is Nikolay Nikitin, PhD. I’m the Analysis Lead on the AI Institute of ITMO College and an open-source fanatic. I typically see a lot of my colleagues failing to search out the time and power to create open repositories for his or her analysis papers and to make sure they’re of correct high quality. On this article, I’ll talk about how we may also help clear up this downside utilizing OSA, an AI software developed by our group that helps the repository turn out to be a greater model of itself. For those who’re sustaining or contributing to open supply, this put up will prevent effort and time: you’ll learn the way OSA can routinely enhance your repo by including a correct README, producing documentation, organising CI/CD scripts, and even summarizing the important thing strengths and weaknesses of the undertaking.
There are lots of completely different documentation enchancment instruments. Nevertheless, they give attention to completely different particular person elements of repository documentation. For instance, the Readme-AI software generates the README file, however it doesn’t account for added context, which is necessary, for instance, for repositories of scientific articles. One other software, RepoAgent, generates full documentation for the repository code, however not README or CI/CD scripts. In distinction, OSA considers the repository holistically, aiming to make it simpler to grasp and able to run. The software was initially made for our colleagues in analysis, together with biologists and chemists, who typically lack expertise in software program engineering and trendy growth practices. The primary intention was to assist them make the repository extra readable and reproducible in just a few clicks. However OSA can be utilized on any repository, not solely scientific ones.
Why is it wanted?
Scientific open supply faces challenges with the reuse of analysis outcomes. Even when code is shared with scientific papers, it’s hardly ever accessible or full. This code is normally troublesome to learn; there is no such thing as a documentation for it, and typically even a primary README is lacking, because the developer meant to put in writing it on the final second however didn’t have time. Libraries and frameworks typically lack primary CI/CD settings corresponding to linters, automated exams, and different high quality checks. Due to this fact, it’s not possible to breed the algorithm described within the article. And it is a large downside, as a result of if somebody publishes their analysis, they do it with a want to share it with the group
However this downside isn’t restricted to science solely. Skilled builders additionally typically delay writing readme and documentation for lengthy intervals. And if a undertaking has dozens of repositories, sustaining and utilizing them could be sophisticated.
Ideally, every repository ought to be simple to run and user-friendly. And infrequently the posted developments typically lack important parts corresponding to a transparent README file or correct docstrings, which could be compiled into full documentation utilizing customary instruments like mkdocs.
Primarily based on our expertise and evaluation of the issue, we tried to recommend an answer and implement it because the Open Supply Advisor software – OSA.
What’s the OSA software?
OSA is an open-source Python library that leverages LLM brokers to enhance open-source repositories and make them simpler to reuse.
The software is a package deal that runs through a command-line interface (CLI). It can be deployed regionally utilizing Docker. By specifying an API key in your most popular LLM, you’ll be able to work together with the software through the console. It’s also possible to strive OSA through the general public net GUI. There’s brief introduction to principal concepts of repository enchancment with OSA:
How does OSA work?
The Open Supply Advisor (OSA) is a multi-agent software that helps enhance the construction and usefulness of scientific repositories in an automatic method. It addresses widespread points in analysis initiatives by dealing with duties corresponding to producing documentation (README information, code docstrings), creating important information (licenses and necessities), and suggesting sensible enhancements to the repository. Customers merely present a repository hyperlink and may both obtain an routinely generated Pull Request (PR) with all advisable adjustments or assessment the recommendations regionally earlier than making use of them.
OSA can be utilized in two methods: by cloning the repository and operating it by means of a command-line interface (CLI), or through an online interface. It additionally presents three working modes: primary, computerized, and superior, that are chosen at runtime to suit completely different wants. In primary mode, OSA applies a small set of ordinary enhancements with no additional enter: it generates a report, README, group documentation, and an About part, and provides widespread folders like “exams” and “examples” in the event that they’re lacking. Superior mode provides customers full guide management over each step. In computerized mode, OSA makes use of an LLM to investigate the repository construction and the prevailing README, then proposes an inventory of enhancements for customers to approve or reject. An experimental multi-agent conversational mode can also be being developed, permitting customers to specify desired enhancements in free-form pure language through the CLI. OSA interprets this request and applies the corresponding adjustments. This mode is presently beneath lively growth.
