AI DEVELOPMENT WORKSPACE
RAG Chunking Preview
Inspect actual chunk slices, offsets and overlap before building retrieval.
Input is processed on this site only when you run the tool. The application does not save inputs or results.
input.txt
Result result.json
A little clarity goes here.
Run the tool to see your result.
Result preview
UTF-8 text, at most 32768 bytes per input. Tab leaves the editor. Plain text disables syntax colors. Browser textareas may normalize line endings to LF.
{"preview":{"inputMode":"text","options":{"chunkUnit":{"type":"enum","values":["characters","words"],"default":"characters"},"size":{"type":"integer","min":1,"max":4096,"default":1000},"overlap":{"type":"integer","min":0,"max":4095,"default":100}},"inputLabels":[]}}{"preview":{"input":"Build useful tools.\nKeep the boundaries clear.\nTest with fictional data.","options":{"chunkUnit":"words","size":6,"overlap":2}}}A QUICK WALKTHROUGH
A familiar workflow,
with room to focus.
Paste text and choose the chunk units, size and overlap.
Preview the chunks and their ranges.
Expand chunks; Source data retains every complete chunk for copying or downloading.
When to use RAG Chunking Preview
Review where text windows start and end before experimenting with a retrieval pipeline outside this site.
Read the result
The view shows the shared chunker’s exact slices and overlap. It does not create embeddings, run retrieval or measure semantic relevance; use Source data to inspect the complete boundaries independently.
SEE IT IN PRACTICE
Worked examples
Sample input. A result you can inspect.
Preview example
Input
Build useful tools. Keep the boundaries clear. Test with fictional data.
- Chunk units
- Whitespace-delimited words
- Chunk size
- 6
- Overlap units
- 2
Expected result
{"unit":"words","size":6,"overlap":2,"totalUnits":11,"chunks":[{"index":1,"unitStart":0,"unitEnd":6,"byteStart":0,"byteEnd":40,"codePointStart":0,"codePointEnd":40,"overlapUnits":0,"text":"Build useful tools.\nKeep the boundaries "},{"index":2,"unitStart":4,"unitEnd":10,"byteStart":25,"byteEnd":67,"codePointStart":25,"codePointEnd":67,"overlapUnits":2,"text":"the boundaries clear.\nTest with fictional "},{"index":3,"unitStart":8,"unitEnd":11,"byteStart":52,"byteEnd":72,"codePointStart":52,"codePointEnd":72,"overlapUnits":2,"text":"with fictional data."}]}{"kind":"chunks","title":"3 chunks prepared","unit":"words","size":6,"overlap":2,"total":3,"chunks":[{"index":1,"unitStart":0,"unitEnd":6,"byteStart":0,"byteEnd":40,"codePointStart":0,"codePointEnd":40,"overlapUnits":0,"text":"Build useful tools.\nKeep the boundaries "},{"index":2,"unitStart":4,"unitEnd":10,"byteStart":25,"byteEnd":67,"codePointStart":25,"codePointEnd":67,"overlapUnits":2,"text":"the boundaries clear.\nTest with fictional "},{"index":3,"unitStart":8,"unitEnd":11,"byteStart":52,"byteEnd":72,"codePointStart":52,"codePointEnd":72,"overlapUnits":2,"text":"with fictional data."}],"note":"Ranges start at zero and exclude the end position. Full exact chunks remain available in source data.","schema":1}Formats, limits and common errors
Visible overlap is a text-window setting, not a promise that a retrieval system will perform well.
- At most 32,768 UTF-8 bytes, 200 chunks and 262,144 aggregate chunk-text bytes. Excess is rejected, not truncated.
- Size 1–4,096, overlap 0–4,095 and overlap strictly smaller than size. Character means Unicode code point, not a displayed grapheme or token.
- Words are Unicode nonwhitespace runs. Leading whitespace belongs to the first unit and following whitespace to the preceding unit. Whitespace-only text is one unit; empty text gives no chunks. Text is retained exactly.
- Offsets are zero-based, half-open byte, code-point and selected-unit ranges in the submitted text. Browser textareas may normalize line endings. No sentence, semantic, token-budget or retrieval-quality guarantee.
- The preview uses the same PHP chunker as Text Chunker. It does not generate embeddings, query an index, retrieve results or assess RAG quality.
- Input: 32,768 UTF-8 bytes per text field. Output: at most 524,288 UTF-8 bytes; serialized responses have a separate 1 MiB ceiling. Size/work failures return no partial result.