RAG Chunk Comparator

What it does

How you split a document changes what your retriever finds. The RAG Chunk Comparator chunks one document three ways — fixed-size (the naive default), sentence-aware (boundaries never cut a sentence), and markdown-aware (sections stay together, headings carried as context) — then compares chunk counts, size spread, and the share of boundaries that land on sentence ends. Mid-sentence cuts are the classic recall killer: half a thought embeds badly. Runs 100% client-side.

How to use it

1. Paste the document your retriever will ingest.
2. Set the target chunk size in tokens (300–500 is a common starting point).
3. Tap the three stat tiles to switch strategies and read each one’s chunks — the sentence- boundary percentage tells you which strategies cut thoughts in half.

Examples

A 200-token policy document at a 60-token target:

fixed      4 chunks · 41–60 tok · 33% sentence boundaries
sentence   4 chunks · 38–58 tok · 100% sentence boundaries
markdown   3 chunks · 52–60 tok · 100% sentence boundaries (headings kept)

Fixed wins on uniformity; markdown wins on coherence — every chunk is a complete section with its heading attached.

Good to know

• Sentence and markdown strategies trade exact size for coherence: a single sentence larger than the target becomes its own oversize chunk rather than being split.
• Markdown chunks carry the nearest heading, which doubles as cheap metadata for citation and filtering downstream.
• Token counts are prose-heuristic estimates (~15% tolerance).
• Runs 100% client-side. Your document never leaves the browser.
• Related: Embedding Chunk Planner, Token Estimator, Context Window Planner.