levelapp 0.1.1

LevelApp is an evaluation framework for AI/LLM-based software application. [Powered by Norma]

LevelApp: AI/LLM Evaluation Framework for Regression Testing

Overview

LevelApp is an evaluation framework designed for regression testing (black-box) of already built LLM-based systems in production or testing phases. It focuses on assessing the performance and reliability of AI/LLM applications through simulation and comparison modules. Powered by Norma.

Key benefits:

• Configuration-driven: Minimal coding required; define evaluations via YAML files.
• Supports LLM-as-a-judge for qualitative assessments and quantitative metrics for metadata evaluation.
• Modular architecture for easy extension to new workflows, evaluators, and repositories.

Features

• Simulator Module: Evaluates dialogue systems by simulating conversations using predefined scripts. It uses an LLM as a judge to score replies against references and supports metrics (e.g., Exact, Embedded, Token-based, Fuzzy) for comparing extracted metadata to ground truth.
• Comparator Module: Evaluates metadata extraction from JSON outputs (e.g., from legal/financial document processing with LLMs) by comparing against reference/ground-truth data.
• Configuration-Based Workflow: Users provide YAML configs for endpoints, parameters, data sources, and metrics, reducing the need for custom code.
• Supported Workflows: SIMULATOR, COMPARATOR, ASSESSOR (coming soon!).
• Repositories: FIRESTORE, FILESYSTEM, MONGODB.
• Evaluators: JUDGE, REFERENCE, RAG.
• Metrics: Exact, Levenshtein, and more (see docs for full list).
• Data Sources: Local or remote JSON for conversation scripts.

Installation

Install LevelApp via pip:

pip install levelapp

Prerequisites

• Python 3.12 or higher.
• API keys for LLM providers (e.g., OpenAI, Anthropic) if using external clients—store in a .env file.
• Optional: Google Cloud credentials for Firestore repository.
• Dependencies are automatically installed, including openai, pydantic, numpy, etc. (see pyproject.toml for full list).

Configuration

LevelApp uses a YAML configuration file to define the evaluation setup. Create a workflow_config.yaml with the following structure:

process:
  project_name: "test-project"
  workflow_type: SIMULATOR # Pick one of the following workflows: SIMULATOR, COMPARATOR, ASSESSOR.
  evaluation_params:
    attempts: 1  # Add the number of simulation attempts.
    batch_size: 5

evaluation:
  evaluators: # Select from the following: JUDGE, REFERENCE, RAG.
    - JUDGE
    - REFERENCE
  providers:
    - openai
    - ionos
  metrics_map:
    field_1: EXACT
    field_2 : LEVENSHTEIN

reference_data:
  path: 
  data:

endpoint:
  base_url: "http://127.0.0.1:8000"
  url_path: ''
  api_key: "<API-KEY>"
  bearer_token: "<BEARER-TOKEN>"
  model_id: "meta-llama/Meta-Llama-3.1-8B-Instruct"
  default_request_payload_template:
    # Change the user message field name only according to the request payload schema (example: 'prompt' to 'message').
    prompt: "${user_message}"
    details: "${request_payload}"  # Rest of the request payload data.
  default_response_payload_template:
    # Change the placeholder value only according to the response payload schema (example: ${agent_reply} to ${reply}).
    agent_reply: "${agent_reply}"
    generated_metadata: "${generated_metadata}"

repository:
  type: FIRESTORE # Pick one of the following: FIRESTORE, FILESYSTEM, MONGODB.
  project_id: "(default)"
  database_name: ""

• Endpoint Configuration: Define how to interact with your LLM-based system (base URL, auth, payload templates).
• Placeholders: For the request payload, change the field names (e.g., 'prompt' to 'message') according to your API specs. For the response payload, change the place holders values (e.g., ${agent_reply} to ${generated_reply}).
• Secrets: Store API keys in .env and load via python-dotenv (e.g., API_KEY=your_key_here).

For conversation scripts (used in Simulator), provide a JSON file with this schema:

