Late Chunker

Split text into chunks based on a late-bound token count

The LateChunker implements the late chunking strategy described in the Late Chunking paper. It builds on top of the RecursiveChunker and uses document-level embeddings to create more semantically rich chunk representations.

Instead of generating embeddings for each chunk independently, the LateChunker first encodes the entire text into a single embedding. It then splits the text using recursive rules and derives each chunk’s embedding by averaging relevant parts of the full document embedding. This allows each chunk to carry broader contextual information, improving retrieval performance in RAG systems.

API Reference

To use the LateChunker via the API, check out the API reference documentation.

Installation

LateChunker requires the sentence-transformers library to be installed, and currently only supports SentenceTransformer models. You can install it with:

The LateChunker uses RecursiveRules to determine how to chunk the text. The rules are a list of RecursiveLevel objects, which define the delimiters and whitespace rules for each level of the recursive tree. Find more information about the rules in the Additional Information section.

pip install "chonkie[st]"

For installation instructions, see the Installation Guide.

Initialization

from chonkie import LateChunker

chunker = LateChunker(
    embedding_model="nomic-ai/modernbert-embed-base",
    chunk_size=2048,
    rules=RecursiveRules(),
    min_characters_per_chunk=24,
)

You can also initialize the LateChunker using a recipe. Recipes are pre-defined rules for common chunking tasks. Find all available recipes on our Hugging Face Hub here.

from chonkie import LateChunker

# Initialize the late chunker to chunk Markdown
chunker = LateChunker.from_recipe("markdown", lang="en")

# Initialize the late chunker to chunk Hindi texts
chunker = LateChunker.from_recipe(lang="hi")

Parameters

embedding_model · str · default: "nomic-ai/modernbert-embed-base"

SentenceTransformer model to use for embedding

chunk_size · int · default: "2048"

Maximum number of tokens per chunk

rules · RecursiveRules · default: "RecursiveRules()"

Rules to use for chunking

min_characters_per_chunk · int · default: "24"

Minimum number of characters per sentence

Usage

Single Text Chunking

text = """First paragraph about a specific topic.
Second paragraph continuing the same topic.
Third paragraph switching to a different topic.
Fourth paragraph expanding on the new topic."""

chunks = chunker(text)

for chunk in chunks:
    print(f"Chunk text: {chunk.text}")
    print(f"Token count: {chunk.token_count}")
    print(f"Embedding shape: {chunk.embedding.shape}")

Batch Chunking

texts = [
    "First document about topic A...",
    "Second document about topic B..."
]

batch_chunks = chunker(texts)

for chunk in batch_chunks:
    print(f"Chunk text: {chunk.text}")
    print(f"Token count: {chunk.token_count}")
    print(f"Embedding shape: {chunk.embedding.shape}")

Return Type

LateChunker returns chunks as Chunk objects:

@dataclass
class Chunk:
    text: str           # The chunk text
    start_index: int    # Starting position in original text
    end_index: int      # Ending position in original text
    token_count: int    # Number of tokens in chunk
    context: Optional[Context] = None    # Optional context metadata
    embedding: Union[list[float], "np.ndarray", None] = None  # Optional embedding vector

As of version 1.3.0, LateChunker returns the base Chunk type instead of the specialized LateChunk type. The embedding is automatically populated by the LateChunker during the chunking process.

Additional Information

LateChunker uses the RecursiveRules class to determine the chunking rules. The rules are a list of RecursiveLevel objects, which define the delimiters and whitespace rules for each level of the recursive tree.

@dataclass
class RecursiveRules:
    rules: list[RecursiveLevel]

@dataclass
class RecursiveLevel:
    delimiters: Union[None, str, list[str]]
    whitespace: bool = False
    include_delim: Optional[Literal["prev", "next"]]  # Whether to include the delimiter in the previous chunk or the next chunk.

You can pass in custom rules to the LateChunker, or use the default ones. Default rules are designed to be a good starting point for most documents, but you can customize them to your needs.

RecursiveLevel expects the list of custom delimiters to not include whitespace. If whitespace as a delimiter is required, you can set the whitespace parameter in the RecursiveLevel class to True. Note that if whitespace = True, you cannot pass a list of custom delimiters.