~1 minIn draft
RAG internals: what the embeddings layer is actually doing
Past the vector-database marketing: how text becomes geometry, why chunking strategy decides retrieval quality, and where similarity search quietly fails.
Most RAG explainers stop at "embed your documents, search by similarity." The interesting engineering lives one layer down — what the embedding model preserves and discards, why chunk boundaries decide what can ever be retrieved, and the failure modes that only show up when your corpus is financial data instead of a demo wiki.
What this essay will cover
- How text becomes geometry — and what an embedding model quietly throws away
- Chunking as an architecture decision, not a preprocessing step
- Where cosine similarity fails on numbers, dates, and ledger language
- Evaluating retrieval the way you would reconcile a subledger
This essay is in draft. The full write-up lands here when it's ready.
RAG · Embeddings · AI Architecture