An implementation of the OpenAI API for GNU Guile
Andrew Whatson 751cd5db5f Bump version to 0.2 | il y a 1 an | |
---|---|---|
build-aux | il y a 1 an | |
openai | il y a 1 an | |
.dir-locals.el | il y a 1 an | |
.envrc | il y a 1 an | |
.gitignore | il y a 1 an | |
COPYING | il y a 1 an | |
Makefile.am | il y a 1 an | |
README.md | il y a 1 an | |
bootstrap | il y a 1 an | |
configure.ac | il y a 1 an | |
guix.scm | il y a 1 an | |
openai.scm | il y a 1 an | |
pre-inst-env.in | il y a 1 an |
This is a GNU Guile client for the OpenAI API. It currently implements the Chat and Image APIs, with support for chat response streaming and inline image display in a Geiser REPL.
An up-to-date GNU Guix package definition is provided in the repository, so you can install it into your profile with the following:
guix package -f path/to/guile-openai/guix.scm
Alternatively, you can jump straight into an interactive environment by loading it with Guile Studio:
guix shell guile guile-studio -f path/to/guile-openai/guix.scm -- guile-studio
A conventional build system is also provided:
cd path/to/guile-openai
./bootstrap
./configure && make
The library is designed for convenient interactive use, so many useful settings are defined as parameters. Reasonable defaults are provided, so all you need to get started is an API key and a Guile REPL.
To get an API key, you'll need to sign up for an account with OpenAI. Currently new accounts get a certain number of free requests, but ultimately the OpenAI API is a paid service.
(use-modules (openai))
;; Set your OpenAI API key
(openai-api-key "MYVERYSECRETAPIKEY")
;; Simple chat example
(openai-chat "Why is software freedom important?")
;; Simple image example
(openai-image "A cute drawing of a kid in a gnu onesie programming a
robot with their laptop computer")
;; Generate a prompt to generate an image
(let ((rsp (openai-chat "A short paragraph describing a cute mascot for
the GNU Guile Scheme programming language")))
(display rsp)
(newline)
(openai-image (chat-content rsp)))
;; Generate an image and a variation of that image
(let* ((original (openai-image "A whimsical scheme wizard."))
(variation (openai-image-variation original)))
(values original variation))
;; Interactive chat example
(use-modules (ice-9 readline))
(define default-prompt
"You are an AI assistant embedded in a GNU Guile Scheme REPL, which
answers user enquiries in a helpful and playful manner while strongly
advocating for Free Software and the principles of user freedom.")
(define* (interactive-chat #:optional (prompt default-prompt))
(let loop ((msgs `((system . ,prompt))))
(let ((query (readline "> ")))
(if (or (eof-object? query)
(string-null? query))
msgs
(let* ((msgs (append msgs `((user . ,query))))
(chat (openai-chat msgs)))
(newline)
(display chat)
(display "\n\n")
(let* ((text (chat-content chat))
(msgs (append msgs `((assistant . ,text)))))
(loop msgs)))))))
Parameter for the URI of the OpenAI API endpoint.
Parameter for the API key to use for requests.
Parameter for an optional organization ID to include in requests.
Parameter for additional HTTP headers to include in requests.
Parameter for a default value for the #:user
argument of requests.
Send a chat completion request. Returns a chat record.
The prompt can be a string, which will be sent as a user message.
Alternatively, prompt can be a list of (role . content)
pairs,
where content is a string and role is a symbol system
, user
, or
assistant
.
The keyword arguments correspond to the request parameters described in the chat completion request documentation:
#:n
The number of responses to generate, returned as multiple values.
#:stream?
Whether to stream the response(s), defaults to #t
.
#:model
A symbol or string identifying the model to use. Defaults to
gpt-3.5-turbo
, but if you're lucky you might be able to use
gpt-4
or gpt-4-32k
here.
#:temperature
The sampling temperature to use, a number between 0 and 2.
#:top-p
An alternative sampling parameter, a number between 0 and 1.
#:user
An optional username to associate with this request.
The #:stop
, #:max-tokens
, #:logit-bias
, #:presence-penalty
,
#:frequency-penalty
parameters are implemented but untested.
Parameter for the default #:model
value.
Parameter for the default #:temperature
value.
Parameter for the default #:top-p
value.
Parameter for the default #:stream?
value.
