Overview - Opik — ContextQMD

<Note> In Opik 2.0, Experiments and Evaluation Suites are project-scoped. Make sure to specify a `project_name` when creating datasets and running experiments. </Note>

Evaluation in Opik helps you assess and measure the quality of your LLM outputs across different dimensions. It provides a framework to systematically test your prompts and models against datasets, using various metrics to measure performance.

Opik also provides a set of pre-built metrics for common evaluation tasks. These metrics are designed to help you quickly and effectively gauge the performance of your LLM outputs and include metrics such as Hallucination, Answer Relevance, Context Precision/Recall and more. You can learn more about the available metrics in the Metrics Overview section.

<Tip> If you are interested in evaluating your LLM application in production, please refer to the [Online evaluation guide](/v1/production/rules). Online evaluation rules allow you to define LLM as a Judge metrics that will automatically score all, or a subset, of your production traces. </Tip> <Note> **New: Multi-Value Feedback Scores** - Opik now supports collaborative evaluation where multiple team members can score the same traces and spans. This reduces bias and provides more reliable evaluation results through automatic score aggregation. [Learn more →](/v1/tracing/annotate_traces#multi-value-feedback-scores) </Note>

Running an Evaluation

Each evaluation is defined by a dataset, an evaluation task and a set of evaluation metrics:

Dataset: A dataset is a collection of samples that represent the inputs and, optionally, expected outputs for your LLM application.
Evaluation task: This maps the inputs stored in the dataset to the output you would like to score. The evaluation task is typically a prompt template or the LLM application you are building.
Metrics: The metrics you would like to use when scoring the outputs of your LLM

To simplify the evaluation process, Opik provides two main evaluation methods: evaluate_prompt for evaluation prompt templates and a more general evaluate method for more complex evaluation scenarios.

<Tip> **TypeScript SDK Support** This document covers evaluation using Python, but we also offer full support for TypeScript via our dedicated TypeScript SDK. See the [TypeScript SDK Evaluation documentation](/reference/typescript-sdk/evaluation/overview) for implementation details and examples. </Tip> <Tabs> <Tab value="Evaluating Prompts" title="Evaluating Prompts">

To evaluate a specific prompt against a dataset:

python

import opik
from opik.evaluation import evaluate_prompt
from opik.evaluation.metrics import Hallucination

# Create a dataset that contains the samples you want to evaluate
opik_client = opik.Opik()
dataset = opik_client.get_or_create_dataset("Evaluation test dataset", project_name="my-project")
dataset.insert([
    {"input": "Hello, world!", "expected_output": "Hello, world!"},
    {"input": "What is the capital of France?", "expected_output": "Paris"},
])

# Run the evaluation
result = evaluate_prompt(
    dataset=dataset,
    messages=[{"role": "user", "content": "Translate the following text to French: {{input}}"}],
    model="gpt-3.5-turbo",  # or your preferred model
    scoring_metrics=[Hallucination()],
    project_name="my-project"
)

</Tab> <Tab value="Evaluating RAG applications and Agents" title="Evaluating RAG applications and Agents">

For more complex evaluation scenarios where you need custom processing:

python

import opik
from opik.evaluation import evaluate
from opik.evaluation.metrics import ContextPrecision, ContextRecall

# Create a dataset with questions and their contexts
opik_client = opik.Opik()
dataset = opik_client.get_or_create_dataset("RAG evaluation dataset", project_name="my-project")
dataset.insert([
    {
        "input": "What are the key features of Python?",
        "context": "Python is known for its simplicity and readability. Key features include dynamic typing, automatic memory management, and an extensive standard library.",
        "expected_output": "Python's key features include dynamic typing, automatic memory management, and an extensive standard library."
    },
    {
        "input": "How does garbage collection work in Python?",
        "context": "Python uses reference counting and a cyclic garbage collector. When an object's reference count drops to zero, it is deallocated.",
        "expected_output": "Python uses reference counting for garbage collection. Objects are deallocated when their reference count reaches zero."
    }
])

def rag_task(item):
    # Simulate RAG pipeline
    output = "<LLM response placeholder>"

    return {
        "output": output
    }

# Run the evaluation
result = evaluate(
    dataset=dataset,
    task=rag_task,
    scoring_metrics=[
        ContextPrecision(),
        ContextRecall()
    ],
    experiment_name="rag_evaluation",
    project_name="my-project"
)

</Tab> <Tab value="Using the Playground" title="Using the Playground"> You can also use the Opik Playground to quickly evaluate different prompts and LLM models.

To use the Playground, you will need to navigate to the [Playground](/v1/prompt_engineering/playground) page and:

1. Configure the LLM provider you want to use
2. Enter the prompts you want to evaluate - You should include variables in the prompts using the `{{variable}}` syntax
3. Select the dataset you want to evaluate on
4. Click on the `Evaluate` button

You will now be able to view the LLM outputs for each sample in the dataset:

![Playground](/img/evaluation/playground_evaluation.gif)

</Tab> </Tabs>

Analyzing Evaluation Results

Once the evaluation is complete, Opik allows you to manually review the results and compare them with previous iterations.

In the experiment pages, you will be able to:

Review the output provided by the LLM for each sample in the dataset
Deep dive into each sample by clicking on the item ID
Review the experiment configuration to know how the experiment was Run
Compare multiple experiments side by side

Analyzing Evaluation Results in Python

To analyze the evaluation results in Python, you can use the EvaluationResult.aggregate_evaluation_scores() method to retrieve the aggregated score statistics:

python

import opik
from opik.evaluation import evaluate_prompt
from opik.evaluation.metrics import Hallucination

# Create a dataset that contains the samples you want to evaluate
opik_client = opik.Opik()
dataset = opik_client.get_or_create_dataset("Evaluation test dataset", project_name="my-project")
dataset.insert([
    {"input": "Hello, world!", "expected_output": "Hello, world!"},
    {"input": "What is the capital of France?", "expected_output": "Paris"},
])
# Run the evaluation
result = evaluate_prompt(
    dataset=dataset,
    messages=[{"role": "user", "content": "Translate the following text to French: {{input}}"}],
    model="gpt-5",  # or model that you want to evaluate
    scoring_metrics=[Hallucination()],
    verbose=2,  # print detailed statistics report for every metric
    project_name="my-project",
)

# Retrieve and print the aggregated scores statistics (mean, min, max, std) per metric
scores = result.aggregate_evaluation_scores()
for metric_name, statistics in scores.aggregated_scores.items():
    print(f"{metric_name}: {statistics}")

You can use aggregated scores to compare the performance of different models or different versions of the same model.

Experiment-Level Metrics

In addition to per-item metrics, you can compute experiment-level aggregate metrics that are calculated across all test results. These experiment scores are displayed in the Opik UI alongside feedback scores and can be used for sorting and filtering experiments.

<Tip> Learn more about computing experiment-level metrics in the [Evaluate your LLM application](/v1/evaluation/evaluate_your_llm#computing-experiment-level-metrics) guide. </Tip>

Learn more

You can learn more about Opik's evaluation features in: