Skip to content

Getting Started

Delphyne Logo Delphyne Logo

Delphyne is a programming framework for building reliable and modular LLM applications. It is based on a new paradigm named oracular programming, where high-level problem-solving strategies are expressed as nondeterministic programs whose choice points are annotated with examples and resolved by LLMs. Delphyne combines three languages:

  • A strategy language embedded in Python that allows writing nondeterministic programs that can be reified into (modular) search trees.
  • A policy language for specifying ways to navigate such trees (with LLM guidance) by composing reusable search primitives.
  • A demonstration language for describing successful and unsuccessful search scenarios to be used as training or prompting examples. A dedicated language server allows writing demonstrations interactively and keeping them synchronized with evolving strategies.

Installation

First, download the Delphyne repository and enter it:

git clone git@github.com:jonathan-laurent/delphyne.git
cd delphyne
git checkout v0.7.0  # latest stable version

Then, to install the Delphyne library in your current Python environment:

pip install -e .

Note that Python 3.12 (or more recent) is required. Once this is done, it should be possible to run import delphyne inside a Python interpreter. Next, you should build the Delphyne vscode extension. For this, assuming you have Node.js installed (version 22 or later), run:

cd vscode-ui
npm install
npx vsce package

The last command should create a delphyne-xxx.vsix extensions archive, which can be installed in vscode using the Extensions: Install from VSIX command (use Ctrl+Shift+P to search for this command).

Testing your installation

To test your installation, open VSCode and set the examples/find_invariants folder as your workspace root. Click on the Delphyne logo on the Activity Bar to start the Delphyne extension, and open the demonstration file abduct_and_branch.demo.yaml. Then, open the command palette (Ctrl+Shift+P) and run the command Delphyne: Evaluate All Demonstrations in File. Diagnostics should then appear to indicate that all tests passed (but no warning or error). Note that adding new demonstrations requires installing why3py, as explained in the example's README.