What chunk size should I start with?

For question answering, 256 to 512 tokens with 10 to 20 percent overlap is a good start. Larger chunks keep more context but match questions less sharply. Try two sizes on a few real questions and keep the one that finds the right passage.

Does it change my text?

No. Every chunk is an exact slice of your document with its start and end position, so you can always find it again in the source.

Are my files uploaded?

No. This tool runs entirely in your browser tab. Files are read from your disk into memory and the result is generated locally.

Is it free? Are there limits or watermarks?

Free and open source (MIT). There are no file-size caps, daily quotas, accounts or watermarks.

Does it work offline?

Yes, once the page (and any engine it needs) has loaded, it keeps working without a connection thanks to the service worker cache.

Home / AI Builder / RAG Text Chunker

RAG Text Chunker

Split documents into overlapping chunks by characters, words or tokens for embeddings and RAG, then export JSONL, JSON or CSV.

Free, no uploads, no sign-up. Runs 100% in your browser.

How RAG Text Chunker works

Everything runs in your browser. Prompts, datasets and model files are never uploaded. A recursive splitter tries Markdown headings, then paragraphs, lines, sentences and words, so each chunk breaks at the largest natural boundary that fits your size. Overlap repeats the end of one chunk at the start of the next. Every chunk is an exact slice of your text with its start and end position.

Input: Text, TXT, MD, HTML. Output: JSONL, JSON, CSV.

How to use RAG Text Chunker

Paste a document or drop a file.
Pick how to split and whether size is in characters, words or tokens.
Set the chunk size and overlap, then press Split.
Check the chunks and download JSONL, JSON or CSV.

Tips

256 to 512 tokens with 10 to 20 percent overlap is a common start for question answering.
Markdown mode keeps each heading at the start of its chunk, which helps retrieval.

RAG Text Chunker FAQ

What chunk size should I start with?: For question answering, 256 to 512 tokens with 10 to 20 percent overlap is a good start. Larger chunks keep more context but match questions less sharply. Try two sizes on a few real questions and keep the one that finds the right passage.
Does it change my text?: No. Every chunk is an exact slice of your document with its start and end position, so you can always find it again in the source.
Are my files uploaded?: No. This tool runs entirely in your browser tab. Files are read from your disk into memory and the result is generated locally.
Is it free? Are there limits or watermarks?: Free and open source (MIT). There are no file-size caps, daily quotas, accounts or watermarks.
Does it work offline?: Yes, once the page (and any engine it needs) has loaded, it keeps working without a connection thanks to the service worker cache.

Related tools

Token Counter: Count tokens with the real GPT-4o, GPT-4, Llama 3, Mistral, Qwen, Gemma or DeepSeek tokenizer, see each token, context use and cost.
Embedding Playground: See how an AI embedding model measures meaning: a similarity grid, the closest pairs, a 2D meaning map and search, all on your device.
MCP Config Builder: Build Model Context Protocol server configs for Claude Desktop, Cursor, VS Code, Windsurf, Claude Code and Gemini CLI, or convert one app's config to another.
Fine-tune Dataset Checker: Validate a JSONL fine-tuning file line by line: roles, empty replies, duplicates, token limits, then download clean train and validation splits.