I've been playing around with the new Qwen3 models recently (from Alibaba). They’ve been leading a bunch of benchmarks recently, especially in coding, math, reasoning tasks and I wanted to see how they work in a Retrieval-Augmented Generation (RAG) setup. So I decided to build a basic RAG chatbot on top of Qwen3 using LlamaIndex.

Here’s the setup:

Model: Qwen3-235B-A22B (the flagship model via Nebius Ai Studio)

RAG Framework: LlamaIndex

Docs: Load → transform → create a VectorStoreIndex using LlamaIndex

Storage: Works with any vector store (I used the default for quick prototyping)

UI: Streamlit (It's the easiest way to add UI for me)

One small challenge I ran into was handling the <think> </think> tags that Qwen models sometimes generate when reasoning internally. Instead of just dropping or filtering them, I thought it might be cool to actually show what the model is “thinking”.

So I added a separate UI block in Streamlit to render this. It actually makes it feel more transparent, like you’re watching it work through the problem statement/query.

Nothing fancy with the UI, just something quick to visualize input, output, and internal thought process. The whole thing is modular, so you can swap out components pretty easily (e.g., plug in another model or change the vector store).

Here’s the full code if anyone wants to try or build on top of it: GitHub: https://github.com/Arindam200/awesome-ai-apps/tree/main/rag_...

And I did a short walkthrough/demo here: YouTube: https://www.youtube.com/watch?v=L7P8RcKcdzI

Would love to hear if anyone else is using Qwen3 or doing something fun with LlamaIndex or RAG stacks. What’s worked for you?