One other key power of OSA is its flexibility with language fashions. It really works with common suppliers like OpenRouter and OpenAI, in addition to native fashions corresponding to Ollama and self-hosted LLMs operating through FastAPI.
OSA additionally helps a number of repository platforms, together with GitHub and GitLab (each GitLab.com and self-hosted cases). It may possibly alter CI/CD configuration information, arrange documentation deployment workflows, and appropriately configure paths for group documentation.
an experimental multi-agent system (MAS), presently beneath lively growth, that serves as the premise for its computerized and conversational modes. The system decomposes repository enchancment right into a sequence of reasoning and execution phases, every dealt with by a specialised agent. Brokers talk through a shared state and are coordinated by means of a directed state graph, enabling conditional transitions and iterative workflows.
README era
OSA features a README era software that routinely creates clear and helpful README information in two codecs: a regular README and an article-style README. The software decides which format to make use of by itself, for instance, if the consumer offers a path or URL to a scientific paper by means of the CLI, OSA switches to the article format. To start out, it scans the repository to search out crucial information, specializing in core logic and undertaking descriptions, and takes under consideration the folder construction and any current README.
For the usual README, OSA analyzes the important thing undertaking information, repository construction, metadata, and the primary sections of an current README if one is current. It then generates a “Core Options” part that serves as the muse for the remainder of the doc. Utilizing this data, OSA writes a transparent undertaking overview and provides a “Getting Began” part when instance scripts or demo information can be found, serving to customers rapidly perceive methods to use the undertaking.
In article mode, the software creates a abstract of the related scientific paper and extracts related data from the primary code information. These items are mixed into an Overview that explains the undertaking targets, a Content material part that describes the primary elements and the way they work collectively, and an Algorithms part that explains how the carried out strategies match into the analysis. This strategy retains the documentation scientifically correct whereas making it simpler to learn and perceive.
Documentation era
The documentation era software produces concise, context-aware documentation for capabilities, strategies, lessons, and code modules. The documentation era course of is as follows:
(1) Reference parsing: Initially, a TreeSitter-driven parser fetches imported modules and resolves paths to them for every explicit supply code file, forming an import map that may additional be used to find out technique and performance requires the international modules utility. By implementing such an strategy, it’s comparatively simple to rectify interconnections between completely different elements of the processed undertaking and to tell apart between inside aliases. Together with the import maps, the parser additionally preserves basic data such because the processing file, an inventory of occurring lessons, and standalone capabilities. Every class accommodates its title, attributes record, decorators, docstring, record of its strategies, and every technique has its particular particulars that are of the identical construction as standalone capabilities, that’s: technique title, docstring, return sort, supply code and alias resolved international technique calls with a reputation of the imported module, class, technique, and path to it.
(2) Preliminary docstrings era for capabilities, strategies, and lessons: With a parser having a construction shaped, an preliminary docstrings era stage is ongoing. Solely docstrings that lack lessons, strategies, and capabilities are processed at this stage. Here’s a basic description of what the ‘what’ technique does. The context is generally the strategy’s supply code, since at this level, forming a basic description of the performance is essential. The onward immediate contains details about the strategy’s arguments and interior decorators, and it trails with the supply code of the referred to as international strategies to offer extra context for processing technique utility. A neat second right here is that class docstrings are generated solely in any case their docstring-lacking strategies are generated; then class attributes, their strategies’ names, and docstrings are offered to the mannequin.
(3) Era of “the primary concept” of the undertaking utilizing descriptions of elements derived from the earlier stage.
(4) Docstrings replace utilizing generated “principal concept”: Therefore, all docstrings for the undertaking are presumably current, era of the primary concept of the undertaking could be carried out. Primarily, the immediate for the concept consists of docstrings for all lessons and capabilities, together with their significance rating primarily based on the speed of prevalence of every part within the import maps talked about earlier than, and their place within the undertaking hierarchy decided by supply path. The mannequin response is returned in markdown format, summarizing the undertaking’s elements. As soon as the primary concept is acquired, the second stage of docstring era begins, throughout which all the undertaking’s supply code elements are processed. At this second, the important thing focus is on offering the mannequin with an unique or generated docstring on the preliminary stage docstring with the primary concept to elaborate on ‘why’ this part is required for the undertaking. The supply code for the strategies can also be being offered, since an expanded undertaking narrative could immediate the mannequin to right some factors within the unique docstring.
