MadcowD/ell
Ell
Ell is a Python library that treats prompts as callable functions, transforming a Python function into a language model program where the docstring defines the system message and the return value defines the user message. It provides a framework for writing language model programs that can accept and return images, audio, and text natively within message objects, and supports chaining multiple model calls into a single function for multi-step reasoning or test-time compute.
The library decouples model selection from client instantiation through a registry that supports multiple provider backends, and includes a provider plugin interface for routing model calls to different backends. It automatically tracks every change to a prompt function with generated commit messages and full history, and logs the full input and output of every language model call for later analysis or dataset creation. A local web-based dashboard visualizes and compares prompt outputs across versions to catch regressions and track behavior over time.
Ell offers a model registry with temporary overrides for testing or routing, and a configuration system for setting global defaults for model parameters, storage, logging, and versioning behavior. The library persists prompt versions, invocations, and metadata to a local filesystem store for offline browsing and comparison.
Features
- Prompt Programming Libraries - Encapsulates prompts as callable Python functions with automatic versioning, logging, and multimodal support.
- Prompt - Transforms a Python function into a language model program using its docstring and return value to define system and user messages.
- AI Model Clients - Provides a client resolution mechanism that returns an OpenAI client and fallback flag for a given model name.
- Prompt Chaining - Implements prompt chaining by composing multiple language model programs into sequential sub-tasks with output passing.
- Model Provider Plugins - Provides a plugin interface for registering third-party inference services and custom model providers.
- Multimodal Prompt Embedding - Integrates images, audio, and external documents into prompts via file paths or encoded data using rich type coercion.
- Provider Abstraction Layers - Decouples model selection from client instantiation with a registry for multiple provider backends.
- Multimodal Prompting - Accepts and returns images, audio, and text natively within message objects for multimodal model interactions.
- Versioned Prompt Variants - Tracks every change to a prompt function automatically with generated commit messages and full history.
- Prompt Programming Libraries - Defines a prompt as a callable Python function that combines system instructions, user messages, and model parameters into a reusable subroutine.
- Function-Based - Ships a mechanism to chain several language model calls into a single function for multi-step reasoning.
- Reasoning Pipelines - Chains several language model calls into a single function for multi-step reasoning or test-time compute.
- Prompt Version Trackers - Automatically tracks every change to a prompt function with generated commit messages and full history.
- Prompt Versioning Engines - Captures prompt versions by combining static code analysis with runtime invocation snapshots, generating commit messages automatically.
- Language Model Programming Libraries - Treats prompts as callable Python functions, transforming a Python function into a language model program.
- Model Client Fallback Resolutions - Returns an OpenAI client and a fallback flag for a given model name, decoupling model selection from client instantiation.
- Multimodal Type Coercions - Accepts and returns images, audio, and other media alongside text by coercing them into a unified message object format.
- Prompt and Agent Versioning - Tracks prompt iterations and analyzes performance shifts over time via a local dashboard.
- Temporary Model Registry Overrides - Maintains a global model configuration map that can be temporarily swapped within a context manager for testing or routing.
- Global Default Configurations - Sets global defaults for model parameters, storage, logging, and versioning behavior in a single initialization call.
- Local Interaction Stores - Persists prompt versions, invocations, and metadata to a local filesystem store for offline browsing and comparison.
- Multimodal Output Selectors - Allows returning either plain text or structured message objects with tool calls and multimodal content from a single function.
- Invocation Logging - Captures and stores the full input and output of every language model call for later analysis, fine-tuning, or dataset creation.
- Prompt Output Comparison Dashboards - Provides a local dashboard to visualize, compare, and track prompt outputs across different versions and runs.
- Prompt Version Dashboards - Launches a web-based UI to visualize, compare, and debug prompt versions and their outputs over time.
