Course 07 — RAG & vector search

Retrieval youcan actually see.

Retrieval-augmented generation lives or dies on retrieval. Here every chunk is a point on a map of meaning; your question drops onto it and pulls in its nearest neighbours. Change the chunking, the k, or the ranking — and watch the right answer arrive, or go missing.

Start lesson one Open the sandbox

11 lessons · 5 chapters · free

RAG — a question finding its chunks
  1. 01 · Read

    Each lesson is one idea, explained in a few short paragraphs — no walls of theory.

  2. 02 · Do

    Type into a live environment and watch it react. Tab completes, ↑ recalls, undo takes back any step.

  3. 03 · Match

    Reach the goal any way you like — you're judged on the result, not the keystrokes. Par is the shortest route.

01

Why retrieve

Keyword search, its blind spots, and searching by meaning instead.

  1. 01 Search by keyword The obvious way to find text — and its blind spot.
  2. 02 Search by meaning Vector search finds answers that share no words with the question.
02

Embeddings

Text as points in space, and similarity as distance.

  1. 03 Find it without its words The map of meaning, and a question placed on it.
  2. 04 Similarity is a number Cosine similarity between two pieces of text.
03

Chunking

How you cut documents decides what can be found.

  1. 05 Chunks that are too big Small chunks retrieve precise answers.
  2. 06 Answers cut in half Overlap keeps a sentence that straddles a boundary in one piece.
04

Ranking

Top-k, hybrid search and reranking — getting the best chunk first.

  1. 07 How many chunks? k — when one answer needs two pages.
  2. 08 Hybrid search When the exact words matter, combine keyword and vector search.
  3. 09 Rerank the results A slower, smarter model reorders the shortlist.
05

The prompt

Grounding the model, citing sources, and knowing when not to answer.

  1. 10 Ground the model The retrieved chunks become the prompt — with sources to cite.
  2. 11 Know when not to answer A similarity threshold keeps off-topic questions away from the model.

Every chunk was embedded ahead of time with the open all-MiniLM-L6-v2 sentence model, so the geometry is real. Questions you type yourself are embedded in the browser by averaging that same model's word vectors — close to, but not exactly, a full encoding. The lessons' own questions use exact embeddings.