(5) Hierarchical modules description era ranging from the underside to the highest.
(6) Utilizing Mkdocs and GitHub pages for automated documentation pushing and streaming: Closing stage of the docstring pipeline, contemplating a recursive traversal throughout the undertaking’s modules and submodules. Hierarchy is predicated on the supply path; at every leaf-processing degree, a beforehand parsed construction is used to create an outline of which submodule is used, in accordance with the primary concept. As processing strikes to greater ranges of the hierarchy, generated submodules’ summaries are additionally used to offer extra context. The mannequin returns summaries in Markdown to make sure seamless integration with the mkdocs documentation era pipeline. The whole schema of the strategy is described within the picture under.
CI/CD and construction group
OSA presents an automatic CI/CD setup that works throughout completely different repository internet hosting platforms. It generates configurable workflows that make it simpler to run exams, examine code high quality, and deploy initiatives. The software helps widespread utilities corresponding to Black for code formatting, unit_test for operating exams, PEP8 and autopep8 for fashion checks, fix_pep8 for computerized fashion fixes, pypi_publish for publishing packages, and slash_command_dispatch for dealing with instructions. Relying on the platform, these workflows are positioned within the acceptable areas, for instance, .github/workflows/ for GitHub or a .gitlab-ci.yml file within the repository root for GitLab.
Customers can customise the generated workflows utilizing choices like –use-poetry to allow Poetry for dependency administration, –branches to outline which branches set off the workflows (by default, principal and grasp), and code protection settings through --codecov-token and --include-codecov.
To make sure dependable testing, OSA additionally reorganizes the repository construction. It identifies check and instance information and strikes them into standardized exams and examples directories, permitting CI workflows to run exams persistently with out extra configuration.
Workflow information are created from templates that mix project-specific data with user-defined settings. This strategy retains workflows constant throughout initiatives whereas nonetheless permitting flexibility when wanted.
OSA additionally automates documentation deployment utilizing MkDocs. For GitHub repositories, it generates a YAML workflow within the .github/workflows listing and requires enabling learn/write permissions and choosing the gh-pages department for deployment within the repository settings. For GitLab, OSA creates or updates the .gitlab-ci.yml file to incorporate construct and deployment jobs utilizing Docker photos, scripts, and artifact retention guidelines. Documentation is then routinely revealed when adjustments are merged into the primary department.
The way to use OSA
To start utilizing OSA, select your repository with draft code that’s incomplete or underdocumented. Optionally, embrace a associated scientific paper or one other doc describing the library or algorithm carried out within the chosen repo. The paper is uploaded as a separate file and used to generate the README. It’s also possible to specify the LLM supplier (e.g., OpenAI) and the mannequin title (corresponding to GPT-4o).
OSA generates suggestions for bettering the repository, together with:
- A README file generated from code evaluation, utilizing customary templates and examples
- Docstrings for lessons and strategies which can be presently lacking, to allow computerized documentation era with MkDocs
- Primary CI/CD scripts, together with linters and automatic exams
- A report with actionable suggestions for bettering the repository
- Contribution tips and information (Code of Conduct, pull request and situation templates, and so on.)
You possibly can simply set up OSA by operating:
pip set up osa_toolAfter organising the surroundings, you must select an LLM supplier (corresponding to OpenAI or a neighborhood mannequin). Subsequent, you must add GIT_TOKEN (GitHub token with customary repo permissions) and OPENAI_API_KEY (when you use OpenAI-compatible API) as surroundings variables, or you’ll be able to retailer them within the .env file as properly. Lastly, you’ll be able to launch OSA straight from the command line. OSA is designed to work with an current open-source repository by offering its URL. The fundamental launch command contains the repository tackle and non-obligatory parameters such because the operation mode, API endpoint, and mannequin title:
osa_tool -r {repository} [--mode {mode}] [--api {api}] [--base-url {base_url}] [--model {model_name}]OSA helps three working modes:
- auto (default) – analyzes the repository and creates a custom-made enchancment plan utilizing the specialised LLM agent.
- primary – applies a predefined set of enhancements: generates a undertaking report, README, group tips, an “About” part, and creates customary directories for exams and examples (if they’re lacking).
- superior – permits guide choice and configuration of actions earlier than execution.
Extra CLI choices can be found here. You possibly can customise OSA by passing these choices as arguments to the CLI, or by choosing desired options within the interactive command-line mode.
