RAG & Embeddings

Put a number on the retrieval

Everyone can draw the pipeline. Almost nobody can say what theirs scores. This is the arithmetic: what two vectors actually rank, what a chunk setting actually produces, what the index quietly dropped, and what recall at five means on a result you are holding.

Start this projectSee the modules
Modules
5
Drills
10
Build steps
4
Time
~4h

Walk out able to

Read a retrieval system you have never seen, write down four integers about it, and say which line of its configuration produced each one.

What you drill

01

Similarity, by hand

Cosines, lengths, and the ranking one division changes.

02

Search and what it drops

Approximate recall, and fusing two ranked lists.

03

Chunks and their edges

Counts, overlap cost, and the boundary that ate the answer.

04

Measuring, then pricing

Recall, precision, first hit — and what moving them costs.

What you get

Skills that stick

Every project mixes short lessons, focused drills and real build steps, so you finish with working code and the judgement behind it.

You will practise

  • Cosines, lengths and what normalising moves
  • Index recall, and fusing two ranked lists
  • Chunk counts, overlap cost and boundaries
  • Recall, precision, first hit, and the price

Afterwards you can

  • Compute a cosine from two stated vectors
  • Count what an approximate index dropped
  • Say which chunk the answer landed in
  • Score a ranked result and price a fix

Modules

Step by step

Each module builds on the last. Read a short lesson, drill it until it's second nature, then put it to work.

  1. 1

    Measure meaning with a vector

    ~42 min

    Compute a similarity by hand, and say which ranking normalising changes.

    • LessonWhat retrieval actually decides4 min
    • LessonRank passages by hand13 min
    • DrillCompute a cosine as a percent13 min
    • DrillSay what normalising changed12 min
  2. 2

    Search without scanning

    ~58 min

    Count what an approximate index dropped, fuse two rankings, and rank a corpus yourself.

    • LessonTwo stages, and what each costs13 min
    • DrillCount what the index dropped11 min
    • DrillFuse two ranked lists13 min
    • Build stepRank a corpus you generated12 min
    • Build stepSee the two metrics disagree9 min
  3. 3

    Split without losing the answer

    ~1h

    Count the chunks a setting produces, and name the chunk a given sentence landed in.

    • LessonLose the sentence that answered14 min
    • DrillCount what a setting produces13 min
    • DrillFind the chunk the answer is in14 min
    • Build stepChunk by sentence, and count11 min
    • Build stepRead what greedy packing left8 min
  4. 4

    Name the stage that failed

    ~38 min

    Read a symptom, say which stage lost the answer, and price the fix in its own unit.

    • LessonRead the symptom, name the stage13 min
    • DrillName the stage that failed13 min
    • DrillCost the cascade12 min
  5. 5

    Measure it, or it is a demo

    ~1h

    Compute the ranking metrics by hand, pick the right one, and score a whole system cold.

    • LessonThree numbers, not one14 min
    • DrillScore a ranked result13 min
    • DrillPick the metric that fits11 min
    • LessonScore a system you have not seen22 min

More in AI & machine learning

Keep going

All AI & machine learning projects

LLM Fundamentals

Tokens are countable. A window is arithmetic.

Intermediate~4h5 modules

PyTorch: Training

Write a loop that learns, and fix one that does not.

Advanced~4h4 modules

PyTorch: Tensors & Autograd

Say the gradient before backward() does.

Intermediate~4h4 modules

Ready when you are.

Work through it at your own pace and pick up right where you left off.

Start RAG & EmbeddingsAll projects