RAGLab
Why RAGLab
Retrieval-augmented generation (RAG) has become one of those terms everyone uses and almost no one can fully unpack. Ask a practitioner what happens between “upload a document” and “get a grounded answer,” and most can name the pieces — chunking, embeddings, a vector store, a prompt — but few have built each piece by hand. The frameworks handle it, and the frameworks are good at their job. That’s exactly the problem: good abstractions are invisible, and invisible mechanics are hard to reason about when something goes wrong in production.
RAGLab is a monorepo where I build the same RAG problem more than once, each project taking a different set of tradeoffs, so the tradeoffs stay visible instead of hidden behind a library’s defaults. Every project under projects/ is fully standalone — its own dependencies, its own .env, its own README — so they stay directly comparable and independently runnable.
Projects in this series
- RAG From Scratch — no RAG framework, no vector-database service. Chunking, embeddings, cosine similarity, and prompt assembly, written out plainly. (project page)
More entries land here as new projects ship.
What’s next
Later RAGLab projects will explore the tradeoffs the first one deliberately skipped: a real ANN-backed vector database, hybrid lexical + semantic retrieval, and reranking — each self-contained, each directly comparable back to the baseline.