lingograph

lingograph is a Go library for building LLM pipelines. It provides a flexible, composable way to create complex conversation flows using pipeline combinators.

Core Concepts

lingograph is built around the following concepts:

Actor

An actor is a component that processes conversation history and generates responses. Actors can be:

• OpenAI LLM invocations with system prompts
• Custom implementations with specialized behavior

OpenAI-based actors can seamlessly invoke Go functions. These functions receive structured Go data directly—no need to manually parse JSON or other formats. This is made possible through the OpenAI Functions API, combined with Go reflection to minimize boilerplate.

Pipeline

A pipeline represents the overall structure for processing history and generating responses. Pipelines are built from actors and composed using combinators:

• Chain: executes a sequence of steps in order
• Parallel: runs multiple pipelines concurrently
• While: repeats a pipeline while a predicate over the store holds true
• If: conditionally executes one of two pipelines based on a predicate over the store

Store

The store provides a way to maintain state between pipeline steps. It supports type-safe variables that can be shared across different parts of the pipeline. Store variables can be modified from within functions called by actors and are particularly useful with conditional pipelines, allowing you to branch the execution flow based on runtime conditions.

Quick Start

With a working Go installation, run the following command in your project directory:

go get github.com/vasilisp/lingograph

Then explore the examples directory to see how the core concepts fit together.

Example Projects

The following projects demonstrate lingograph:

• velora: An AI-powered command-line workout tracker and coach. It uses lingograph to build complex pipelines for analyzing workout data and producing training recommendations.

• wikai: A Git-powered note-taking app with AI capabilities. It builds a RAG system utilizing lingograph pipelines with heavy usage of the (OpenAI) function interface.