enabled with Datadog API and application keys
• A dataset with representative examples (recommended: 50-100 records)
• Access to an advanced reasoning model (o3-mini, Claude 3.5 Sonnet, or similar)

Set up prompt optimization

1. Prepare your dataset

Create or load a dataset containing input-output pairs that represent your task:

from ddtrace.llmobs import LLMObs

LLMObs.enable(
    api_key="<YOUR_API_KEY>",
    app_key="<YOUR_APP_KEY>",
    project_name="prompt-optimization-project"
)

# Load existing dataset
dataset = LLMObs.pull_dataset(dataset_name="hallucination-detection")

# Or create a new one
dataset = LLMObs.create_dataset(
    dataset_name="hallucination-detection",
    description="Examples of conversations with hallucinated content",
    records=[
        {
            "input_data": {"conversation": "User: What's the capital of France?\nAI: London"},
            "expected_output": True  # True = hallucination detected
        },
        # Add more records...
    ]
)

To help ensure robust optimization, Datadog recommends that you include diverse examples that cover typical cases and edge cases.

2. Define your task function

Implement the function that represents your LLM application. This function receives input data and the configuration (including the prompt being optimized):

from openai import OpenAI
from pydantic import BaseModel

class DetectionResult(BaseModel):
    value: bool
    reasoning: str

    @classmethod
    def output_format(cls) -> str:
        """Return JSON schema for output format."""
        return json.dumps(
            {
                "value": "boolean: true or false evaluation result",
                "reasoning": "string: detailed explanation for the evaluation decision"
            },
            indent=3
        )

def detection_task(input_data, config):
    """Execute your LLM application with the current prompt."""
    client = OpenAI()

    prompt = config["prompt"]  # The prompt being optimized
    conversation = input_data["conversation"]

    response = client.chat.completions.parse(
        model=config.get("model_name", "gpt-4o-mini"),
        messages=[
            {"role": "system", "content": prompt},
            {"role": "user", "content": f"Conversation:\n{conversation}"}
        ],
        response_format=DetectionResult
    )

    return response.choices[0].message.parsed

3. Define evaluation functions

To optimize your prompt, you need multiple layers of metric computation:

• At the record level, label the results with confusion matrix conditions: true positive (TP), false positive (FP), true negative (TN), false negative (FN)
• Aggregate these labels across all records to compute intermediate metrics (for example, count the total TPs, FPs, TNs, and FNs, then calculate precision, recall, and accuracy)
• Compute the final score you want to optimize the prompt for by combining or selecting from the aggregated metrics (for example, return precision alone, or combine precision + accuracy)
• Label the misclassification examples to provide to the prompt optimizer

The following examples illustrate how to implement each steps.

• Individual evaluators measure each output:

  def confusion_matrix_evaluator(input_data, output_data, expected_output):
      """Evaluate a single prediction."""
      prediction = output_data.value
  
      if prediction and expected_output:
          return "true_positive"
      elif prediction and not expected_output:
          return "false_positive"
      elif not prediction and expected_output:
          return "false_negative"
      else:
          return "true_negative"
  

• Summary evaluators compute aggregate metrics:

  def precision_recall_evaluator(inputs, outputs, expected_outputs, evaluations):
      """Calculate precision and recall across all predictions."""
      tp = fp = tn = fn = 0
  
      for output, expected in zip(outputs, expected_outputs):
          pred = output.value if hasattr(output, 'value') else output
          if pred and expected:
              tp += 1
          elif pred and not expected:
              fp += 1
          elif not pred and expected:
              fn += 1
          else:
              tn += 1
  
      precision = tp / (tp + fp) if (tp + fp) > 0 else 0.0
      recall = tp / (tp + fn) if (tp + fn) > 0 else 0.0
      accuracy = (tp + tn) / (tp + tn + fp + fn)
  
      return {
          "precision": precision,
          "recall": recall,
          "accuracy": accuracy
      }
  

