LLM Text Chunker

How to use

1. 1Paste your document into the input area, or click Load sample to try a sample RAG-style text.
2. 2Pick a strategy. Recursive separators is the safest default for documents with paragraphs and headings; Characters is best for raw text; Tokens is best when you have a token budget in mind.
3. 3Set the chunk size and overlap. A common starting point is 1000 characters with 100 overlap, or use one of the presets at the top.
4. 4Scroll the chunk preview to read every chunk with source offsets, character and word counts, and approximate token counts. Use the per-chunk Copy button to grab one chunk.
5. 5Pick JSON, JSONL, or Plain text for the export format, then click Copy export or Download to save the full set of chunks to your machine.

About this tool

LLM Text Chunker splits a long document into smaller, overlapping pieces that fit into an embedding model, a retrieval-augmented generation pipeline, or a constrained prompt context. Paste the text on the left, pick a strategy, and see every chunk numbered and ready to copy. Five strategies cover the common cases: Characters slices a fixed character window with a configurable overlap, useful when the input has no clear structure; Tokens uses the same window sized in approximate tokens (a 4 chars per token heuristic, matching the rule of thumb published by OpenAI); Words groups whitespace-delimited words into a fixed count with overlap, useful for prose; Paragraphs packs whole paragraphs into a chunk under the size budget so headings and lists stay intact; and Recursive separators implements the LangChain RecursiveCharacterTextSplitter algorithm, trying double newline first, then single newline, sentence-ending punctuation, semicolons, commas, single spaces, and finally character-level fallback, which is the default used by most RAG pipelines. The tool reports the number of chunks, average and maximum chunk size in characters and approximate tokens, the smallest and largest chunk, and an Over budget count of any chunk that exceeded the size limit because no smaller separator was available. Each chunk shows its source offset, character and word counts, and a one-click Copy button; the full set can be exported as JSON for direct ingestion into a vector store, JSONL for streaming embedding APIs, or plain text with chunk separators for human review. Quick presets cover the four most common configurations: Embeddings (1000 characters, 100 overlap, recursive) for general RAG, Long context (4000 characters, 400 overlap) for long-context models, Tweet-size (280 characters), and SMS chunks (160 characters). The token counts here are estimates; pair this tool with the GPT Token Counter when you need an exact BPE count against a specific model. Everything runs locally in your browser so the documents you split (knowledge base content, internal docs, private corpora) never leave your device or pass through a third-party API.

Free to use. Works in your browser. No signup, no login.