The RAG pipeline, stage by stage

A RAG system is a pipeline. Knowing the stages — and that each is a place quality can be won or lost — is the map for the whole course.

1. Ingestion / parsing. Load and clean your source documents (PDFs, HTML, database rows, code). This is underrated: if your PDF parser mangles tables or drops headings, everything downstream suffers. Garbage in, garbage retrieved.

2. Chunking. Split documents into retrievable units, usually a few hundred tokens each. How you split matters enormously (Module 2) — chop badly and you sever the context that makes a passage meaningful.

3. Embedding. Convert each chunk into a vector — a list of numbers that encodes its meaning — using an embedding model. Similar meanings produce nearby vectors.

4. Indexing. Store those vectors in a vector database with an approximate-nearest-neighbor index (usually HNSW) plus metadata. Often you build a keyword (BM25) index alongside — you'll want both.

5. Retrieval. At query time, embed the question and find the most similar chunks (and/or keyword-match them). This is the "R" in RAG.

6. Reranking. Take the retrieved candidates and rescore them with a more precise model, keeping only the best few. This two-stage "retrieve wide, rerank tight" pattern is one of the biggest quality levers (Module 2).

7. Generation. Put the top chunks in the prompt and have the model answer — ideally citing which chunk each claim came from.

The classic hybrid version, in Anthropic's own steps: chunk → build both keyword (BM25) and semantic indexes → get top matches from each → fuse and dedupe → add the top-K to the prompt.

Two things to internalize now. First, retrieval quality is where most RAG systems live or die — a perfect model can't answer from chunks you failed to fetch. Second, the pipeline is tunable at every stage, which is why evaluation (Module 3) matters: you'll be changing chunk sizes, swapping embedding models, and adding rerankers, and you need to measure whether each change actually helped. Keep this pipeline in your head; every lesson slots into one of these stages.