Case study
NudaUI Semantic Search (RAG)
NudaUI Semantic Search (RAG) is a project by Santiago Gómez de la Torre. This case study covers the problem it solves, what his role was and what came of it.
Updated on
The problem
NudaUI had 1,022 components and keyword search could not find what people described, so “a loader with dots” never led to “Pulse Dots”. It needed search by meaning.
My role
I built the full RAG pipeline without RAG frameworks, with Voyage embeddings, hand-written cosine similarity, a FastAPI service deployed on Vercel and my own golden set of 45 queries across 15 categories, labelled by hand, to measure every change.
Stack
- RAG
- Embeddings
- Python
- FastAPI
- Evals
The outcome
Adding each component's CSS to its embedding raised hit@1 from 66.7% to 80%, hit@5 to 95.6% and recall@5 from 54.9% to 59.4%. The measurement also showed a drop in text effects, from 100% to 67%, which is documented. It is live on nudaui.dev.
Links
- Source, blog.sgomez.dev/rag-busqueda-semantica-nudaui
- Repository, github.com/sgomez-dev/nudaui-rag