The five playgrounds from the lesson, on one page. Everything runs in your browser.
Playgrounds 02 and 05 use real precomputed OpenAI vectors; the rest is
exact browser code.
Playground 01
When do you actually need RAG?
Corpus scale explorerdrag the slider
0
tokens
0
chunks @800
$0
to embed once
0
of a 2M window
Embedding cost assumes text-embedding-3-small at
$0.02 per million tokens. One-off, not per question.
Playground 02
The neighbourhood map
Real embeddings · text-embedding-3-smallclick any word to make it the probe
cat
probe word
—
closest
—
furthest
These are genuine 1,536-dimension OpenAI vectors, precomputed by
tools/export_slide_data.py. The map is a 2-D projection of all 1,536.
Playground 03
Watch the chunks re-cut
Chunking playgroundoverlap is highlighted in each chunk
0
chunks
0
avg chars
0%
duplicated
0
est. tokens
Playground 04
Ask, and watch the prompt build
Retrieval playgroundhighlighted chunks are the ones sent
annual leavelaptop budgetwhen is paydaywhat is the capital of France
0
prompt tokens
0
document tokens
0%
sent to model
0.00
top score
Scoring here uses character-trigram vectors so the page runs with no
API key. The mechanism is identical; demo 07 does the same thing with real embeddings.
Playground 05
Flip the model, break the system
Same store · same question · different query modelreal precomputed scores