Category: RAG & Vector Databases

Retrieval-augmented generation is what you reach for when a model has to answer from your documents rather than from its training data. The pipeline is short to describe and long to get right: split the source material into chunks, turn each chunk into a vector, store the vectors somewhere you can search quickly, retrieve the passages that match a question, and hand those passages to the model. Every one of those five steps has a way of quietly ruining the answers. This section works through them one at a time, with runnable Python and with retrieval quality measured rather than assumed. Chunking comes first, because chunk size and overlap decide what the retriever can find at all. Embeddings come next — the choice of model sets the ceiling on semantic matching, and a local sentence-transformer is often enough. Then storage, then the retrieval strategy itself, where combining vector search with classic keyword scoring reliably beats either one alone. This is the newest section on the blog and it is being built out deliberately. If you are assembling a RAG system for the first time, read it in the order below rather than jumping to the database comparison — the storage layer is rarely what makes retrieval bad.