Opik Optimizer

Purpose

Design, run, and interpret Opik Optimizer workflows for prompts, tools, and model parameters with consistent dataset/metric wiring and reproducible evaluation.

When to use

Use this skill when a user asks for:

• Choosing and configuring Opik Optimizer algorithms for prompt/agent optimization.
• Writing ChatPrompt-based optimization runs and custom metric functions.
• Optimizing with tools (function calling or MCP), selected prompt roles, or prompt segments.
• Tuning LLM call parameters with optimize_parameter.
• Comparing optimizer outputs and interpreting OptimizationResult.

Workflow

1. Select optimizer strategy (MetaPromptOptimizer, FewShotBayesianOptimizer, HRPO, etc.) based on the target optimization goal.
2. Build prompt/dataset/metric wiring and validate placeholder-field alignment.
3. Run prompt, tool, or parameter optimization with explicit controls (nthreads, nsamples, max_trials, seed).
4. Inspect OptimizationResult and compare score deltas against initial baselines.
5. Summarize recommendations, risks, and next experiments.

Inputs

• Target optimization objective (prompt/tool/parameter) and success metric.
• Dataset source and expected schema fields.
• Model/provider constraints and runtime limits.
• Optional scope constraints (optimize_prompts segments, tool fields, project names).

Outputs

• Optimizer run configuration and rationale.
• Result interpretation (score, initial_score, history trends).
• Recommended next changes and follow-up experiment plan.

Use the reference files in this skill for details before implementing code:

• references/algorithms.md
• references/promptagentworkflow.md
• references/example_patterns.md

Opik Optimizer quickstart

1. Install and import:

pip install opik-optimizer

from opik_optimizer import ChatPrompt, MetaPromptOptimizer, HRPO, FewShotBayesianOptimizer
from opik_optimizer import datasets

1. Build a prompt and metric:

from opik.evaluation.metrics import LevenshteinRatio

prompt = ChatPrompt(
    system="You are a concise answerer.",
    user="{question}",
)

def metric(dataset_item: dict, output: str) -> float:
    return LevenshteinRatio().score(
        reference=dataset_item["answer"],
        output=output,
    ).value

1. Load dataset and run:

dataset = datasets.hotpot(count=30)

result = MetaPromptOptimizer(model="openai/gpt-5-nano").optimize_prompt(
    prompt=prompt,
    dataset=dataset,
    metric=metric,
    n_samples=20,
    max_trials=10,
)
result.display()

Core workflow you should follow

1. Pick optimizer class:

- Few-shot examples + Bayesian selection: FewShotBayesianOptimizer - LLM meta-reasoning: MetaPromptOptimizer - Genetic + MOO / LLM crossover: EvolutionaryOptimizer - Hierarchical reflective diagnostics: HierarchicalReflectiveOptimizer (HRPO) - Pareto-based genetic strategy: GepaOptimizer - Parameter tuning only: ParameterOptimizer

1. Define a single ChatPrompt (or dict of prompts for multi-prompt cases).
2. Provide a dataset from opik_optimizer.datasets.
3. Provide metric callable with signature (datasetitem, llmoutput) -> float (or ScoreResult/list of ScoreResult).
4. Set optimizer controls (nthreads, nsamples, max_trials, seed, etc.).
5. Run one of:

- optimizeprompt(...) for prompt/system behavior changes. - optimizeparameter(...) for model-call hyperparameters.

1. Inspect OptimizationResult (score, initialscore, history, optimizationid, getoptimizedparameters).

Key execution details to enforce

• Prefer explicit project_name for Opik tracking if you are using org-level observability.
• Keep placeholders in prompts aligned with dataset fields (for example {question}).
• Start with optimize_prompts="system" or "user" when scope should be constrained.
• Keep model names in MetaPrompt/reasoning calls provider-compatible for your account.
• Validate multimodal input payloads by preserving non-empty content segments only.
• For small datasets, use nsamples and nsamples_strategy carefully; over-allocation auto-falls back to full set.

Tooling and segment-based control

• Tools can be optimized with MCP/function schema fields, not only by changing prompt wording.
• For fine-grained text updates, use optimizeprompts values and helper functions from promptsegments:

- extractpromptsegments(ChatPrompt) to inspect stable segment IDs. - applysegmentupdates(ChatPrompt, updates) for deterministic edits.

• Tool optimization is distinct from prompt optimization.

Runnable examples live upstream in the Opik repo:

• https://github.com/comet-ml/opik/tree/main/sdks/opikoptimizer/src/opikoptimizer

If you need local runnable scripts, vendor the upstream examples into a scripts/ folder and keep references one level deep.

Common mistakes to avoid

• Passing empty dataset or mismatched placeholder names.
• Mixing deprecated constructor arg numthreads with nthreads.
• Assuming tool optimization is the same as agent function-calling optimization.
• Running ParameterOptimizer.optimize_prompt (it raises and should not be used).

Next actions

• For in-depth behavior and per-class parameter tables: references/algorithms.md
• For exact optimizeprompt signatures, prompts, tool constraints, and result usage: references/promptagent_workflow.md
• For pattern examples and source-backed workflows: references/example_patterns.md