Type a tool name, for example "keyword density" or "transcript".
Split long text into chunks that fit a context window or a retrieval pipeline, by tokens, characters or words.
Your text stays in your browser. Nothing is uploaded or saved.
Working. Large texts can take a few seconds.
The exact tokenizer could not be loaded, so token counts are estimates (about 4 characters per token). Reload the page to try again.
Save all chunks
Token counts use o200k_base, the tokenizer OpenAI publishes for its GPT models. Other models count differently, so leave some room below their limit.
An LLM text chunker splits long text into smaller, numbered pieces that fit a model's context window or a retrieval index. Paste text or drop a file, choose a size in tokens, characters or words, set an overlap, and you get copy-ready chunks you can export as JSON, JSONL, CSV or a plain text file.
Text chunking is the process of cutting a long document into smaller segments so a language model or a search index can handle each one. Models accept only a limited number of tokens per request, and retrieval systems work better when each stored passage covers one idea.
A token is the unit a model reads. It is often a short word or part of a longer word, and English text averages roughly four characters per token. Because models count in tokens, not words, a 1,000-word article can land anywhere from about 1,300 to 1,500 tokens depending on vocabulary and formatting.
Chunking has two settings that matter most:
You can split a document in a few clicks. Here is the full flow, using the labels you see on the page.
Say you have a 2,400-token Markdown guide that you want to index for a question-answering bot. You choose Tokens, 512 from Common token sizes, an overlap of 50, paragraph-first cutting and the heading option turned on.
Each chunk after the first begins with the last 50 tokens of the one before it. That means every chunk adds at most about 460 new tokens. Divide 2,400 by roughly 460 and you get a little over five, so expect about six chunks. Because the tool prefers to cut at paragraph ends, chunks usually come in a bit under 512 rather than exactly at it.
The summary might show an average chunk of around 430 tokens and a largest chunk just under 512. That is a healthy result: nothing exceeds your limit and the pieces are fairly even. If a heading such as "Feeding the starter" sits above a paragraph, the heading stays attached to that paragraph instead of being stranded at the end of the previous chunk.
If you need the chunks in a pipeline, click JSONL to get one chunk per line, which is a common format for batch jobs and embedding scripts.
The results panel shows the chunk count, the average chunk and the largest chunk, followed by each chunk as a card. Use these numbers to check whether your settings did what you intended.
This is normal and means the splitter used the space available. Remember that a model also needs room for your instructions and its reply. If you plan to paste a chunk into a chat with a 4,000-token prompt budget, choose a size below that, such as 3,000, not 4,000.
A big gap usually means short paragraphs or headings forced early cuts. Try Sentences, then words for even pieces, or raise the size so more paragraphs fit together.
Very small pieces lose context and bloat an index. Increase the size, or reduce overlap. Overlap that is too high repeats a lot of text, which inflates your total token count.
Switch Try to cut at from the exact size to a paragraph or sentence option. The exact setting cuts at the limit, even in the middle of a word, so use it only when a strict length matters more than readability.
These are common guidelines, not rules. Test with your own content and questions.
| Goal | Chunk size | Overlap |
|---|---|---|
| Retrieval (RAG) passages | 256 to 512 tokens | 10 to 20% |
| Pasting long text into a chat | 2,000 to 4,000 tokens | 0 to 5% |
| Summarizing in large passes | 8,000 to 16,000 tokens | Small or none |
Anyone who works with text that is longer than a model can read in one go will benefit. These are the situations where it earns its place:
Good chunks come from matching the settings to the job, not from picking the biggest number. These habits prevent most problems.
If the tool shows that the chunk size was raised or lowered, your number fell outside the allowed range. The minimum is 20 for every unit. The maximum is 100,000 for tokens and words, and 1,000,000 for characters. If the overlap was reduced, it was too large for the chunk size, so lower it or increase the size. If a file is rejected, it is either larger than 10 MB or could not be read, so paste the text directly or split the file first.
The chunker reports token counts with o200k_base, the tokenizer OpenAI publishes for its GPT models. Other model families count tokens differently, so the same chunk may be a little larger or smaller elsewhere. Keep a safety margin of 5 to 10 percent below the limit of any non-OpenAI model.
If the exact tokenizer cannot load, the page shows a note and counts become estimates of about four characters per token. Reload the page to try again before relying on tight limits.
When a text produces many chunks, the page previews them as cards and may not list every one on screen. Use the export buttons under Save all chunks to get the complete set. Also note that the tool splits text only. It does not judge meaning, so very different topics inside one paragraph can still end up in the same chunk.
Once your content is chunked, you can check how crawlers see the original page with the AI crawler view, or run the AI search readiness checker. For more related utilities, browse the AI SEO tools hub.
Paste your text into the box above, pick a size and overlap that suit your model, and click Split text to get clean, numbered chunks you can copy or export in seconds.
Your report has been submitted. We will look into this issue.
A VPN changes your network exit. It does not guarantee anonymity.
Themes, plugins, code, video, audio, graphics & more from 90,000+ authors.
Web Development
Design & Media
Unlimited Access
Explore Envato Market