What are Recipes?

Recipes are the core of Adaptive Engine, they are the fundamental building blocks that enable you to write custom scripts for various AI workflows. Recipes interact directly with the Adaptive Engine compute plane, exposing LLM methods in a similar fashion to API endpoints (e.g. .generate(), .train_grpo()).

Overview of flows

Built-in Recipes

The Adaptive team provides built-in recipes for the most common AI flows, covering: Model Tuning
  • Supervised fine-tuning
  • RL from preference or metric feedback
  • RL with graders
Model Evaluation
  • Evaluation with graders
These pre-built recipes offer a quick start for standard workflows and include a wide range of parametrization configuration, allowing you to easily customize them to fit your specific needs.

Custom Recipes

Beyond the built-in options, you can write your own custom recipes using the building blocks provided in the adaptive_harmony package. This approach allows you to create very diverse scripts and AI workflows tailored to your specific requirements. The adaptive_harmony package provides a comprehensive set of components and utilities that you can combine and orchestrate to build complex AI pipelines, integrate with external systems, or implement specialized logic that isn’t covered by the pre-built recipes. Recipes are designed to be flexible and can be adapted for various use cases. You can write custom logic for:
  • Model tuning
  • Model evaluation
  • Custom reusable graders
  • Synthetic data generation
  • Red teaming
  • and much more…

How It Works

With recipes, you define a set of high-level set of instructions that control the sequence of actions, while Adaptive Engine manages the underlying orchestration, compute resources, and infrastructure for you. Whether you’re working directly in the UI or through the SDK, recipes give you the ability to customize and control your AI workflows within the Adaptive Engine platform.