🟠 Draft | Preloads and concatenate files from a specified directory to build a prompt for your LLM
Go to file
Glenn Y. Rolland e3e091974d
continuous-integration/drone/push Build is passing Details
doc: use the new config format
2024-01-24 14:18:00 +01:00
misc doc: add example prompts 2024-01-23 16:23:11 +01:00
spec fix: make unit tests work with CODE_PRELOADER_DETECT=no 2024-01-05 11:48:12 +01:00
src fix: ensure files are sorted by name 2024-01-23 16:23:32 +01:00
static feat: add defaults (template & config) as files 2024-01-04 23:03:21 +01:00
.code_preloader.yml feat: use embedded filesystem for default files 2024-01-04 23:14:14 +01:00
.drone.yml ci: add unit testing 2024-01-04 11:58:55 +01:00
.editorconfig chore: add editorconfig 2024-01-02 12:32:59 +01:00
.gitattributes chore: enable LFS 2024-01-04 11:53:26 +01:00
.gitignore chore: do not ignore docs dir 2024-01-04 11:52:43 +01:00
LICENSE doc: add LICENSE 2023-12-29 15:55:16 +01:00
Makefile chore: do not force owner root in install (allow local install) 2024-01-03 18:01:36 +01:00
README.md doc: use the new config format 2024-01-24 14:18:00 +01:00
Vagrantfile feat: add vagrant provisioning scripts 2024-01-03 14:06:58 +01:00
shard.lock feat: use embedded filesystem for default files 2024-01-04 23:14:14 +01:00
shard.yml feat: use embedded filesystem for default files 2024-01-04 23:14:14 +01:00

README.md

Build Status License LGPL3.0-or-later Donate on patreon

This project is available on our self-hosted server and on CodeBerg and GitHub as mirrors. For the latest updates and comprehensive version of our project, please visit our primary repository at: https://code.apps.glenux.net/glenux/code-preloader.

Code-Preloader

Code-Preloader is a specialized tool designed to streamline the process of working on coding projects with interactive large language models (LLM) like ChatGPT, Claude, Mixtral 8x7B, etc.

It preloads and concatenates files from a specified directory, allowing for the seamless integration of customized prompts.

This tool is ideal for those who seek an alternative to tools like GitHub Copilot, enabling a tailored interaction with your favorite LLM.

Features

  • Preload and concatenate files from a given directory.
  • Customizable header and footer prompts for your LLM.
  • Simple command-line interface for easy operation.

Limitations

Note: This tool is optimized for smaller codebases. For larger repositories, performance may not be optimal due to processing constraints and the nature of interactions with LLMs.

Prerequisites

Before installing and using Code-Preloader, make sure your system meets the following requirements:

  1. Crystal Language: Code-Preloader is written in Crystal. Ensure you have the Crystal programming language installed on your system. For installation instructions, refer to the official Crystal language website.

  2. Required Libraries: The following libraries are necessary for the proper functioning of Code-Preloader:

    • libevent: Used for asynchronous event notification.
    • libyaml: Required for YAML parsing.
    • libmagic: Utilized for file type detection.
    • make: Used to define compilation rules

On a Debian-based system, you can install these libraries using the following command:

sudo apt-get install libevent-dev libyaml-dev libmagic-dev make

Installation

To get started with Code-Preloader, ensure that you have the prerequisites installed on your system (see above).

Then follow these steps to install:

git clone https://code.apps.glenux.net/glenux/code-preloader
cd code-preloader
make prepare
make build
sudo make install                 # either to install system-wide
make install PREFIX=$HOME/.local  # or to install as a user

Usage

Packing directory content

Run Code-Preloader with the following command-line options:

Usage: code-preloader pack [options] DIR ...

Global options:
    --version                        Show version
    -h, --help                       Show this help

Pack options:
    -i REGEXP, --ignore=REGEXP       Ignore file or directory
    -o FILE, --output=FILE           Write output to FILE
    -H FILE, --header-prompt=FILE    Load header prompt from FILE
    -F FILE, --footer-prompt=FILE    Load footer prompt from FILE
    -c FILE, --config=FILE           Load parameters from FILE

Basic Use Case

To preload all files in the src directory and output to result.txt, while ignoring the git the bin directory, and the result file itself:

./bin/code-preloader pack -o result.txt -i .git -i result.txt -i bin/ src

Advanced Use Case

To preload all files in the src directory and output to clipboard, prepending and appending prompts, while ignoring the git the bin directory, and the result file itself:

./bin/code-preloader pack \
    -i .git -i bin/ -i result.txt -i prompts \
    -H prompts/context.txt -F prompts/request-readme.txt \
    src \
    | ctrlc

Note: the command ctrlc in previous command is an alias to xclip -selection clipboard -i

