embeddings

README

go-embeddings

This project provides an implementation of API clients for fetching embeddings from various LLM providers.

Currently supported APIs:

• OpenAI
• Cohere
• Google Vertex
• VoyageAI
• Ollama
• AWS Bedrock

You can find sample programs that demonstrate how to use the client packages to fetch the embeddings in cmd directory of this project.

Finally, the document package provides an implementation of simple document text splitters, heavily inspired by the popular Langchain framework. It's essentially a Go rewrite of character and recursive character text splitters from the Langchain framework with minor modifications, but more or less identical results.

Environment variables

[!NOTE] Each client package lets you initialize a default API client for a specific embeddings provider by reading the API keys from environment variables

Here's a list of the env vars for each supported client

OpenAI

• OPENAI_API_KEY: Open AI API token

Cohere

• COHERE_API_KEY: Cohere API token

Google Vertex AI

• VERTEXAI_TOKEN: Google Vertex AI API token (can be fetch by gcloud auth print-access-token once you've authenticated)
• VERTEXAI_MODEL_ID: Embeddings model (at the moment only textembedding-gecko@00 or multimodalembedding@001 are available)
• GOOGLE_PROJECT_ID: Google Project ID
• VOYAGE_API_KEY: VoyageAI API key

Voyage

• VOYAGE_API_KEY: Voyage AI API key

AWS Bedrock

[!IMPORTANT] You must enable access to Bedrock embedding models See here: https://docs.aws.amazon.com/bedrock/latest/userguide/model-access.html#add-model-access

• AWS_REGION: AWS region

Usual AWS env vars as read by the AWS SDKs i.e. AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY, etc.

nix

The project provides a simple nix flake tha leverages gomod2nix for consistent Go environments and builds.

To get started just run

nix develop

And you'll be dropped into development shell.

In addition, each command is exposed as a nix app so you can run them as follows:

nix run ".#vertexai" -- -help

NOTE: gomod2nix vendors dependencies into nix store so every time you add a new dependency you must run gomod2nix generate that updates gomod2nix.toml

Contributions

Yes please!

Documentation

Overview

Package embeddings provides an implementation for fetching vector embeddings from various LLM providers.

Index

• type Base64
  • func (s Base64) Decode() (*Embedding, error)
• type Embedder
• type Embedding
  • func (e Embedding) ToFloat32() []float32

Types

type Base64

type Base64 string

Base64 is base64 encoded embedding string.

func (Base64) Decode

func (s Base64) Decode() (*Embedding, error)

Decode decodes base64 encoded string into a slice of floats.

type Embedder

type Embedder[T any] interface {
	// Embeddings fetches embeddings and returns them.
	Embed(context.Context, T) ([]*Embedding, error)
}

Embedder fetches embeddings.

type Embedding

type Embedding struct {
	Vector []float64 `json:"vector"`
}

Embedding is vector embedding.

func (Embedding) ToFloat32

func (e Embedding) ToFloat32() []float32

ToFloat32 returns Embedding verctor as a slice of float32.