中文

RAG Chunking Playground

Preview RAG chunks with six strategies, real token counts, source offsets, and exportable results.

Local RAG laboratory

Chunking workspace

No uploads · Worker processing
584 UTF-16 code unitsInput is not persisted
or drop a local file here (up to 5 MB)
Strategy & parameters
Supports \n, \r, \t, and \\ escapes; never treated as regular expressions.

o200k_base · RAG estimate Real tiktoken counts; estimates unless your target model also uses o200k_base.

Offsets are UTF-16 [start, end) positions in the original input; character counts are UTF-16 too. Fixed-character splitting measures grapheme clusters.

Inspect results

Chunk list

0Chunks
0Average chars
0Minimum chars
0Maximum chars
No chunks to display. Adjust the input or settings to begin.
Text is processed only in this tab and cleared on refresh.
About this tool

Made for the task at hand.

Tune retrieval-ready passages before sending them to an embedding pipeline. Compare six deterministic splitting strategies, inspect true o200k_base token counts, and locate every chunk in the original input without uploading your document.

How to use

  1. Paste a document or import a local TXT or Markdown file, then choose a chunking strategy.
  2. Adjust chunk size, overlap, minimum length, separators, whitespace normalization, and short-chunk merging.
  3. Locate or copy individual chunks, review length statistics, and export the complete result as JSON or JSONL.
Your work stays yours.
Tool inputs are processed locally in your browser and are never sent to KitVerse servers. Clear the workspace when using a shared device.
Common questions
Are the token counts exact for my RAG model?

The tool uses the real o200k_base tiktoken encoder. Counts are exact for models using that encoding and clearly labeled as estimates for other RAG or embedding models.

What do start and end offsets mean after whitespace normalization?

They remain zero-based UTF-16 [start, end) ranges in the original input. Normalized chunk text can differ from the selected source range because repeated whitespace was collapsed.

Can headings inside code fences create chunks?

No. The Markdown strategy tracks backtick and tilde fences so heading-like lines inside examples remain part of their surrounding section.

Related tools · AI & LLM

New

Tokenizer Playground

See exactly how text becomes tokens, then compare the cost across popular LLMs.

Free

Token Counter

Count characters and LLM tokens with a lightweight, private text workspace.

Free

Token Cost Calculator

Estimate input and output spend with transparent per-million-token pricing.