RAG Chunking Visualizer
Inspect local document chunk boundaries, overlap and duplicated token estimates before building a retrieval pipeline.
Split up to 64 KiB UTF-8 locally. Token mode uses the existing UTF-8 weighted estimate, not an exact provider tokenizer. No embeddings or API calls.
Enter a document and choose chunk boundaries.
No verified local GPT-6.1 encoding is bundled. This is a text heuristic, not the provider tokenizer. At most 256 chunks and 256 KiB total chunk text including overlap. Reduce overlap if those limits are reached.
Boundaries affect retrieval.
Small chunks can isolate facts but lose surrounding context. Larger chunks preserve context but may dilute retrieval precision. Overlap repeats material between adjacent chunks. There is no universally ideal size; evaluate against your documents and queries.
Fixed characters or estimated tokens.
Character mode counts Unicode code points, preserving emoji surrogate pairs. Token mode uses the same UTF-8 weighted heuristic as the Token Counter: ASCII contributes roughly one quarter token per character, and other text roughly one token per two UTF-8 bytes. This is not a provider tokenizer, exact vocabulary or semantic analysis. Combining characters can cross boundaries; no normalization or paragraph-aware splitting is applied.
See the overlap cost.
Repeated text has both an underline and an explicit label. Ranges use zero-based, exclusive-end character offsets. Indexed tokens sum independently rounded chunk estimates; repeated-overlap tokens estimate the repeated spans separately. Their totals can differ slightly from subtracting source tokens because of rounding. Token overlap is a maximum target adjusted to whole code points.
Local and bounded.
Use up to 64 KiB UTF-8, 256 chunks and 256 KiB total chunk text including duplication. Size is 1–8,192 characters or 2–2,048 estimated tokens; overlap must be smaller. Ten chunks appear per page. Nothing runs on each keystroke. No embeddings, model inference, uploads, persistence or document URLs are used.