​
Athina in CI/CD Pipelines

You can use Athina evals in your CI/CD pipeline to catch regressions before they get to production. Following is a guide for setting athina-evals in your CI/CD pipeline.

All codes described here are present in our GitHub repository as well.

​
GitHub Actions

We’re going to use GitHub Actions to create our CI/CD pipelines. GitHub Actions allow us to define workflows that are triggered by events (pull request, push, etc.) and execute a series of actions. Our GitHub Actions are defined under our repository’s .github/workflows directory.

We have defined a workflow for the evals too. The workflow file is named athina_ci.yml. The workflow is triggered on every push to the main branch.

name: CI with Athina Evals

on:
  push:
    branches:
      - main

jobs:
  evaluate:
    runs-on: ubuntu-latest

    steps:
      - uses: actions/checkout@v3

      - name: Set up Python
        uses: actions/setup-python@v4
        with:
          python-version: '3.9'

      - name: Install Dependencies
        run: |
          python -m pip install --upgrade pip
          pip install -r requirements.txt  # Install project dependencies
          pip install athina  # Install Athina Evals

      - name: Run Athina Evaluation and Validation Script
        run: python -m evaluations.run_athina_evals
        env:
          OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
          ATHINA_API_KEY: ${{ secrets.ATHINA_API_KEY }}

​
Athina Evals Script

The run_athina_evals.py script is the entry point for our Athina Evals. It is a simple script that uses the Athina Evals SDK to evaluate and validate the Rag Application.

For example we are testing if the response from the Rag Application answers the query using the DoesResponseAnswerQuery evaluation from athina.

eval_model = "gpt-3.5-turbo"
df = DoesResponseAnswerQuery(model=eval_model).run_batch(data=dataset).to_df()
# Validation: Check if all rows in the dataframe passed the evaluation
all_passed = df['passed'].all()
if not all_passed:
    failed_responses = df[~df['passed']]
    print(f"Failed Responses: {failed_responses}")
    raise ValueError("Not all responses passed the evaluation.")
else:
    print("All responses passed the evaluation.")

You can also load a golden dataset and run the evaluation on it.

with open('evaluations/golden_dataset.jsonl', 'r') as file:
  raw_data = file.read().split('\n')
  data = []
  for item in raw_data:
    item = json.loads(item)
    item['context'], item['response'] = app.generate_response(item['query'])
    data.append(item)

You can also run a suite of evaluations on the dataset.

eval_model = "gpt-3.5-turbo"
eval_suite = [
  DoesResponseAnswerQuery(model=eval_model),
  Faithfulness(model=eval_model),
  ContextContainsEnoughInformation(model=eval_model),
]

# Run the evaluation suite
batch_eval_result = EvalRunner.run_suite(
  evals=eval_suite,
  data=dataset,
  max_parallel_evals=2
)

# Validate the batch_eval_results as you want.

​
Secrets

We are using GitHub Secrets to store our API keys. We have two secrets, OPENAI_API_KEY and ATHINA_API_KEY. You can add these secrets to your repository by navigating to Settings > Secrets > New repository secret.

​
Further reading

We have more examples and details in our GitHub repository

Running Evals via SDKOpen Source Evaluations