Imagine you've just opened a repository you've never seen before.
What do you check first, and what immediately makes you think, "This project is well maintained"?
Every week there's a new model, framework, or benchmark.
But what's the one issue you think deserves more attention from model providers or tooling developers?
I'm interested in hearing what experienced builders think is still missing.
I've been learning RAG over the past few weeks and recently built a small chatbot using Python, LangChain, and Qdrant.
A few things surprised me:
Chunking strategy had a much bigger impact than I expected. Retrieval quality often mattered more than the LLM itself. Embeddings only really clicked for me after experimenting with different retrieval results and seeing how they affected responses.
Before building it, I thought the difficult part would be prompting. In practice, most of my time went into improving retrieval and understanding why relevant information wasn't being returned.
For those who have built RAG systems in production or for personal projects:
What was the hardest concept or problem for you when getting started?
I'd love to hear what challenged you and what eventually made it click.