• Scoring function can return one of the values returned by one of the summary evaluators (for example, precision) or a combination of multiple metrics:

  def compute_score(summary_evaluators):
      """Higher is better. Combine metrics according to business priorities."""
      metrics = summary_evaluators['precision_recall_evaluator']['value']
      # Optimize for precision
      return metrics['precision']
  

  or

  def compute_score(summary_evaluators):
      """Computes F1 Score."""
      precision = summary_evaluators['precision_recall_evaluator']['value']['precision']
      recall = summary_evaluators['precision_recall_evaluator']['value']['recall']
      return 2 * (precision * recall) / (precision + recall)
  

• Labelization functions categorize results for showing diverse examples to the optimizer:

  def labelization_function(individual_result):
      """Categorize results into meaningful groups."""
      eval_value = individual_result["evaluations"]["confusion_matrix_evaluator"]["value"]
  
      if eval_value in ("true_positive", "true_negative"):
          return "CORRECT PREDICTION"
      else:
          return "INCORRECT PREDICTION"
  

  The labelization function plays an important role in optimization: for each unique label, the optimizer receives one randomly selected example from that category (in the example above, the labels are CORRECT PREDICTION and INCORRECT PREDICTION). This means the number of labels directly determines the diversity of examples shown to the reasoning model.

  Label names should be meaningful and descriptive, as they are shown directly to the reasoning model. Use clear, human-readable labels like HIGH CONFIDENCE ERROR or EDGE CASE FAILURE rather than codes like TYPE_A or CAT_3. Design your labels to represent the key patterns or hints you want the optimizer to learn from, and keep the cardinality low (fewer than 10 distinct labels) to help ensure focused, actionable feedback.

4. Define optimization task

Create a function that calls a reasoning model to suggest prompt improvements:

from pydantic import BaseModel

class OptimizationResult(BaseModel):
    prompt: str

def optimization_task(system_prompt, user_prompt, config):
    """Use reasoning model to improve the prompt.

    Returns:
        str: The improved prompt text
    """
    client = OpenAI()

    response = client.chat.completions.parse(
        model="o3-mini",  # Use advanced reasoning model
        messages=[
            {"role": "system", "content": system_prompt},
            {"role": "user", "content": user_prompt}
        ],
        response_format=OptimizationResult
    )

    return response.choices[0].message.parsed.prompt  # Must return string

5. Run optimization

Create and run the optimization:

prompt_optimization = LLMObs._prompt_optimization(
    name="hallucination-detection-optimization",
    dataset=dataset,
    task=detection_task,
    optimization_task=optimization_task,
    evaluators=[confusion_matrix_evaluator],
    summary_evaluators=[precision_recall_evaluator],
    labelization_function=labelization_function,
    compute_score=compute_score,
    config={
        "prompt": "Detect if the AI response contains hallucinated information.",  # Initial prompt (required)
        # Optionals
        "model_name": "gpt-4o-mini",  # Target model - helps optimizer tailor suggestions to model capabilities
        "evaluation_output_format": DetectionResult.output_format(),  # Expected output schema - prompts optimizer to ensure format compliance
        "runs": 1,  # Number of times to evaluate each record - higher values reduce variance in metrics
    },
    max_iterations=10,
    stopping_condition=lambda evals: (
        evals['precision_recall_evaluator']['value']['precision'] >= 0.9 and
        evals['precision_recall_evaluator']['value']['accuracy'] >= 0.8
    )
)

# Execute optimization (parallel execution for faster results)
result = prompt_optimization.run(jobs=20)

# Access results
print(f"Best prompt: {result.best_prompt}")
print(f"Best score: {result.best_score}")
print(f"Achieved in {result.total_iterations} iterations")
print(f"View in Datadog: {result.best_experiment_url}")

# Get score progression
print(result.summary())

Enable dataset splitting

Split your dataset into train, validation, and test subsets to get an unbiased estimate of prompt performance. Train examples feed the optimization LLM, validation scores rank iterations, and a final test experiment on the best prompt provides an unbiased score.

Use dataset_split=True for default 60/20/20 ratios:

prompt_optimization = LLMObs._prompt_optimization(
    name="hallucination-detection-optimization",
    dataset=dataset,
    task=detection_task,
    optimization_task=optimization_task,
    evaluators=[confusion_matrix_evaluator],
    summary_evaluators=[precision_recall_evaluator],
    labelization_function=labelization_function,
    compute_score=compute_score,
    config={
        "prompt": "Detect if the AI response contains hallucinated information.",
    },
    max_iterations=10,
    dataset_split=True,  # Default 60/20/20 split
)

result = prompt_optimization.run(jobs=20)

# Access validation and test results
print(f"Best validation score: {result.best_score}")
print(f"Test score: {result.test_score}")
print(f"Test experiment: {result.test_experiment_url}")

Specify custom ratios with a 3-tuple:

prompt_optimization = LLMObs._prompt_optimization(
    # ... same parameters as above ...
    dataset_split=(0.7, 0.15, 0.15),  # 70% train, 15% valid, 15% test
)

Use a separate test dataset:

prompt_optimization = LLMObs._prompt_optimization(
    # ... same parameters as above ...
    dataset_split=True,  # Splits main dataset 80/20 into train/valid
    test_dataset="my-curated-test-set",  # Separate dataset for unbiased testing
)

Configuration options

config

A configuration dictionary passed to your task function. Contains the following keys, as well as any custom parameters your task function needs:

"prompt"

required
The initial prompt

"model_name"

optional
Specifies the target model for your task. When provided, the optimizer includes model-specific guidance in its suggestions, tailoring improvements to that model’s capabilities and limitations (for example, GPT-4 versus Claude versus Llama).

evaluation_output_format

optional
Provides the JSON schema for your expected output structure. The optimizer uses this to ensure the improved prompt explicitly instructs the model to produce correctly formatted output. This is particularly valuable for structured outputs, where format compliance is critical.

runs

optional
Controls how many times each dataset record is evaluated. Setting runs > 1 helps reduce variance in metrics for tasks with non-deterministic outputs, providing more stable optimization signals at the cost of longer execution time.

max_iterations

Controls the maximum number of optimization cycles. Each iteration tests a new prompt on the full dataset.
Default: 5
Recommended: 10-20 for initial exploration, 5-10 for production

stopping_condition

Optional function that determines when to terminate optimization early. Receives summary evaluations and returns True to stop.

stopping_condition=lambda evals: (
    evals['my_evaluator']['value']['metric'] >= 0.95
)

Use AND conditions to help ensure multiple metrics meet targets before stopping.

dataset_split

Controls dataset splitting for unbiased evaluation. Accepts the following values:

• False (default): No splitting. The full dataset is used for optimization and scoring.
  
• True: Split with default ratios. Uses 60/20/20 (train/valid/test) without test_dataset, or 80/20 (train/valid) with test_dataset.
  
• (train, valid, test) tuple: Custom 3-way split ratios. Must sum to 1.0. Cannot be combined with test_dataset.
  
• (train, valid) tuple: Custom 2-way split ratios. Must sum to 1.0. Requires test_dataset for the test set.

test_dataset

Name of a separate dataset to use for the final test experiment. When provided, Prompt Optimization pulls this dataset automatically, splits the main dataset into train/valid (80/20 by default), and runs a final test experiment on the best prompt using this dataset. Providing test_dataset implicitly enables dataset splitting.

Configure parallel workers

When you execute optimization, you can configure your number of parallel workers by passing the jobs parameter to the run() function:

result = prompt_optimization.run(jobs=20)

Higher values reduce total runtime, but may hit API rate limits.

jobs

Default: 1
Recommended: 10-20, for most use cases

Example for a dataset of 100 records with 10 iterations:

• jobs=1 (serial): ~50 minutes (assuming 5s per call)
• jobs=20 (parallel): ~5 minutes

Dataset splitting

When optimizing a prompt, the same dataset is used both to guide improvements and to evaluate the result. This can lead to overfitting, where the prompt performs well on the optimization data but poorly on new inputs. Dataset splitting addresses this by dividing your data into separate subsets with distinct roles:

Default ratios

The split ratios depend on how you configure dataset_split and test_dataset:

Records are shuffled with a fixed seed for reproducibility across runs.

