Text Chunker for LLMs
Split long documents into overlapping, token-sized chunks for search and retrieval, and test which chunk answers a question.
Token counts are estimates (about 4 characters per token). Splits happen at paragraph and sentence boundaries.
Test a Question
Type a question to see which chunks match it best. This is keyword matching (BM25), not semantic search, so it can miss paraphrases and synonyms. It is a quick way to check whether your chunk size and overlap keep the right passage together.
What It Is For
Models can only read so much at once, and document-search systems (often called RAG) work by splitting documents into chunks, finding the few chunks relevant to a question, and sending only those. This tool does the splitting step, with a token budget per chunk and an optional overlap, and lets you test whether a sample question finds the right chunk.
Example
A 100,000-token manual and 20 questions: sending the whole manual each time costs 2,000,000 input tokens, while sending the three most relevant 500-token chunks per question costs 30,000. The chunking itself costs nothing; the saving comes from sending less.
Why Overlap?
An idea often straddles a boundary. With a 50-token overlap, the end of each chunk is repeated at the start of the next, so a sentence cut at the boundary still appears whole in at least one chunk.
Questions
Is my text uploaded?
No. Splitting and matching happen in your browser.
How big should chunks be?
A common starting point is 300 to 800 tokens with 10 to 15 percent overlap. Smaller chunks give more precise matches, larger chunks keep more context.
Are the token counts exact?
No, they are estimates of about 4 characters per token. Use the Token Counter for exact OpenAI counts.
How does the question test work?
It scores each chunk by how many of your question words it contains, giving more weight to rarer words (the BM25 method). Real systems usually use embeddings, which also match different wording with the same meaning, so treat this as a rough check.