As soon as launched, OSA performs an preliminary evaluation of the repository and shows key data: basic undertaking particulars, the present surroundings configuration, and tables with deliberate and inactive actions. The consumer is then prompted to both settle for the instructed plan, cancel the operation, or enter an interactive enhancing mode.
In interactive mode, the plan could be modified: actions toggled on or off, parameters (strings and lists) adjusted, and extra choices configured. The system guides the consumer by means of every motion’s description, doable values, and present settings. This course of continues till the consumer confirms the ultimate plan.
This CLI-based workflow ensures flexibility, from totally automated processing to express guide management, making it appropriate for each fast preliminary assessments and detailed undertaking refinements.
OSA additionally contains an experimental conversational interplay mode that permits customers to specify desired repository enhancements utilizing free-form pure language through the CLI. If the request is ambiguous or insufficiently associated to repository processing, the system iteratively requests clarifications and permits the connected supplementary file to be up to date. As soon as a legitimate instruction is obtained, OSA analyzes the repository, selects the suitable inside modules, and executes the corresponding actions. This mode is presently beneath lively growth.
When OSA finishes, it creates a pull request (PR) within the repository. The PR contains all proposed adjustments, such because the README, docstrings, documentation web page, CI/CD scripts, сontribution tips, report, and extra. The consumer can simply assessment the PR, make adjustments if wanted, and merge it into the undertaking’s principal department.
Let’s take a look at an instance. GAN-MFS is a repository that gives a PyTorch implementation of Wasserstein GAN with Gradient Penalty (WGAN-GP). Right here is an instance of a command to launch OSA on this repo:
osa_tool -r github.com/Roman223/GAN_MFS --mode auto --api openai --base-url https://api.openai.com/v1 --model gpt-4.1-miniOSA made a number of contributions to the repository, together with a README file generated from the paper’s content material.
OSA additionally added a License file to the pull request, in addition to some primary CI/CD scripts.
OSA added docstrings to all lessons and strategies the place documentation was lacking. It additionally generated a structured, web-based documentation website utilizing these docstrings.
The generated report contains an audit of the repository’s key elements: README, license, documentation, utilization examples, exams, and a undertaking abstract. It additionally analyzes key sections of the repository, corresponding to construction, README, and documentation. Primarily based on this evaluation, the system identifies key areas for enchancment and offers focused suggestion.
Lastly, OSA interacts with the goal repository through GitHub. The OSA bot creates a fork of the repository and opens a pull request that features all proposed adjustments. The developer solely must assessment the recommendations and alter something that appears incorrect. For my part, that is a lot simpler than writing the identical README from scratch. After assessment, the repository maintainer efficiently merged the pull request. All adjustments proposed by OSA can be found here.
Though the variety of adjustments launched by the OSA is important, it’s troublesome to evaluate the general enchancment in repository high quality. To do that, we determined to look at the repository from a safety perspective. The scorecard software permits us to judge the repository utilizing the aggregated metric. Scorecard was created to assist open supply maintainers enhance their safety greatest practices and to assist open supply shoppers decide whether or not their dependencies are protected. The mixture rating takes under consideration many repository parameters, together with the presence of binary artifacts, CI/CD exams, the variety of contributors, and a license. The aggregated rating of the unique repository was 2.2/10. After the processing by OSA, it rose to three.7/10. This occurred because of the addition of a license and CI/CD scripts. This rating should still appear too low, however the repository being processed isn’t meant for integration into massive initiatives. It’s a small software for producing artificial knowledge primarily based on a scientific article, so its safety necessities are decrease.
What’s Subsequent for OSA?
We plan to combine a RAG system into OSA, primarily based on greatest practices in open-source growth. OSA will examine the goal repository with reference examples to establish lacking elements. For instance, if the repository already has a high-quality README, it received’t be regenerated. Initially, we used OSA for Python repositories, however we plan to help extra programming languages sooner or later.
You probably have an open repository that requires enchancment, give OSA a strive! We’d additionally recognize concepts for brand spanking new options that you may go away as points and PRs.
For those who want to use OSA in your works, it may be cited as:
Nikitin N. et al. An LLM-Powered Device for Enhancing Scientific Open-Supply Repositories // Championing Open-source DEvelopment in ML Workshop@ ICML25.
