Every LLM feature looks great in the first ten prompts you try during a demo. The gap that shows up in production isn't usually the model, it's that nobody built a way to know when it's wrong.
Before we ship an LLM feature, we build the evaluation set first: real examples pulled from the product, with a clear right answer, checked automatically on every change. It sounds like extra work up front, and it is, it's also the difference between catching a regression in code review versus catching it from an angry support ticket.
For retrieval-augmented features specifically, the retrieval step is where most failures hide. Teams spend weeks tuning the prompt and days on the thing that decides what the model even sees.
- LLM
- AI
- RAG
Written by Mustafa Halvadwala
Engineering Lead at HASANI.
