DeepEval – Unit Testing for LLMs

👩‍⚖️ DeepEval

Evals provides a Pythonic way to run offline evaluations on your LLM pipelines so you can launch comfortably into production. The guiding philosophy is a "Pytest for LLM" that aims to make productionizing and evaluating LLMs as easy as software engineering.

Documentation

We highly recommend getting started through our documentation here: https://docs.confident-ai.com/docs/

Join our discord: https://discord.gg/a3K9c8GRGt

Why DeepEval?

Deepeval aims to make writing tests for LLM applications (such as RAG) as easy as writing Python unit tests.

For any Python developer building production-grade apps, it is common to set up PyTest as the default testing suite as it provides a clean interface to quickly write tests.

However, it is often uncommon for many machine learning engineers as their feedback is often in the form of an evaluation loss.

With the advent of agents, LLMs and AI, there is yet to be a tool that can provide software-like tooling and abstractions for machine learning engineers where the feedback loop of these iterations can be significantly reduced.

It is therefore important then to build a new type of testing framework for LLMs to ensure engineers can keep iterating on their prompts, agents and LLMs while being able to continuously add to their test suite.

Introducing DeepEval.

Installation

QuickStart

Individual Test Cases

# test_example.py
from deepeval.test_utils import assert_llm_output

def generate_llm_output(input: str):
    expected_output = "Our customer success phone line is 1200-231-231."
    return expected_output

def test_llm_output(self):
    input = "What is the customer success phone line?"
    expected_output = "Our customer success phone line is 1200-231-231."
    output = generate_llm_output(input)
    assert_llm_output(output, expected_output, metric="entailment")
    assert_llm_output(output, expected_output, metric="exact")

Once you have set that up, you can simply call pytest

python -m pytest test_example.py

# Output
Running tests ... ✅

Bulk Test Cases

You can run a number of test cases which you can define either through CSV or through our hosted option.

from deepeval import BulkTestRunner, TestCase

class BulkTester(BulkTestRunner):
    @property
    def bulk_test_cases(self):
        return [
            TestCase(
                input="What is the customer success number",
                expected_output="1800-213-123",
                tags=["Customer success"]
            ),
            Testcase(
                input="What do you think about the models?",
                expected_output="Not much - they are underperforming.",
                tags=["Machine learning"]
            )
        ]

tester = BulkTester()
tester.run(callable_fn=generate_llm_output)

From CSV

You can import test cases from CSV.

import pandas as pd
df = pd.read_csv('sample.csv')
from deepeval import TestCases
# Assuming you have the column names `input`, `expected_output`
class BulkTester(BulkTestRunner):
    @property
    def bulk_test_cases(self):
        return TestCases.from_csv(
            "sample.csv",
            input_column="input",
            expected_output_column="output",
            id_column="id"
        )

Setting up metrics

Setting up custom metrics

To define a custom metric, you simply need to define the measure and is_successful property.

from deepeval.metric import Metric
class CustomMetric(Metric):
    def measure(self, a, b):
        return a > b
    def is_successful(self):
        return True

metric = CustomMetric()

Integrate tightly with LangChain

We integrate DeepEval tightly with common frameworks such as Langchain and lLamaIndex.

Synthetic Query Generation

Generating synthetic queries allows you to quickly evaluate the queries related to your prompts. We help developers get up and running with a lot of example queries.

Dashboard (Coming soon)

Once you have added a ground truth, you should be able to see a dashboard that contains information about the pipeline and the run.

RoadMap

Authors

Built by the Confident AI Team

from Hacker News https://ift.tt/KcuMbDm