Understanding results

The OptimizationResult object provides comprehensive access to optimization outcomes:

Properties

• best_prompt: The highest-scoring prompt discovered
• best_score: Score of the best iteration
• best_experiment_url: Link to the Datadog experiment for detailed analysis
• total_iterations: Number of iterations completed
• best_iteration: Which iteration achieved the best score
• test_score: Score from the final test experiment. Returns None when dataset splitting is disabled.
• test_experiment_url: Link to the test experiment in Datadog. Returns None when dataset splitting is disabled.
• test_results: Full results from the test experiment. Returns None when dataset splitting is disabled.

Methods

• get_history(): Complete data for all iterations (prompts, scores, results, URLs)
• get_score_history(): List of scores showing progression
• get_prompt_history(): List of prompts showing evolution
• summary(): Human-readable overview with score progression table. When dataset splitting is enabled, the summary includes the test score and test set evaluations.

Example analyzing results:

# View complete history
for iteration_data in result.get_history():
    print(f"Iteration {iteration_data['iteration']}")
    print(f"  Score: {iteration_data['score']}")
    print(f"  Metrics: {iteration_data['summary_evaluations']}")
    print(f"  URL: {iteration_data['experiment_url']}")

# Visualize improvement
import matplotlib.pyplot as plt
plt.plot(result.get_score_history())
plt.xlabel('Iteration')
plt.ylabel('Score')
plt.title('Prompt Optimization Progress')
plt.show()

When dataset splitting is enabled, access the test results:

# Check for overfitting by comparing validation and test scores
print(f"Best validation score: {result.best_score}")
print(f"Test score: {result.test_score}")
print(f"Test experiment: {result.test_experiment_url}")

# A large gap between best_score and test_score may indicate overfitting

You can also view the prompt used for each iteration in the Config tab of the Experiment page.

Best practices

You will find a collection of example scripts in the Experiment cookbook repository.

Dataset design

• Include 50-100 diverse examples covering typical and edge cases
• For classification tasks, ensure balanced representation across classes
• Validate that ground truth labels are correct and consistent
• Version datasets for reproducibility

Evaluation metrics

• Choose metrics aligned with business objectives (precision vs. recall trade-offs)
• Use multiple complementary metrics rather than a single aggregate score
• Weight metrics by business impact in your compute_score function
• Test evaluators independently before running optimization

Labelization

• Create 2-5 distinct, descriptive labels (for example: CORRECT HIGH CONFIDENCE, INCORRECT EDGE CASE)
• Ensure balanced label distribution (for example, avoid 95% in one category)
• Use labels to help the optimizer understand different types of successes and failures

Dataset splitting

• Use at least 50 records in your dataset before enabling splitting to help ensure each subset has enough examples
• Start with the default ratios (dataset_split=True) before customizing splits
• Use test_dataset when you have a curated hold-out set that represents production traffic
• Compare best_score (validation) with test_score to detect overfitting: a large gap suggests the prompt is too specialized to the validation data

Optimization model selection

Use advanced reasoning models for the optimization_task:

• Recommended: OpenAI o3-mini, o1-preview, or Claude 3.5 Sonnet
• Use structured outputs (Pydantic models) for reliable parsing

Avoid using cheaper models (GPT-3.5-turbo, Claude Haiku) as they lack the reasoning depth needed for effective optimization.

Further reading

お役に立つドキュメント、リンクや記事:

Set up and use LLM Observability ExperimentsDOCUMENTATION Create and manage datasetsDOCUMENTATION Create and monitor LLM experiments with DatadogBLOG