Type predicate for chat responses.
Return the chat response as a string.
Return the chat response as a stream of strings.
Send an image generation request. Returns an image record.
The prompt must be a string.
The keyword arguments correspond to the request parameters described in the image generation request documentation:
#:n
The number of images to generate, returned as multiple values.
#:size
A number specifying the generated image size, either 256, 512,
or 1024. The symbols 256x256
, 512x512
and 1024x1024
are
also supported.
#:format
How the image data will be returned, either url
(hosted on a
cloud server), or b64
(embedded in the response message, the
default). In either case, the image data will be fetched and
stored locally for display.
#:user
An optional username to associate with this request.
(openai-image-edit image prompt [keyword arguments])
Send an image edit request. Returns an image record.
The image can be an image record or a string path to the image file to be varied. The API requires the image to be a valid PNG file, less than 4MB, and square.
The *prompt** must be a string describing the changes to be made to the image.
The keyword arguments correspond to the request parameters described in the image edit request documentation:
#:mask
An image record or a string path to an image file to use as the edit mask. The model will only edit pixels which are fully transparent (ie. alpha 0) in the mask image. The mask must have the same dimensions as IMAGE. The default is fully opaque, so no pixels will be edited and a copy of the original image will be returned.
#:n
, #:size
, #:format
, #:user
See openai-image
.
(openai-image-variation image [keyword arguments])
Send an image variation request. Returns an image record.
The image can be an image record or a string path to the image file to be varied. The API requires the image to be a valid PNG file, less than 4MB, and square.
The keyword arguments correspond to the request parameters described in the image variation request documentation:
#:n
, #:size
, #:format
, #:user
See openai-image
.
(openai-default-image-size 512)
Parameter for the default #:size
value.
Parameter for the default #:format
value.
Type predicate for image responses.
The URL of the image, if using the url
format.
The temporary filename containing the image data.
Send an embedding request. Returns an embedding record.
The input must be a string to be embedded.
The keyword arguments correspond to the request parameters described in the embedding request documentation:
#:model
A symbol or string identifying the model to use. Defaults to
text-embedding-ada-002
.
#:user
An optional username to associate with this request.
(openai-default-embedding-model 'text-embedding-ada-002)
Parameter for the default #:model
value.
Type predicate for embedding responses.
The embedding vector generated by the model.
Send an edit request. Returns an edit record.
The input must be a string, which is the text to be edited. The instruction is a string describing the changes to be made.
The keyword arguments correspond to the request parameters described in the edit request documentation:
#:model
A symbol or string identifying the model to use. Defaults to
text-davinci-edit-001
.
#:n
The number of responses to generate, returned as multiple values.
#:temperature
The sampling temperature to use, a number between 0 and 2.
#:top-p
An alternative sampling parameter, a number between 0 and 1.
(openai-code-edit input instruction [keyword arguments])
Identical to openai-text-edit
, except it defaults to the
code-davinci-edit-001
model.
Parameter for the default #:model
value (text edit).
Parameter for the default #:temperature
value (text edit).
Parameter for the default #:top-p
value (text edit).
Parameter for the default #:model
value (code edit).
Parameter for the default #:temperature
value (code edit).
Parameter for the default #:top-p
value (code edit).
Type predicate for edit responses.
The edited content generated by the model.
Send a moderation request. Returns a moderation record.
The input must be a string to be classified by the moderation model.
The keyword arguments correspond to the request parameters described in the moderation request documentation:
#:model
A symbol or string identifying the model to use. The default is
*unspecified*
, which is equivalent to text-moderation-latest
.
(openai-default-moderation-model unspecified)
Parameter for the default #:model value
.
Type predicate for moderation responses.
Whether the content was in violation of OpenAI's content policy.
The list of categories which caused this content to be flagged.
An association list of the scores for this content in each category.
The moderation score for this content in the specified category.
Copyright © 2023 Andrew Whatson whatson@tailcall.au
guile-openai is free software: you can redistribute it and/or modify it under the terms of the GNU Affero General Public License as published by the Free Software Foundation, either version 3 of the License, or (at your option) any later version.
guile-openai is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU Affero General Public License for more details.
You should have received a copy of the GNU Affero General Public License along with guile-openai. If not, see https://www.gnu.org/licenses/.
See COPYING for details.