Chunk slicer
Drag the chunk-size and overlap sliders and watch a RAG pipeline's precision, recall, and cost move in step — with an F1 curve that rises then crashes, so the sweet spot is visible, not guessed. Switch chunking strategy (fixed / sentence / semantic) and see boundaries stop cutting sentences in half. Runs entirely in your browser (0 uploads, works offline).
How to use it
- Set the chunk size, overlap %, and strategy.
- Watch the precision / recall / F1 bars and the cost readout move, and the F1 curve mark where you are versus the optimum.
- Read the X-ray: the precision-context-cost triangle you can't max all at once, and roughly where F1 peaks.
What this clears up (the fundamentals)
- Bigger chunks aren't better — F1 rises then crashes with chunk size. There's an interior sweet spot, not a monotonic "more context wins."
- The triangle is real — tiny chunks lose the surrounding context a fact needs (recall drops); huge chunks bury the answer in irrelevant text (precision drops) and cost more tokens per query. You trade among three linked quantities.
- Boundaries matter — fixed-size cuts slice sentences (and facts) in half; sentence and semantic boundaries recover precision by keeping meaning intact.
- Overlap buys recall for cost — repeating tokens across chunk edges catches facts that straddle a boundary, at the price of more tokens retrieved per query.
Where it's used
A RAG-engineering literacy tool — the fastest way to feel why retrieval quality depends on chunking, and to find a starting chunk size before you index a corpus. It's a miatz build-lab concept playable; learn to build it yourself.
FAQ
Why does F1 rise then crash with chunk size?
Because precision and recall move in opposite directions as chunks grow. Recall climbs (a bigger chunk is more likely to contain the whole answer) while precision falls (more irrelevant text dilutes the match). Their harmonic mean, F1, peaks where the two balance.
What does overlap do?
It repeats some tokens across neighbouring chunk boundaries so a fact that straddles a boundary still lands wholly inside at least one chunk — buying recall at the cost of more tokens stored and retrieved.
Which strategy should I use?
Sentence and semantic boundaries avoid cutting a thought in half, which recovers precision over naive fixed-size slicing. They cost a little more to compute but usually lift retrieval quality at the same size.
Is anything uploaded?
No. The slicing and scoring all run in your browser — nothing is transmitted or stored. Turn off your Wi-Fi and it still works.
Limits
A teaching model using smooth precision/recall curves and a single test query — real retrieval quality depends on your embeddings, corpus, and query distribution. Use it to build the triangle intuition, then measure on your own eval set.
Related
Part of the miatz playgrounds. Explore the rest from the playground catalog.
Bookmark this page (Ctrl+D, or ⌘D on Mac) — it works offline the next time you need it.