How the Fedora Editorial Assistant Works

All right, so in this part of the presentation we are going to take an article and we will try to move from raw data to grade up review. But first we will try to pull or like to fetch articles from the Federal uh com blog and then also the Federal magazine using WordPress API and we will end up like sort of setting the guidelines as the ground truth and then do ingesting and like chunking using Ram Rag that runs like Duckling internally to do the parsing and the chunking. Uh there is a lesson that I just want that we learned and that I just want to mention here is that the more like higher the number of chunk or like the chunk size is uh it's it's likely that you will run into like uh like short short like storage issue, right? Uh because at like at first we were running it entirely like on CPU which actually sort of like led my mentors to provisioning um a GPU environment in which now we do the work. So the the fourth step is like building the vector store uh uh Ram Rag in this case actually sort of produces um OCI images like open container images that are hosted or like published on Quay from which you actually now do the your query and then at the end we did model evaluation using like Rag Eval which is a Rag evaluation tool. Um so we do have the repo itself. So here's the repo for this tool which you can actually test out. Um So basically here is what the pipeline actually sort of looks like. So you have the WordPress REST API that fetches the articles from like Federal Account blog and then Federal magazine and then you actually end up getting like raw data and then like source data in JSON and HTML format. And then these are actually sort of converted into markdown because that's what you actually use for the writing enhancement. And then Ram actually with Duckling that runs internally. So at the end you build the vector store and you get these uh OCI images that I was there on Quay. And which would actually find here on Quay uh under like this Fedora Quay uh like repo. So you'd find all of these uh uh the three OCIs. I like the three open container images. So the first one for the community blog, the magazine is in an editorial just means combining both the community blog and the magazine. Um I'll come back to this. So basically um the problem with chunks as you see is something you were actually going to read uh here in the repo. And then at first you just try to clone the repo and you CD into it. Uh make sure you create like um an environment like a virtual environment. And then the requirements.txt file is already created in the repo, so you can install it. And then you set your environment variable with this. So this is how you fetch your articles using [clears throat] the WordPress API. Uh uh the files are also in the repo. And you stage the corpus. You can try to regenerate like a quality report. It's uh just analyze the data set quality before you start like your writing enhancement. And then here is how you actually build the right vector store. We just truncate 100 because at first uh which is what you see here, we run into like uh memory outage uh because of the the the the number of tokens that it takes. But now we do have a working environment environment provided by the community through my mentors and then we increase the chunk size. Um so you can push them uh which I just showed you uh on Quay a few seconds ago. And you can also now try to run it directly using the three uh Streamlit uh app. Which is here. And then the app actually uh it's either you can check it against like Fedora community blog articles or like Fedora magazine articles or like both publications. So, you have this text area where you actually paste a a draft article and then you just click on review article, it gives you like a review after after it runs. Uh we we chose two models for now. Um So, we chose the Gmail, which is the recommended one uh for now and then like Granite. Um and then as you can see here, I actually took uh Fedora article and then I pasted it here and then after that it says, "Okay, uh first of all, there is no featured image." So, this one failed. Um the says, "Okay, more tags than easy space after this one." and so on. Like at the end, it will actually provide you with a final verdict. Like it needs a revision, of course. So, it is that the draft has a good tone and structure, gameful but is extremely brief and lacks all necessary metadata and publishing like elements. So, you can download like the review uh any author can actually download the review to a txt file and make the improvement in their draft and resubmit again, making sure they all work and then yeah, have it published in the Fedora Account Lab or like magazine. Um the other thing is you can actually run it directly uh again. Uh Uh first of all, like when you get back here, there is a doc folder a doc folder in which you see this guide.md that gives you like a guide of how to reproduce this work on your local machine or like uh locally. Anyways, um so, these are like the installations that you have to do like the prerequisites. Um and then you clone the repo, you see the into it and then this is the expected output when you just try to find uh what's in the repo. Like creating the Python the the virtual environment and then you activate it. Uh so, here you try to install like the requirements. And then yeah, creating just like a .env and then like uh now creating like your environment variables. And after that, uh just editing this one these ones and then you can now fetch the articles here. If you run like the fetch article directly, it will fetch both. If you want to fetch only like um either like the com blog articles or like the magazine articles. So, you could do that too by giving this {dash} {dash} source and then com blog or like magazine. Uh so, this is the expected output. Uh it would actually maybe tell you that okay, there is this much number of articles that were fetched and state and saved into data {slash} raw {slash} com blog and 40 magazine, too. Then you can verify the output with this. Um so, you could analyze like the data set quality, as I said uh before. And then here like the expected output would just give you a uh like after the analysis, how many of those articles actually uh have good quality for your writing enhancement. And then how many of them like passed and how many of them like failed. Uh so, you can do the review and then now you pull the models like G mine granite. And then now you build your the vector store and then you push them to Qdrant. You make you make sure they're on Qdrant. Um now you can go ahead and try and test it, right? So, you could test them either like on using the Streamlit app, which is actually really intuitive for anybody who does not have maybe like a tech background who just want like a GUI, a graphical user interface, where they would actually come and uh paste in the draft article and then just click on uh review article. Or you could do it using the CLI, right? I just prepared this to save us some time. So, you run it from a llama run, right? And then you give it the uh the Qdrant image and then you call the model. So, you wait for it and then it works also like um on the command line interface. So, basically ask you the question, uh what tone should a future article in your blog article actually use? It says that the tone has to be a Yeah, the tone of a future blog article should be a careful balance between technical authority and community warmth. So, it gives you the pillars, what are called what is the core philosophy, what you have to avoid here. Uh adjusting the tone of my topic, you know, for technical tutorials, and how to guide debugging, and so on. So, like at least you see that you can also run it like if you're really like tech savvy, you can run it using the CLI. Another way of doing it is you can run it using like uh because when you check there is a Yeah. Okay, like a Jupyter notebook that you can actually try to launch. Uh So, I'm just going to copy this one here and then come back and paste it here. Click. And then it will launch a notebook where you can actually sort of visualize everything, right? So, here you try to do all of the imports, like just the setup. Now, here is how it where you fetch the articles using WordPress API. And after that, you do the quality report analysis to remove articles that are not of good quality, you know, for your writing enhancement enhancement. And chunking strategy. And so on. Now, the write testing, you have all of these, which were actually over there previously what I was showing in the repo. Um so, yeah, you could actually try to run this and also visualize using the notebook directly. Or you could use the CLI or the Streamlit app, which is the recommended one for everyone regardless of your your competencies like in in in in in tech. So, basically, that's it for all this one. And you can do some iterations. Um You can also refine like your system prompt. So, here's the full pipeline, and yeah. So, basically, that's it. And then when you check this docs folder inside the repository, there are like guidelines on how to set it up, how to reproduce it, and also if you if I want to contribute, there is a contribution guide. Uh and also like other details on uh what's needed for this project to scale or to be reproduced. Thank you so much for listening and uh uh there is a next video that will give a lot more details about like the internal aspect of this one. Thank you.