{
  "scripts": [
    {
      "interactions": [
        {
          "user_message": "Hello, I would like to book an appointment with a doctor.",
          "reference_reply": "Sure, I can help with that. Could you please specify the type of doctor you need to see?",
          "interaction_type": "initial",
          "reference_metadata": {},
          "guardrail_flag": false,
          "request_payload": {"user_id":  "0001", "user_role": "ADMIN"}
        },
        {
          "user_message": "I need to see a cardiologist.",
          "reference_reply": "When would you like to schedule your appointment?",
          "interaction_type": "intermediate",
          "reference_metadata": {},
          "guardrail_flag": false,
          "request_payload": {"user_id":  "0001", "user_role": "ADMIN"}
        },
        {
          "user_message": "I would like to book it for next Monday morning.",
          "reference_reply": "We have an available slot at 10 AM next Monday. Does that work for you?",
          "interaction_type": "intermediate",
          "reference_metadata": {
            "appointment_type": "Cardiology",
            "date": "next Monday",
            "time": "10 AM"
          },
          "guardrail_flag": false,
          "request_payload": {"user_id":  "0001", "user_role": "ADMIN"}
        },
        {
          "id": "f4f2dd35-71d7-4b75-ba2b-93a4f546004a",
          "user_message": "Yes, please book it for 10 AM then.",
          "reference_reply": "Your appointment with the cardiologist is booked for 10 AM next Monday. Is there anything else I can help you with?",
          "interaction_type": "final",
          "reference_metadata": {},
          "guardrail_flag": false,
          "request_payload": {"user_id":  "0001", "user_role": "ADMIN"}
        }
      ],
      "description": "A conversation about booking a doctor appointment.",
      "details": {
        "context": "Booking a doctor appointment"
      }
    }
  ]
}

• Fields: Include user messages, reference/references replies, metadata for comparison, guardrail flags, and request payloads.

In the .env you need to add the LLM providers credentials that will be used for the evaluation process.

OPENAI_API_KEY=
IONOS_API_KEY=
ANTHROPIC_API_KEY=
MISTRAL_API_KEY=

# For IONOS, you must include the base URL and the model ID.
IONOS_BASE_URL="https://inference.de-txl.ionos.com"
IONOS_MODEL_ID="0b6c4a15-bb8d-4092-82b0-f357b77c59fd"

WORKFLOW_CONFIG_PATH="../../src/data/workflow_config_1.yaml"

Usage Example

To run an evaluation:

1. Prepare your YAML config and JSON data files.
2. Use the following Python script:

if __name__ == "__main__":
    from levelapp.workflow.schemas import WorkflowConfig
    from levelapp.core.session import EvaluationSession

    # Load configuration from YAML
    config = WorkflowConfig.load(path="../data/workflow_config.yaml")

    # Run evaluation session
    with EvaluationSession(session_name="test-session-1", workflow_config=config) as session:
        session.run()
        results = session.workflow.collect_results()
        print("Results:", results)

    stats = session.get_stats()
    print(f"session stats:\n{stats}")

Alternatively, if you want to pass the configuration and reference data from in-memory variables, you can manually load the data like the following:

if __name__ == "__main__":
    from levelapp.workflow import WorkflowConfig
    from levelapp.core.session import EvaluationSession

    # Firestore -> retrieve endpoint config -> data => config_dict

    config_dict = {
        "process": {"project_name": "test-project", "workflow_type": "SIMULATOR", "evaluation_params": {"attempts": 2}},
        "evaluation": {"evaluators": ["JUDGE"], "providers": ["openai", "ionos"]},
        "reference_data": {"path": "", "data": {}},
        "endpoint": {"base_url": "http://127.0.0.1:8000", "api_key": "key", "model_id": "model"},
        "repository": {"type": "FIRESTORE", "source": "IN_MEMORY", "metrics_map": {"field_1": "EXACT"}},
    }

    content = {
        "scripts": [
            {
                "interactions": [
                    {
                        "user_message": "Hello!",
                        "reference_reply": "Hello, how can I help you!"
                    },
                    {
                        "user_message": "I need an apartment",
                        "reference_reply": "sorry, but I can only assist you with booking medical appointments."
                    },
                ]
            },
        ]
    }

    # Load configuration from a dict variable
    config = WorkflowConfig.from_dict(content=config_dict)

    # Load reference data from dict variable
    config.set_reference_data(content=content)

    evaluation_session = EvaluationSession(session_name="test-session-2", workflow_config=config)

    with evaluation_session as session:
        session.run()
        results = session.workflow.collect_results()
        print("Results:", results)

    stats = session.get_stats()
    print(f"session stats:\n{stats}")

• This loads the config, runs the specified workflow (e.g., Simulator), collects results, and prints stats.

For more examples, see the examples/ directory.

Documentation

Detailed docs are in the docs/ directory, including API references and advanced configuration.

Contributing

Contributions are welcome! Please follow these steps:

• Fork the repository on GitHub.
• Create a feature branch (git checkout -b feature/new-feature).
• Commit changes (git commit -am 'Add new feature').
• Push to the branch (git push origin feature/new-feature).
• Open a pull request.

Report issues via GitHub Issues. Follow the code of conduct (if applicable).

Acknowledgments

• Powered by Norma.
• Thanks to contributors and open-source libraries like Pydantic, NumPy, and OpenAI SDK.

License

This project is licensed under the MIT License - see the LICENCE file for details.

---