ell

ell is an open-source, MIT-licensed Python library for prompt engineering, created by William Guss. It treats prompts as functions called language model programs (LMPs), and adds versioning, call storage, and a local visualization tool on top of regular Python code.

What It Is

ell is a lightweight, functional prompt engineering framework. An LMP is an encapsulated function that returns either a string prompt or a list of messages for a language model. The @ell.simple decorator yields plain string outputs, while @ell.complex returns Message objects for tool calls and multimodal responses. Installation is via pip install -U "ell-ai[all]".

Versioning, Tracing, and Studio

Calling ell.init(store='./logdir') makes ell version LMPs through static and dynamic analysis and write them to a local store. The docs say it generates commit messages automatically with gpt-4o-mini. It can also optionally save every language model call locally, which allows building invocation datasets and comparing outputs across versions. Ell Studio, launched with ell-studio --storage ./logdir, is a local open-source tool for prompt version control, monitoring, and visualization.

Composition and Multimodality

Because LMPs are ordinary functions, they can be chained for multi-call patterns such as drafting several candidates and having another LMP select the best. The documentation also covers tool usage, structured outputs, and multimodal inputs such as PIL images inside messages. It lists evaluations as a new feature. The docs say ell can be adopted incrementally, migrating from langchain one function at a time.