Skip to content
StackUnseen
Subscribe

Kit · Lab

Chunking lab

Where a chunk ends decides what the model can know. Move the sliders and watch the answer get cut in half, then put back together.
Try

The default: the hotel sentence lands across a boundary.

No chunk holds the whole answer: the split cut it in two. Try a bigger chunk or some overlap.2 of 5 chunks sent · 80 words

  1. Chunk 1 · score 0.20

    Travel and expenses policy, version 4. Who this covers. This policy applies to all employees and contractors travelling on company business. Managers approve trips before booking; finance approves anything over the limits below. Booking. Book travel through the company travel

  2. Chunk 2 · score 0.60 · retrieved #2

    desk at least fourteen days ahead where possible. Economy class is the default for flights under six hours. Premium economy is allowed for flights over six hours with manager approval. Hotels. Hotels are reimbursed up to 180 pounds per night

  3. Chunk 3 · score 1.00 · retrieved #1

    in London and 140 pounds per night elsewhere in the United Kingdom. Outside the United Kingdom the limit is the local rate published by finance each January. Breakfast is included in the nightly limit when the hotel charges it separately.

  4. Chunk 4 · score 0.20

    Meals. Meals are reimbursed up to 45 pounds per day when travelling overnight, and up to 20 pounds for a day trip. Alcohol is not reimbursed. Claims. Submit claims within thirty days of returning, with itemised receipts. Claims older than

  5. Chunk 5 · score 0.00

    ninety days are not paid.

Edit the document

What it shows

A teaching model, not a benchmark: chunks are fixed word windows and ranking is shared word roots, the way a keyword retriever scores. Real systems split on structure, rank with embeddings and a reranker, and still fail the same way at the boundaries.

Read on: Why chunk boundaries shape answer quality and Chunking and context packaging, or open the Retrieval component on the map.