Compare chunking strategies side-by-side

Drop a file or paste into the editor to compare how two algorithms split the same source. This tab extracts PDF and Word to text first, and nothing is uploaded. Those same strategy names are one pip install away in Python.

Drop a file

.txt, .md, .pdf, .docx, .csv, .json, .html, or code. PDF and Word extract here. Nothing is uploaded.

Choose a file

No file yet

Which strategy to pick

recursive_character is the RAG default: paragraphs, then lines, then spaces. It is not hierarchical recursive.
recursive keeps parent and child chunks. The list is a tree. pip compare is still how you run three or more.
token_based uses cl100k_base windows. The tokenizer loads once in this tab.
markdown_chunker follows headers. python_code follows module-level def/class (tree-sitter in this tab, stdlib ast in pip). Naive B is fixed_size.
overlapping_window slides with overlap. The amber band is the shared range.
regex_custom is for speaker turns and log lines. The match starts the next chunk.
Drop .md, .py, .json, .csv, .pdf, or .docx. PDF and Word are extracted to text first. You can always switch.
semantic loads MiniLM from Hugging Face (~25 MB) the first time you pick it. Cuts will not match pip [ml]. Comparing three or more strategies is pip compare.

Strategy A

Strategy B

Python library

This tab is two columns. N-way compare (three or more) and a strategy you wrote yourself run in pip, not in this JS UI.

N-way compare · Bring your own · Logs · All docs

Python

Two strategies here. Three or more: chunking-strategy compare FILE -s a,b,c. Your own chunker: @register_chunker then the same compare / benchmark. Grey names stay pip-only until a real JS port exists. Cap 150 MB. Python streams from disk. Docs

Document Find

A

B

How it works

Drop a file or paste

Files stay in this tab. Nothing is uploaded. Type, paste, or drop a file, including PDF and Word. The chunks update from the text the browser extracted.

Compare strategies

Two algorithms, same source. Click a chunk to underline its range. Link copies A vs B and the knobs, never your file.

Copy the pip command

The snippet on the left is already Python, with the same strategy names and knobs. Grey names in the dropdowns ship in the library, not this tab.

Where chunking sits in RAG

Retrieval can only be as good as the pieces you embed. This lab is that first cut: before vectors, before the index, before the model.

DOCUMENT

CHUNKING this playground

EMBED

INDEX

RETRIEVE

LLM

Chunking is the first cut. Bad windows here become bad retrieval later.

Use it in your pipeline

The browser lab caps at 150 MB. It extracts PDF and Word to text here. The Python library streams from disk, with the same strategy names, 40+ algorithms, unlimited file size, plus layout, images, OCR, and directories. N-way compare and a custom strategy are pip; the docs cover both.

Source on GitHub

The library, issues, and the pip-only strategies this tab does not fake, including embeddings and PDF layout. Token counts here are cl100k_base after you run token_based once, never a chars/4 guess.

Install from PyPI

Install chunking-strategy from pip. Same names as this lab, then it streams from disk: 40+ algorithms, PDF layout, directories, and no 150 MB cap.

Docs

N-way compare, bring-your-own strategies, every registry name, CLI, and what this tab will not fake.