Video summary
Francois Uray, an Outreachy intern with the Fedora project, introduces "Editorial Guidelines Am I a Llama," an AI assistant designed to help authors adhere to the specific standards of the Fedora Community Blog and Fedora Magazine. The primary challenge this tool addresses is the time editors spend providing repetitive feedback to new authors who often overlook established guidelines during the drafting process. By automating this check, the tool allows writers to identify issues before submission, significantly reducing the back-and-forth communication required for revisions.
The solution leverages a Retrieval-Augmented Generation (RAG) architecture that reads both the official editorial guidelines and real published articles to evaluate drafts against them. This approach was chosen because it grounds the AI's feedback in actual documentation rather than relying solely on pre-trained knowledge, ensuring accuracy even when guidelines change frequently. Unlike traditional fine-tuning methods that require retraining models for every update, RAG allows developers to simply refresh the corpus and vector database with new documents, making the system highly adaptable and efficient over time.
To ensure privacy, eliminate API costs, and align with open-source values, the tool runs entirely on local machines using open-source models like Llama. The technical stack combines these models with Docling for document processing, Core for orchestration, and Streamlit to create an accessible graphical user interface. The project structure is organized into clear sections including the main application file, a data folder containing raw guidelines and reports, scripts for automation, model configurations, and documentation for contributors who wish to reproduce or extend the tool.
The user interface allows authors to paste their draft articles directly into a text area, where the system instantly reviews them against criteria such as the presence of featured images, correct tagging, and proper header usage. For instance, if a draft lacks a required image, the tool immediately flags it as needing revision while confirming that other elements like tags are compliant. Authors can download the detailed review as a text file to make necessary improvements and resubmit their work, creating a streamlined workflow that empowers writers to produce high-quality content independently without waiting for manual editorial feedback.
Read the full video transcript
Uh, my name's Francois Uray, and I'm
an Outreachy May 2026 intern with the
Fedora project. And in this
presentation, I'm going to be talking
about Editorial Guidelines Am I a Llama,
uh, which is a tool that we actually
built as part of this internship.
So, what it actually is and why does it
exist? Uh, it's an AI assistant that
actually checks draft articles against
Fedora's editorial guidelines, uh,
running entirely on your local machine
using open models.
So, what is the problem that we actually
trying to solve here? Uh,
Fedora Community Blog and Fedora
Magazine actually have editorial
guidelines or standards.
New authors don't always know all of
them, and then they draft articles
and submit them. Editors now spend a lot
of time on repeated feedback, allowing
these new authors to meet the guidelines
before the drafts are published
officially.
Now, the idea, uh, that we actually came
up with, uh,
in building this tool, is a RAG tool,
uh, that actually reads the guidelines
and real published articles,
and it tells an author whether their
draft actually meets the guidelines or
not.
Why did we choose RAG? Uh,
it's grounded in the actual guidelines,
first of all,
explaining why and how to fix
the guidelines if they are not respected
in the drafted article.
Um, updating the corpus and also not
retraining all the time because with
RAG, you just keep enhancing the RAG
model rather than fine-tuning all the
time.
Uh,
another reason you can think think of,
uh, to use RAG is because guidelines
actually may change all the time, right?
Um, and it's actually easier to update
the corpus and vector database than to
fine-tune a model every single time
there's an update in guidelines.
Why did we choose to go with locals or
like open models? First of all, it's
because of privacy.
And second of all, it's because there is
no API costs, um, associated to them.
And open models actually align better
uh, with open source eaters.
And of course, they are reproducible.
So, anyone can actually take it and
reproduce them.
So,
what are the two publications that this
tool actually covers? Of course, it is
the Fedora Community Blog and the Fedora
Magazine,
each with its own like editorial
guidelines.
So, the overall stack that we actually
use is first of all Rama Lama,
then Docling,
then Core,
Streamlit, what do you have? I mean,
that launches the GUI, the graphical
user interface.
And then small GPT generated unified
formats models.
Now, about the um
project structure. So, basically,
you would actually find like the app.py,
which is the Streamlit app that launches
the the GUI, which we will see in a
second, uh at the root of the project.
And then you have the data folder in
which you will see the raw data, the
source, the guidelines, reports, and
reviews.
And you have the scripts,
which we will get into in in in like in
an upcoming video.
And then the models themselves, the Rama
Lama configuration,
a notebook that actually allows you to
run
and see a graphically the processing.
And then documentation
for
he or she who wants to uh actually uh
use the tool and
um how to contribute or like what are
the things that they really need to
reproduce this tool.
Now, here is for instance the GUI, I
mean, like the Streamlit app that I was
mentioning.
Um so, as you can see, it takes in as an
input um
a draft article
that you actually paste here
in this
like text area.
And then you click on review article.
So, it will review this article against
the Fedora Community Blog guidelines.
For instance, this one doesn't have like
an image.
So, when you check it, it says, "Okay,
it needs revision."
Uh featured image failed.
So, there is no featured image
included, which the person has to do,
the author.
And then add a tags. The tool actually
passed. The headers actually passed
because there are and so on and so
forth. So, like what's really important
is with this is the fact that you
actually run it
rather than waiting for editors to sort
of come back to you with repetitive
feedback all the time. And then the tool
actually gives you feedback, and then
you can actually download the review as
a TXT file, and improve your draft of
article,
and resubmit it again, uh making sure it
actually matches all like meets all of
the guidelines or the standards.
Thank you so much for listening.
We will continue with the next video.