MiniChain

Free

A tiny library for coding with large language models. ![GitHub Repo stars](https://img.shields.io/github/stars/srush/MiniChain?style=social)

FreeFree tier

Inputs: text, codeOutputs: text, code

Type

Open Source

Visit Website GitHub (40 stars)

About MiniChain

MiniChain is a lightweight Python library for building applications with large language models (LLMs) through prompt chaining. It uses a decorator-based approach (@prompt) to annotate functions that call LLMs, enabling developers to chain multiple prompts together in a lazy graph structure similar to PyTorch. The library separates prompt templates (Jinja) from code, supports multiple backends including OpenAI, Hugging Face, Google Search, Python interpreter, Bash, and Manifest-ML (AI21, Cohere, Together), and can implement popular techniques like Retrieval-Augmented QA, chat with memory, information extraction, interleaved code (PAL), search augmentation (Self-Ask), and chain-of-thought reasoning in just a few lines of code. MiniChain is designed as a minimal, digestible alternative to larger frameworks like LangChain, Promptify, and GPTIndex.

Key Features

@prompt decorator to annotate Python functions as LLM calls

Lazy prompt chaining with graph-based execution (similar to PyTorch)

Separation of prompt templates (Jinja) from Python code

Support for multiple backends: OpenAI, Hugging Face, Google Search, Python, Bash, Manifest-ML (AI21, Cohere, Together)

Built-in implementations of popular techniques: RAG, chat with memory, information extraction, PAL, Self-Ask, Chain-of-Thought

Installable via pip and requires only an API key to get started

Pros & Cons

Pros

Extremely simple and lightweight (tiny codebase) compared to LangChain, Promptify, GPTIndex
Intuitive decorator-based API for defining prompts and chains
Supports multiple LLM backends and tools (including Python and Bash execution)
Separation of prompt templates from logic improves maintainability
Free and open source (MIT license inferred from GitHub repo)

Cons

Limited to Python; no direct support for other languages
Requires external API keys (OpenAI, etc.) for most backends
Not as feature-rich as larger frameworks (e.g., no built-in vector stores, agents, or complex memory)
Primarily a single-developer project (srush) with limited community support

Best For

Rapid prototyping of LLM-powered applicationsBuilding multi-step prompt chains for complex reasoning tasksEducational tool for learning prompt engineering and chainingImplementing and experimenting with research papers (e.g., PAL, Self-Ask, CoT)Creating lightweight chatbots or QA systems with memory

FAQ

What is MiniChain?

MiniChain is a tiny library for coding with large language models. It allows you to annotate Python functions as prompts and chain them together, similar to PyTorch for neural networks.

How do I install MiniChain?

Install via pip: 'pip install minichain'. Then set your OpenAI API key with 'export OPENAI_API_KEY="sk-..."'.

What backends does MiniChain support?

MiniChain supports OpenAI (Completions/Embeddings), Hugging Face, Google Search, Python interpreter, Bash, and Manifest-ML (which includes AI21, Cohere, Together).

What can I build with MiniChain?

You can implement retrieval-augmented QA, chat with memory, information extraction, interleaved code (PAL), search augmentation (Self-Ask), and chain-of-thought reasoning in just a few lines of code.

How does MiniChain compare to LangChain?

MiniChain aims to be a tiny, digestible alternative to larger frameworks like LangChain, Promptify, and GPTIndex. It focuses on core prompt chaining functionality without the complexity.