RAG Chunking Visualizer
Paste any text and see how it splits into chunks for a retrieval pipeline. Adjust chunk size, overlap, and strategy, and watch where each chunk starts, ends, and repeats characters from its neighbour. Everything runs in your browser; nothing is sent anywhere.
501 characters (about 126 estimated tokens).
Chunks (2)
- Chunk 0280 chars~70 tokens (estimate)chars 0 to 280
RAG splits a long document into smaller chunks. The name is short for retrieval-augmented generation. An embedding model can then index each chunk on its own. Chunk size and overlap decide how much context each chunk carries. Too small, and one idea gets split across chunks. Too
- Chunk 1261 chars~66 tokens (estimate)chars 240 to 501first 40 chars overlap chunk 0
d one idea gets split across chunks. Too large, and retrieval returns more than the model needs. Overlap repeats a few characters from the end of one chunk at the start of the next. A sentence that sits on a boundary still shows up whole in at least one chunk.
Token counts are estimated at four characters per token and will differ from a model's own tokenizer. The highlighted range shows the characters a chunk repeats from the one before it.
Common questions
- What is chunk overlap and why does it matter?
- Overlap repeats a set number of characters from the end of one chunk at the start of the next. A sentence that falls on a chunk boundary can get cut in half; overlap makes sure that sentence still appears whole inside at least one chunk, which helps retrieval find it.
- How should I pick a chunk size?
- Chunk size is a trade-off. Smaller chunks keep each one focused but can split a single idea across several chunks; larger chunks hold more context but return more text than the model needs. A common starting point is a few hundred characters with a small overlap, then adjust based on how retrieval performs on your own documents.
- Are the token counts exact?
- No. The token counts here are estimated at roughly four characters per token for English text. A model's own tokenizer will give a different number, especially for code, other languages, or unusual formatting. Use the estimate for planning, not for billing.