Search

Search IconIcon to open search

Late Chunking

Last updatedUpdated: by Jakub Žovák · 1 min read

Properties
created 07.02.2025, 16:43
modified 02.08.2026, 10:19
published Empty
topics Chunking
authors Jakub
ai-assisted No

Jina - Late Chunking:

Late Chunking approach we propose in this article first applies the transformer layer of the embedding model to the entire text or as much of it as possible. This generates a sequence of vector representations for each token that encompasses textual information from the entire text. Subsequently, mean pooling is applied to each chunk of this sequence of token vectors, yielding embeddings for each chunk that consider the entire text’s context. Unlike the naive encoding approach, which generates independent and identically distributed (i.i.d.) chunk embeddings, late chunking creates a set of chunk embeddings where each one is “conditioned on” the previous ones, thereby encoding more contextual information for each chunk.

# Comparison

Naive ChunkingLate Chunking
The need of boundary cuesYesYes
The use of boundary cuesDirectly in preprocessingAfter getting the token-level embeddings from the transformer layer
The resulting chunk embeddingsi.i.d.Conditional
Contextual information of nearby chunksLost. Some heuristics (like overlap sampling) to alleviate thisWell-preserved by long-context embedding models