Playground

The lab is two columns on one file

Open the playground. Drop a file or paste. Two strategies run in a worker. Click a chunk and the source underlines that Unicode range. Nothing is uploaded. The snippet on the left is already the pip you would run next.

What you actually do

1. Drop .txt, .md, .py, .json, .csv, .html, code, PDF, or Word. PDF.js and mammoth extract text in the tab. There is no lab strategy named pdf_chunker.
2. The lab picks A and B from the extension. You can switch either side. Grey names are pip-only until a real JS port exists.
3. Knobs match the Python kwargs. The snippet updates as you type.
4. Click a chunk. Focus chips read A 00 / B 01, not hex colors. Find ranks chunks in both lists; that query also stays local.
5. Link copies A vs B and the knobs. It never puts your file in the URL.

Layout

Desktop is a three-column grid: toolbox, an 8px divider, then source on top of A | B. Drag the vertical divider to widen the source, or Tab to it and use ArrowLeft / ArrowRight (16px steps). Drag the horizontal divider under the editor to make the paste box taller, or Tab to it and use ArrowUp / ArrowDown. Width is lab-rail-w, height is lab-paper-h. Sidebar floor 196px, ceiling min(420px, 42% of the app). Source floor 160px. Below 960px the sidebar divider hides and the toolbox stacks under the stage; you can still drag the source taller.

Names that surprise people

• recursive_character is the RAG default: paragraphs, then lines, then spaces. It is not hierarchical recursive.
• recursive is a parent/child tree (paragraph then sentence). The list is a tree. pip compare is still how you run three or more.
• python_code in the tab is tree-sitter (indent fallback). pip uses stdlib ast. Nested methods are separate chunks. Compare on symbol_name / symbol_kind / line_start / line_end.
• token_based loads cl100k_base once. Counts are not chars/4.
• semantic is grey until Transformers.js MiniLM loads from Hugging Face (~25 MB). Cuts will not match pip [ml].
• overlapping_window paints the shared range in amber.
• regex_custom starts the next chunk at the match (speaker turns, log lines).

Hard limits on purpose

• 150 MB. Python streams from disk; the tab will not pretend to.
• Two columns. N-way is pip compare.
• No eval of custom JS. A strategy you wrote is pip register_chunker.
• No analytics, no sendBeacon, no files in the share URL.
• Images, xlsx, pptx, OLE .doc, layout PDFs, audio, and video stay pip-only. See Formats.