43 names in the registry

Same string in the dropdown and in pip

list-strategies --tier lab|python_only|all is the source of truth. Public copy says 40+ because a few names are aliases. The lab runs 22. Everything else is honest grey or pip-only.

chunking-strategy list-strategies --tier lab
chunking-strategy list-strategies --tier python_only
chunking-strategy list-strategies --tier all

Live in the lab lab

Name	What the cut actually is
`fixed_size`	Character windows. Naive baseline. Headings get sliced.
`sentence_based`	Sentence groups. Default splitter `simple_v1` since 0.5.0. Optional cl100k packing via `max_tokens`.
`paragraph_based`	Blank-line / paragraph groups.
`overlapping_window`	Sliding window. Amber band is the overlap.
`recursive_character`	Market “recursive”: paragraphs, then lines, then spaces. Not a tree.
`recursive`	Hierarchical parent/child (paragraph then sentence). The list is a tree.
`token_based`	Real cl100k_base windows. Tokenizer loads once. Never chars/4.
`fixed_length_word`	Word-count windows.
`regex_custom`	Match starts the next chunk (turns, logs).
`markdown_chunker`	ATX headers. Breadcrumb + optional contextualize prefix.
`python_code`	Module-level `def`/`class`. Tab: tree-sitter. pip: stdlib ast.
`javascript_code` `css_code` `go_code` `java_code` `c_cpp_code`	Language-aware code splits.
`json_chunker`	JSON value / array structure, not a raw window.
`csv_chunker`	Row groups; optional header copied onto each chunk.
`xml_html_chunker`	Element-aware markup.
`rolling_hash` `fastcdc`	Content-defined boundaries.
`semantic`	Grey until MiniLM loads in the tab. pip needs `[ml]`. Cuts will not match.

Grey in the tab lab_later

Shown so you know the name exists. Locked until a JS port with a golden exists.

universal_code, rabin_fingerprinting, buzhash, gear_cdc, tttd, ml_cdc, enhanced_pdf_chunker, doc_chunker, boundary_aware, context_enriched, embedding_based, discourse_aware, grid_based_image, patch_based_image, time_based_audio, silence_based_audio, time_based_video.

Pip-only python_only

pdf_chunker (layout, images, tables), universal_document (Tika), adaptive, scene_based_video. The lab extracts PDF/Word to text, then runs a text strategy. It will not grow a fake pdf_chunker.

In this library recursive is a parent/child tree (paragraph then sentence). recursive_character is the separator cascade (paragraphs, then lines, then spaces). They are different strategies. LangChain’s cascade is the second name here.