Quickstart

1
from opik.evaluation.metrics import LevenshteinRatio
2
from opik_optimizer import FewShotBayesianOptimizer
3
from opik_optimizer.demo import get_or_create_dataset
4
from opik_optimizer import (
5
    MetricConfig,
6
    TaskConfig,
7
    from_llm_response_text,
8
    from_dataset_field,
9
)
10
11
# 1. Define your evaluation dataset
12
# You can use a demo dataset for testing, or your own dataset
13
dataset = get_or_create_dataset("tiny-test")
14
15
# 2. Configure the evaluation metric
16
# This example uses Levenshtein distance to measure output quality
17
metric_config = MetricConfig(
18
    metric=LevenshteinRatio(),
19
    inputs={
20
        "output": from_llm_response_text(),  # Model's output
21
        "reference": from_dataset_field(name="label"),  # Ground truth
22
    }
23
)
24
25
# 3. Define your base prompt
26
# Following best practices, provide clear instructions and context
27
system_prompt = """You are an expert assistant. Your task is to answer questions
28
accurately and concisely. Consider the context carefully before responding."""
29
30
# 4. Choose and configure an optimizer
31
# FewShotBayesianOptimizer automatically finds optimal examples to include
32
optimizer = FewShotBayesianOptimizer(
33
    model="gpt-4",  # Choose your preferred model
34
    project_name="Prompt Optimization",
35
    min_examples=2,  # Minimum number of examples to try
36
    max_examples=8,  # Maximum number of examples to try
37
    n_iterations=10  # Number of optimization iterations
38
)
39
40
# 5. Configure the task
41
task_config = TaskConfig(
42
    instruction_prompt=system_prompt,
43
    input_dataset_fields=["text"],  # Input field(s) from your dataset
44
    output_dataset_field="label",   # Output field from your dataset
45
    use_chat_prompt=True,  # Use chat format for modern LLMs
46
)
47
48
# 6. Run the optimization
49
result = optimizer.optimize_prompt(
50
    dataset=dataset,
51
    metric_config=metric_config,
52
    task_config=task_config,
53
)
54
55
# 7. View the results
56
result.display()  # Shows improvement metrics and best prompt