Shipping a small RAG system

I wanted to understand retrieval augmented generation from the ground up, so I built the smallest useful version I could. The pipeline has three stages: chunk documents, retrieve relevant chunks, and pass grounded context to a model.

const chunks = splitIntoChunks(document, { size: 420 })
const results = await vectorStore.similaritySearch(query, 4)
return generateAnswer({ context: results, query })

The biggest lesson was that retrieval quality matters more than prompt cleverness.