🚀 [RFC] Decoupling RAG Modules

[pancacake]

Hi community! Its DeepTutor Team Here👋

🔍 The Problem

Regarding our RAG modules, we have identified two major "tight couplings" in our current architecture that may limit its flexibility.

RAG Module.

The current RAG module utilizes RAG-Anything (MinerU + Knowledge Graphs, https://github.com/HKUDS/RAG-Anything) to build knowledge bases and perform retrieval. However, we recognize that different users have diverse needs, such as simple vector search or high-speed retrieval, which our current setup may not fully cover.

Environment Config.

Our configuration logic inherits directly from LightRAG (https://github.com/HKUDS/LightRAG), making it difficult to manage system-level settings independently and intuitively.

💡 Idea Collection

We want to build a truly modular and extensible system, and we need your ideas to get it right:

• RAG Strategies: Besides our current Knowledge Graph approach, what RAG backends do you need? (e.g., Simple Vector Search with FAISS/Qdrant, Hybrid Search, or custom reranking pipelines?)
• Configuration Management: How should we structure the new system module? What do you prefer?
• ANY Other thoughs are welcome!

Please drop a comment below with your thoughts or suggestions! Your feedback means a lot to us. :)

🚀 Join Us

We are also looking for hands-on collaborators to shape the future of DeepTutor TOGETHER. If you are interested in participating in the future development and maintanance work, feel free to tell us.

Let's build a better DeepTutor together! We hope our fully open-source project could become a gift for the whole community, for everyone to use.🎁

[Andres77872]

Awesome project. I have experience with RAG systems and retrieval. In the end, the approach depends on the project’s needs, but a strong pipeline is: multimodal retrieval (e.g., ColPali) + hybrid search (combined with RRF) → re-ranking → aggregation.

[pancacake]

Roger that! Will take a comprehensive look! Thanks!

[kushalgarg101]

hey @Andres77872 👋, can i do a part of this pipeline ?

Thanks!

[pancacake]

hey @Andres77872 👋, can i do a part of this pipeline ?

Thanks!

Hi, you could leave an e-mail, or add me on X (Bingxi Zhao @Bingxizhao39) if available, for further discussion~

[pancacake]

kushalgarg9425@gmail.com

Great! Basically, you could try to decoupling the RAG module (include rag_tool, kb_init, logging system...), and add several alternative methods of rag (e.g., naive rag.) I'll send you our plan very soon.

[tusharkhatriofficial]

I’m keen to get involved as a long-term contributor and would be excited to collaborate on the ongoing development and maintenance of DeepTutor. I’d love to help make it even better.

[pancacake]

I’m keen to get involved as a long-term contributor and would be excited to collaborate on the ongoing development and maintenance of DeepTutor. I’d love to help make it even better.

hey! do you use X ,discord or any other platform that you are familiar with? maybe we could have a group to discuss some future features.

[tusharkhatriofficial]

ya sure!

x: @tusharkhatri369
discord username: tusharkhatri

[pancacake]

ya sure!

x: @tusharkhatri369 discord username: tusharkhatri

pls join the discord by : https://discord.gg/aka9p9EW

[srwang0506]

Great project! Looking forward to more awesome updates

[mainaalhajiali3-debug]

Great 👌