MDAL (Model-Driven Architecture Layer)

What is it?

MDAL is a specialized normalization layer (proxy) that acts as middleware between applications/users and Large Language Models (LLMs).

The goal of MDAL is to dampen the so-called "Model-Shift Effect": the user experience, tone, and structure of responses stay absolutely consistent, regardless of which AI model (or which model version) is doing the actual generation work in the background.

What does MDAL do? (High-Level)

MDAL intercepts LLM responses and checks them against a pre-defined "character fingerprint" (the target state). MDAL only forwards responses to the user once they meet the quality criteria:

• Style Normalization: MDAL checks formality level, sentence structure, and vocabulary (e.g. consistent formal register).
• Structure Validation: Hard verification of structured outputs (e.g. JSON or XML) against defined schemas.
• Semantic Integrity: An LLM-based transformer smooths the text style when needed, but strictly blocks hallucinations or factual drift (entity check).
• Merciless Escalation Logic: When the model fails to meet requirements, MDAL forces retries. If the model fails repeatedly, MDAL blocks the response with HTTP 503 Service Unavailable rather than passing through bad output ("strictness over silent compromise").

---

How do I start?

The system is lightweight, API-first by design, and ideal for use with e.g. Ollama or OpenAI-compatible endpoints.

1. Install dependencies

All that's needed is a recent Python environment:

pip install -r requirements.txt

2. Start the MDAL Server

MDAL features a built-in configuration web interface (similar to a smart home router setup). You don't need to manually edit YAML files to get started.

# Starts on port 6969 by default
python -m mdal.proxy.server

3. Configure and Run

Open the built-in Control Center in your browser at http://localhost:6969/config.

From the Control Center UI, you can:

• set up your LLM and embedding endpoints with presets for Ollama, OpenAI, Anthropic, and Google
• configure audit logging, validation checks, notifier settings, fingerprint path, and plugin registry path
• start or stop the MDAL proxy and launch the trainer from the UI

Some advanced runtime options such as fallback_llm, max_retries, and language are still configured manually in config/mdal.yaml.

python -m mdal.trainer.trainer --config config/mdal.yaml --input your_chats.json --language de

---

Further Documentation

For in-depth information on the architecture, detailed layer concepts, configuration, and plugin development, see the full user documentation: 👉 Open User Documentation