adithya-s-k

I have been building in AI for the past year, and one of the most frequent questions I get from peers and juniors is, "How do I get started?" To answer that, I began documenting what I’ve learned along the way. From Prompt Engineering and Retrieval-Augmented Generation (RAG) to fine-tuning large language models, deploying AI systems, and working on real-world projects, I’ve organized it all into one comprehensive roadmap.

I truly believe in learning by applying, so the repository is structured in a way that allows you to run any notebook easily. You can dive straight into hands-on projects, experiment with code, and see concepts in action as you learn.

This resource is entirely open source and designed to help anyone interested in AI get started and progress through structured, practical learning. Whether you're a beginner or looking to advance your skills, you’ll find tutorials, insights, and real-world applications to guide you.

I've been working on various Retrieval-Augmented Generation (RAG) projects for a while now, and I've noticed a common pattern: creating the pipeline isn't the hard part anymore. With the growth of frameworks and libraries, setting up a RAG pipeline to handle document retrieval, vector embeddings, and LLM generation is quite straightforward. The tools are there, the codebase is cleaner, and we have better control over the data flow.

But the real challenge comes when you want to put these pipelines in the hands of actual users. Integrating RAG into a user-facing application is far from straightforward. It involves a lot of custom integration work, UI development, backend scaling, handling user interactions, and ensuring the system remains responsive and efficient. And let’s be honest, this is a pain point for most of us who just want to focus on building smarter AI-driven applications.

This is where a RAG SaaS solution comes in. Imagine just plugging your existing pipeline into a platform and having a user-ready application out of the box. No more spending weeks (or even months) building out a custom UI or dealing with the nitty-gritty of server-side optimization. You can deploy, iterate, and share a fully functioning application with your end users or stakeholders in a fraction of the time.

I believe this kind of SaaS approach can dramatically speed up the cycle of innovation for anyone working on RAG solutions. It lets us do what we love—focusing on refining the model, the data retrieval techniques, and the overall pipeline—without getting bogged down in the labor-intensive process of turning it into a production-ready app.

Would love to hear thoughts from others who’ve faced similar hurdles. Is the future of RAG about better pipelines, or better ways to deploy them?

Some of the softwareslike LibreOffice, are used to convert files from one format to another.For Windows,it will require a different approach which hasnt been implemented yet

Yep, the accuracy it currently offers is 80% to 90%. We are actively working on improving the underlying models and there are some major improvments coming soon

Yes, the current implementation of the repository converts any data primarily into strctured markdown text.

The next stage will involve prompt guides or schema-guided structure extraction.

Let's say you are processing a lot of research PDFs and want to convert them into clean markdown that best represents the content. Now, let's say you want to extract the authors, abstracts, captions, and store images.

The extraction engine we are currently working on will help you with that.

Oh, I will do some more research on LlamaParse.

Yep, planning to release it under a commercially permissible license.

We have an active API which we are using for our internal clients, and we are planning to release it soon.

Regarding the ETA of the new model, I don't have a fixed deadline as we are training and testing for a lot of edge cases. Currently, we are doing research and trying to build/train in public on X/Twitter.

Llamaparse currently only parses PDF documents, as far as I know. OmniParse aims to process any data type, from documents and images to videos and websites, and provide the best representation for AI applications.

We have a few dependencies that are licensed under GNU, which is why we have that license. However, I am currently training models to be under the MIT license and plan to replace the current GNU-licensed dependency to eliminate this limitation.