Late Chunker

Late chunking with contextual embeddings for each token

Late Chunker

cURL

curl --request POST \
  --url https://api.chonkie.ai/v1/chunk/late

200

{
  "text": "<string>",
  "start_index": 123,
  "end_index": 123,
  "token_count": 123
}

The Late Chunker generates contextual embeddings for each token before chunking, enabling more semantically aware chunk boundaries.

Examples

Text Input

Python

from chonkie.cloud import LateChunker

chunker = LateChunker(
    embedding_model="minishlab/potion-base-8M",
    chunk_size=512
)

text = "Your text here..."
chunks = chunker.chunk(text)

File Input

Python

from chonkie.cloud import LateChunker

chunker = LateChunker(
    embedding_model="minishlab/potion-base-8M",
    chunk_size=512
)

# Chunk from file
with open("document.txt", "rb") as f:
    chunks = chunker.chunk(file=f)

Request

Parameters

text · string | string[]

The text to chunk. Can be a single string or an array of strings for batch processing. Either text or file is required.

file · file

File to chunk. Use multipart/form-data encoding. Either text or file is required.

embedding_model · string · default: "minishlab/potion-base-8M"

The embedding model to use.

Chosen embedding model must return token level embeddings. Request will fail otherwise

chunk_size · integer · default: "512"

Maximum number of tokens per chunk.

recipe · string · default: "default"

Recursive rules recipe to follow. See all recipes on our Hugging Face repo

lang · string · default: "en"

Language of the document to chunk

Response

Returns

Array of Chunk objects, each containing:

text · string

The chunk text content.

start_index · integer

Starting character position in the original text.

end_index · integer

Ending character position in the original text.

token_count · integer

Number of tokens in the chunk.