Late Chunking: Contextual Chunk Embeddings Using Long-Context Embedding Models

By Michael Gunther, Isabelle Mohr, Daniel James Williams, Bo Wang, Han Xiao. 2024

Subfields: Retrieval-Augmented Generation, Information Retrieval & Search, Representation & Self-Supervised Learning

Citations in the corpus

• IdioLink: Retrieving Meaning Beyond Words Across Idiomatic and Literal Expressions (2026, unknown): “…complementary metrics (more in Appendix D): (i) nDCG@10 (binary relevance), which emphasizes ranking quality at top positions (Lee et al., 2024; Günther et al., 2025); (ii) R– Precision, which measures precision at rank R, the number of relevant documents per query (Wu et al., 2024; Hasan and…”
• Improving Long-Context Retrieval with Multi-Prefix Embedding (2026, unknown): “Late chunking (Günther et al., 2024) pools chunk representations after bidirectional document encoding, while token pooling (Clavié et al., 2024) compresses ColBERT-style token embeddings for short-passage retrieval.”
• MDKeyChunker: What Does One LLM Call per Chunk Buy for Markdown Retrieval? (2026, unknown): “Late chunking [Günther et al., 2024] obtains document-aware chunk embeddings without any LLM call, by embedding the whole document with a long-context model and pooling token embeddings per chunk.”
• Beyond Chunk-Then-Embed: A Comprehensive Taxonomy and Evaluation of Document Chunking Strategies for Information Retrieval (2026, unknown): “Original Claim: Contextualized chunking (embedding the document before segmentation) preserves global context, thereby improving retrieval effectiveness on standard in-corpus benchmarks compared to embedding chunks independently [G¨unther et al.(2024)].”
• Context is Gold to find the Gold Passage: Evaluating and Training Contextual Document Embeddings (2025, unknown): “Specifically, we include Anthropic’s contextual retrieval approach (Anthropic, 2024) 4 , and evaluate Late Chunking (Günther et al., 2024) without specific fine-tuning using modernbert-embed-large .”
• CoRet: Improved Retriever for Code Editing (2025, unknown): “Context has been shown to significantly improve the quality of retrieval (Lewis et al., 2020b; Günther et al., 2024; Borgeaud et al., 2022).”
• Late Code Chunking: A Code Chunking Strategy for Repository-Level Code Completion (2026, unknown): “A primary cause of failure is semantic loss introduced during chunking —the process of splitting source code into separately retrievable fragments (Günther et al., 2025).”