This is a really great use of LLM! Seriously great job! Once it's fully self-hostable (including the LLM model), I will absolutely find it space on the home server. Maybe using Rupeshs fastdcpu as the model and generation backend could work. I don't remember what his license is, though.
I have some suggestions: let's not make people translate to English unless they are learning English. I don't want to be thinking about whether "I'm coming Friday" is correct grammar in English. I want to be thinking about my target language!
Thanks for the suggestion, I’ll definitely try to make the app as language inclusive as possible!
Also, sorry if I might’ve been too vague with the post title.
The app is just similar to Duolingo in terms of structure and the idea, however it’s not specific to language learning but supposed to cater to any subject, really.
For example, I personally use it to study for my university subjects.
Yeah, it's my minor pet peeve with Duolingo, like source language and my language doesn't have/need suffixes like "the" or "a" so I often forget about it, it's soo annoying to fail because of such minor thing, especially when their suggested English often looks terrible
In some languages that's not a minor thing because of the gender. I mean that's a problem of the language which should improve but for now you have to use the gender for good communication
Is there any interest in getting local models to run using this? I’d rather not use Gemini, and then all the data can reside locally (and not require a login).
I’d be happy to work on this, though I’m a python developer not a typescript one.
Yeah, good idea.
It's possible to do that with WebLLM & Langchain. Once Langchain is integrated, it's kinda similar to the Python Version so should be do-able I think.
@Cr4yfish nice project 🙂
I'm a bit worried about the AI part, though, as you'd want an app whose main purpose is "learning" to guarantee, if not the reliability of the material (since anyone can contribute), at least the reliability of the course generation process that it proposes.
As far as I know, this is not possible with current generative AI tools, so what's your plan to make sure hallucinations do not creep in?
Thanks.
My general strategy regarding GenAI and reducing the amount of hallucinations is by not giving it the task to make stuff up, but to just work on existing text - that's why I'm not allowing users to create content without source material.
However, LLMs will be LLMs and I've been testing it out a lot and found already multiple hallucinations.
I built in a reporting system, although only reporting stuff works right now, not viewing reported questions.
That's my short term plan to get a good content quality, at least.
I also want to move away from Vercel AI & Gemini to a Langchain Agent system or Graph maybe, which will increase the output Quality.
Maybe in some parallel Universe this really takes off and many people work on high quality Courses together...
In case OP doesn’t know, if a repo hasn’t got a licence it’s implied it’s licensed under "all rights reserved", so not open source!
You need to https://choosealicense.com
Thanks, haha.
I'd love develop a Native App for it too but this is a zero-budget Project (aside from the Domain).
PlayStore has a one-time fee so that's 25€ for Android + 8€/Month for the IOS AppStore just to have the App on there.
In theory, I could just have a downloadable .apk for Android to circumvent the fee but most people don't want to install a random .apk from the internet. And I'm not developing a Native App for like 3 people excluding myself (I'm an iPhone user).
This post gathered a bit of traction. So hopefully more people help out. F droid is a better marketplace for oss compared to playstore because people downloading from playstore act entitles a little, especially towards oss software.
In theory, I could just have a downloadable .apk for Android to circumvent the fee but most people don't want to install a random .apk from the internet.
You could've considered F-Droid.
Soo, yeah that'll probably not happen :(.
It's sad. Anyway, good luck with future development!
For all projects/apps, I am looking for OIDC, S3 and PgSQL. It's easier to implement these features earlier and these features will make any projects more popular in the self host community.
In theory not an issue. I use Supabase, which you can self host as well.
You can also self host the Mistral Client, but not Gemini. However, I am planning to move away from Gemini towards a more open solution which would also support self hosting, or in-browser AI.
I am looking for OIDC, S3 and PgSQL
Since I use Supabase, it runs on PgSQL and Supabase Storage, which is just an Adapter to AWS S3 - or any S3, really.
For Auth, I use Supabase Auth which uses OAuth 2.0, that's the same as OIDC right?
Very cool. You can check out ollama for hosting local ai model.
OIDC is an extension of OAuth2 that focuses on user authentication rather than user authorization. Once OIDC authenticates a user, it uses OAuth2 specifications to perform authorization.
The easiest way to support oidc is thru using lib from your framework/language. All major language should already have oidc lib. Take a look for authelia which has pretty nice doc. We host lots of app and we don't want to login hundred times for each apps. It's nice to login once only and all apps play nice to each other ;)
I use Gemini, which supports PDF File uploads, combined with structured outputs to generate Course Sections, Levels & Question JSON.
When you upload a PDF, it first gets uploaded to a S3 Database directly from the Browser, which then sends the Filename and other data to the Server. The Server then downloads that Document from the S3 and sends it to Gemini, which then streams JSON back to the Browser. After that, the PDF is permanently deleted from the S3.
Data Privacy wise, I wouldn't upload anything sensitive since idk what Google does with PDFs uploaded to Gemini.
The Prompts are in English, so the output language is English as well.
However, I actually only tested it with German Lecture PDFs myself.
So, yes, it probably works with any language that Gemini supports.
Here is the Source Code for the core function for this feature:
export async function createLevelFromDocument(
{ docName, apiKey, numLevels, courseSectionTitle, courseSectionDescription }:
{ docName: string, apiKey: string, numLevels: number, courseSectionTitle: string, courseSectionDescription: string })
{
const hasCourseSection = courseSectionTitle.length > 0 && courseSectionDescription.length > 0;
// Step 1: Download the PDF and get a buffer from it
const blob = await downloadObject({ filename: docName, path: "/", bucketName: "documents" });
const arrayBuffer = await blob.arrayBuffer();
// Step 2: call the model and pass the PDF
//const openai = createOpenAI({ apiKey: apiKey });
const gooogle = createGoogleGenerativeAI({ apiKey: apiKey });
const courseSectionsPrompt = createLevelPrompt({ hasCourseSection, title: courseSectionTitle, description: courseSectionDescription });
const isPDF = docName.endsWith(".pdf");
const content: UserContent = [];
if(isPDF) {
content.push(pdfUserMessage(numLevels, courseSectionsPrompt) as any);
content.push(pdfAttatchment(arrayBuffer) as any);
} else {
const html = await blob.text();
content.push(htmlUserMessage(numLevels, courseSectionsPrompt, html) as any);
}
const result = await streamObject({
model: gooogle("gemini-1.5-flash"),
schema: multipleLevelSchema,
messages: [
{
role: "user",
content: content
}
]
})
return result;
}
The UI mostly works offline once loaded in due to aggressive caching.
Downloading Course Content was on the initial Roadmap but I removed it since I wasn't sure if anyone would like the feature.
Syncing stuff is a real pain in the ass but I'll implement it if at least a couple people want it.
I don't know how much of a subset I am, but I still use dictionary softwares from Windows 95~2000 era and Android softwares on a completely offline and vanilla VM, partly due to internet randomly going bad, and partly because I am neurotic about digital contents vanishing once support ends.