--- language: en title: "Late chunking" description: "Late chunking creates embeddings for chunks after processing a larger context. The technique aims to preserve references between parts of the same document." canonical: "https://tropk.ai/en/resources/glossary/late-chunking" last_modified: "2026-10-12T16:13:12.018Z" --- # Late chunking Reviewed on 2026-10-06. ## What it means Late [chunking](/en/resources/glossary/chunking) is a technique for generating [embeddings](/en/resources/glossary/embeddings) for chunks using the context of a larger portion of the document. The model first processes the tokens together; their representations are then grouped according to each chunk’s boundaries. In the original method, the average of a passage’s token vectors produces its embedding, preserving dependencies on other parts. ## Why it matters for marketing The concept is relevant when business documents use references such as this plan or the option above, whose meaning depends on an earlier section. Generating a vector in isolation can lose this association. For marketing, the decision is whether to assess if the additional context helps find the correct condition, taking into account the cost and technical feasibility of processing. ## Example in practice Imagine a software company whose manual introduces the Team plan and later states that this option allows users to export reports. The isolated passage does not name the plan. When reviewing search results about exporting with the Team plan, the team identifies this dependency and decides to compare isolated embeddings with late chunking, keeping the same chunking and questions. ## How to apply or verify The person responsible for [retrieval](/en/resources/glossary/information-retrieval) should confirm whether the model and implementation allow token representations to be obtained and grouped by chunk. Record the version, [context window](/en/resources/glossary/context-window), passage boundaries, and grouping method. Compare queries with distant references and with independent information. Also check whether long documents were truncated, since discarded context does not contribute to the representation. ## Important distinctions Late chunking does not involve creating a single vector for the document and then splitting it. Each chunk receives its own vector, derived from the contextualized token representations. It also differs from adding a textual summary to each passage. The original article identifies cases where additional context does not help; the technique does not guarantee an advantage for every task or eliminate the need for evaluation. ## Sources and references - [Late Chunking: Contextual Chunk Embeddings Using Long-Context Embedding Models]() — Jina AI e Weaviate, via arXiv ## Continue a leitura - [Semantic chunking]() - [Embeddings]() - [Passage retrieval]() ## About the author [Vinicius Pacheco]() Responsible for Tropk’s technical content, focusing on crawling, indexing, structured data, and agent access to content. --- [All pages in Markdown]()