j2pipeline 0.4.0

Create LLM pipelines with ease!

J2pipeline

Installation

Create Python 3.12 virtual environment, activate it and run

pip install j2pipeline

Setup

J2pipeline requires an active TCP server listening on port 50027

Server:

• Receives a prompt
• Sends back AI response

Protocol:

1. 16 bytes header (message length)
2. Message (in UTF-8)

For example:

data: bytes = message.encode(encoding='UTF-8')
size: bytes = f'{len(data):16}'.encode(encoding='UTF-8')
server.send(size)
server.send(data)

Server

• For personal use, you might consider installing a Telegram GPT server available at https://github.com/yaz008/TG-GPT-API

• You can write your own server following the described protocol

Usage

Prompt

Prompt creates a function based on a single template file

Params

• path: path to a template file
• process: function that modifies LLM output
• auto_upper: makes sure all arguments passed in the template are in upper case

Example

Suppose we have a translate.j2 file

Translate the following text into {% LANGUAGE %}:

{% TEXT %}

From this template we can than create translate function

from j2pipeline import Prompt

translate: Prompt[str] = Prompt[str](path='translate.j2')

text: str = 'Quelle heure est-il?'
translation: str = translate(language='English', text=text)
print(translation)

In the output we shall see

What time is it?

Pipeline

Pipeline creates a function based on a sequence of templates

Params

• base_path: path to a folder with templates
• reorder: function that takes a list of files and returns reordered one
• process: a dict where keys are template names and values are respective modifiers
• final_process: function that processes the final result
• extension: common extension of the template files

Example

Suppose we have two template files in yoderize folder

• yoda.j2:

You are Yoda from Star Wars, so you must speak like him
Rewrite this text in style of Yoda

{% PROMPT %}

• translate.j2:

Translate the following text into German preserving the style of Yoda:

{% PROMPT %}

Your response must only contain the translation

From this template we can than create yoderize function

from j2pipeline import Pipeline

yoderize: Pipeline[str] = Pipeline[str](base_path='yoderize',
                                        reorder=lambda _: ['yoda', 'translate'],
                                        final_process=lambda text: f'German Yoda:\n{text}')
yoderized: str = yoderize(prompt='I give you an advice: you should be wise')
print(yoderized)

In the end we shall see

German Yoda:
Weise, sein solltest du. Rat, gebe ich dir.

License

J2pipeline is a free, open-source software distributed under the MIT License