Using a config file

You can automatically create an empty configuratio file by running Code-Preloader with the following command-line options:

Usage: code-preloader init [options]

Global options:
    --version                        Show version
    -h, --help                       Show this help

Init options:
    -c FILE, --config=FILE           Load parameters from FILE

Example: Advanced with configuration file

Create an empty configuration file with

./bin/code-preloader init -c .code_preloader.yml

Adapt the configuration file (.code_preloader.yml) to your needs:

---
ignore_list:
  - .git
  - code_preloader.yml
  - bin
  - prompts

prompt:
  header_path: prompts/context.txt
  footer_path: prompts/request-readme.txt
  template_path: null

output_file_path: null

Then you can type a shorter command like:

./bin/code-preloader pack -c .code_preloader.yml src | ctrlc

Note: the command ctrlc in previous command is an alias to xclip -selection clipboard -i

Prompting efficiently with CodePreloader

The goal is to maximize the effectiveness of your interactions with large language models (LLMs) using CodePreloader for software development tasks.

  1. Preload Your Project: Before starting a new task, ensure you preload your project. This prepares the LLM with the necessary context, making it ready to assist you effectively.
  2. Utilize Structure Separators: Use "@@" as a delimiter to clearly separate and introduce new content. This could be for final requests, additional files, or other relevant data.
  3. State Your Goals Clearly: Begin by explicitly informing the LLM of your objectives. A clear and concise explanation of what you aim to achieve helps the model understand your needs better.
  4. Reference Preloaded Files: If necessary, refer to preloaded files by using "@@ File ...". This directs the LLM's attention to specific parts of your project.
  5. Direct Focus Using Specific Phrasing: Encourage the LLM to "focus" or "concentrate" on particular elements of the preloaded context, like files, classes, methods, etc. These specific terms have been observed to enhance the model's attention to the relevant details.
  6. Approach the Task in Stages: Divide your request into multiple stages for clarity and precision:
    • First Request: Ask the LLM to analyze the content based on your instructions and explain the approach it would take for the changes requested. At this stage, request the LLM to refrain from writing code.
    • Second Request: Instruct the LLM to focus on both the content and its previous analysis. Request a proposed fixed version of the specific file, class, or method based on the prior analysis. The more detailed your request, the better the outcome. Ask for changes to be made one at a time.
    • Third and Subsequent Requests: Follow the same pattern as the second request but apply it to different parts of the project that require changes.

Example: improving a README.md

First request:

@@ REQUEST (ANALYZE)

We want to improve the README.md file as for an open-source project following the best practices.

Please focus on the code provided in the « @@ File ... » sections and analyze the existing « @@ File "README.md" », then take in account the content of class Config which analyzes the command line.

In this analysis of the existing « @@ File "README.md" », please provide me with a list of things that you appreciate, a list of things to amplify, a list of things to adjust.

Second request:

@@ REQUEST (WRITE)

We want to improve the README.md file as for an open-source project following the best practices.

Please focus on the code provided in the « @@ File ... » sections and analyze the existing « @@ File "README.md" », then take in account the content of class Config which analyzes the command line.

Please concentrate on the detail of your analysis in your previous message and write a new improved/fixed version of « @@ File "README.md"

Example: writing a new feature

First request:

@@ REQUEST (ANALYZE)

I would like to change CLI parameters structure add add two subcommands:

  • init : which will create an example .code_preloader.yml file (with comments)
  • pack : which will create the packed version of the current directory for LLM prompting

Most of current options (except --version and --help) must become options of the pack subcommand.

I already started some changes to achieve this goal, but it is not finished, and I need your help and expert advises.

Can you please tell me :

  • where the changes should occur (which file? which class? which method?)
  • what kind of changes must be made there?

Please do not write code yet, simply explain.

Second request:

@@ REQUEST (WRITE)

I would like to change CLI parameters structure add add two subcommands:

  • init : which will create an example .code_preloader.yml file (with comments)
  • pack : which will create the packed version of the current directory for LLM prompting

Most of current options (except --version and --help) must become options of the pack subcommand.

Please concentrate on the detail of your analysis in your previous message and write a new improved/fixed version of « @@ File "README.md". Please keep in mind the separation of concerns and the single responsibility principle.

Please provide me a new version of « File ... » which include the requested change?

Contributing

Contributions are what make the open-source community such an amazing place to learn, inspire, and create. Any contributions you make are greatly appreciated.

Troubleshooting and Support

If you encounter any issues or need support, please open an issue in the project's issue tracker. We strive to be responsive and helpful.

License

Distributed under the LGPL-3.0-or-later License. See LICENSE file for more information.

Acknowledgments

  • A special thanks to all contributors and users of this project for their valuable feedback and support.
  • Inspired by the community's need for efficient code preparation tools in the context of AI-assisted programming.