Build my path

4e2b8a7course(app-engineer):curriculum preview · demo

Retrieval That Cites Its Sources

AI that searches your own documents first, then answers only from what it found, with the paragraph to prove it.

track
App Engineer
level
Working
length
5 modules · about 12 hours

Your progress dot moves to this branch

foundationsbuildautomatedeploylead

Prerequisites

  • LLM Apps, Properly
  • Some SQL or search experience helps

4e2b011docs:outcomes and build

You will be able to

  • Chunk documents sensibly
  • Rank chunks and inspect the ranking
  • Answer with citations or refuse
  • Measure retrieval with a test set

Hands-on build

A cited answer box

Answers questions about one document set and links every claim to its source.

4e2b022feat:5 modules

Modules.

  1. module 1

    Why look things up first

    Outcome: Know when retrieval helps and when it does not.

    Lessons: What the model does not know · Retrieval (often called RAG) in one picture · When search alone is enough

    free Retrieval that cites or refuses covers this module interactively.

  2. module 2

    Chunking

    Outcome: Chunks that keep one idea each.

    Lessons: Size vs meaning · Headings as boundaries · Overlap and its cost

    free Retrieval that cites or refuses covers this module interactively.

  3. module 3

    Ranking

    Outcome: A ranked list you can inspect and explain.

    Lessons: Keyword scoring · Embeddings, briefly and honestly · Mixing both

  4. module 4

    Cite or refuse

    Outcome: Answers grounded in retrieved text, or a clear refusal.

    Lessons: Thresholds · Citation format · Saying I don't know

    free Retrieval that cites or refuses covers this module interactively.

  5. module 5

    Measuring it

    Outcome: Recall at k on twenty real questions.

    Lessons: Writing questions with known answers · Recall at k · Fixing the worst three

    free Retrieval that cites or refuses covers this module interactively.

4e2b033docs:sample lesson

Sample lesson · Headings as boundaries

Fixed-size chunks cut sentences in half. Splitting at headings keeps each chunk about one thing, which makes ranking easier and citations readable.

Long sections still need splitting. Split them at paragraph breaks and repeat the heading at the top of each piece so the chunk still says what it is about.

4e2b044test:assessment rubric

Ready to merge when

CriterionReady to mergeNot yet
GroundingEvery answer links to the chunk it used.Answers have no sources.
RefusalQuestions outside the documents get a clear refusal.The model guesses.
MeasurementRecall at k is measured on a written test set.Quality is judged by feel.