What does chunking mean? Origins, examples, data chunking, semantic and late chunking

Chunking means breaking information into smaller, meaningful units. In psychology and learning, it describes how people group items so memory can hold more; in computing and AI, it means splitting data or documents into pieces for processing, such as the chunks retrieved in RAG. I design retrieval systems like RAG.NextUpgrad and teach study habits in campus sessions, so this page covers the meaning, its origin, examples for students and assignments, data chunking, and tutorials for RAG, semantic, late and Docling chunking.

What does chunking mean?

In psychology, chunking is grouping individual pieces of information into larger, meaningful units, so working memory can hold more. In computing and AI, chunking means splitting large data, text or files into smaller pieces for storage, processing or retrieval.

The two meanings share an idea: big things are easier to handle in well-chosen pieces. A phone number is remembered as a few groups rather than ten digits; a long PDF is split into passages so a search system can find the relevant one. In both cases, good chunks are meaningful units, not arbitrary cuts: a chunk that splits a sentence or idea in half is harder to remember and harder to retrieve.

Who gave the concept of chunking?

Psychologist George A. Miller introduced chunking in his famous 1956 paper, The Magical Number Seven, Plus or Minus Two, which argued that short-term memory holds about seven chunks, and that grouping increases how much information fits.

Also asked as: who gave chunking

Later research, notably by Nelson Cowan, suggested working memory capacity is closer to about four chunks. Chase and Simon's studies of chess players in the 1970s showed that experts remember board positions by recognising meaningful patterns as chunks, which explains much expert performance. In computing, chunking emerged independently as a practical technique for handling data in blocks, and in AI it later became central to retrieval pipelines that split documents for search.

What are some examples of chunking?

Everyday examples: remembering a phone number as 98765 43210 instead of ten separate digits, grouping a shopping list by aisle, learning a song verse by verse, and reading words as whole shapes rather than letter by letter.

Other examples include acronyms such as NASA, which pack several words into one unit; card numbers printed in groups of four; and musicians recognising chords instead of individual notes. In software, developers chunk code into functions and modules so each part is understandable on its own. In AI, a RAG system chunks a manual into sections, so a question about warranty retrieves just the warranty section rather than the entire document.

What are chunking examples for students?

Students can chunk by grouping related facts, breaking long topics into subtopics, using acronyms and mnemonics, and studying in focused blocks. For example, learn the periodic table in groups or families rather than as 118 separate elements.

In the classroom, teachers chunk lessons into short segments with a check for understanding after each, split reading into sections with questions, and present multi-step problems one step at a time. For programming, learn a language through small concepts, variables, loops and functions, then combine them in projects. The three-year roadmap I shared at the VSICS induction is itself chunking: a large goal broken into yearly and monthly milestones that students can act on.

What does chunking assignments mean?

Chunking an assignment means breaking a large task, such as a research paper or project, into smaller steps with their own deadlines, so it feels manageable and progress is steady instead of rushed at the end.

Example for a research paper: choose a topic by day 2, gather sources by day 5, outline by day 7, write the introduction and first section by day 10, complete the draft by day 14, and revise by day 17. Teachers often chunk assignments for students who struggle with planning, providing checkpoints and feedback at each stage. The same approach works for software projects, where large features are split into small tasks that can each be finished and tested.

What does chunking data mean?

Chunking data means processing or storing large datasets in smaller pieces, so they fit in memory, transfer reliably or can be processed in parallel. Examples include reading a huge CSV file in chunks or uploading a large file in parts.

In Python, pandas can read files with a chunksize parameter, processing one block of rows at a time. Web uploads use chunked transfer so failures restart from the last piece. Databases and array formats such as Parquet and Zarr store data in chunks for efficient access. In AI, chunking text means splitting documents into passages for embedding; chunk size and overlap strongly affect retrieval quality, because each chunk becomes one searchable unit.

What is a good chunking tutorial for RAG?

Start with recursive character splitting: split on paragraphs, then sentences, then words, targeting a few hundred tokens per chunk with modest overlap. Then evaluate retrieval on real questions and adjust size, overlap and boundaries based on results.

Better chunking respects structure: split by headings, keep tables and code blocks intact, and prepend each chunk with its document title and section heading so it makes sense alone. Store metadata such as page number for citations. Compare settings with a test set of questions and known answer passages, measuring recall at k. Coherent, self-contained chunks help both BM25 and embedding retrieval, which matters in hybrid systems like RAG.NextUpgrad.

What is a semantic chunking tutorial?

Semantic chunking splits text where the meaning changes, rather than at fixed lengths. Embed each sentence, measure similarity between neighbouring sentences, and start a new chunk where similarity drops sharply, signalling a topic shift.

LangChain's experimental SemanticChunker and LlamaIndex's semantic splitter implement this with a breakpoint threshold, such as a percentile of similarity drops. Steps: split text into sentences, embed them, compute cosine distances between consecutive sentences, choose breakpoints above a threshold, and merge sentences between breakpoints. Semantic chunking can improve coherence but costs extra embedding calls and may produce uneven chunk sizes, so cap the maximum length and compare it against simple recursive splitting on your evaluation set.

What is a late chunking tutorial?

Late chunking, introduced by Jina AI in 2024, embeds the whole document first with a long-context embedding model, then splits the resulting token embeddings into chunks and pools each chunk. Each chunk's vector therefore carries context from the entire document.

Traditional chunking embeds each chunk separately, so a chunk saying its revenue grew loses the company name mentioned earlier. Late chunking fixes this by letting every token attend to the full document before pooling. To implement it, use a long-context embedding model that exposes token-level outputs, run the document through once, compute chunk boundaries in token positions, and mean-pool tokens within each boundary. Documents longer than the model's context still need splitting into large windows.

What is a Docling chunking tutorial?

Docling, an open-source document conversion library from IBM, parses PDFs, Word files and slides into a structured document with headings, paragraphs, tables and figures. Its chunkers then split that structure into retrieval-ready chunks.

Convert a file with DocumentConverter, then pass the result to HybridChunker, which follows the document hierarchy and also respects a token limit from your embedding model's tokenizer, merging small neighbouring pieces and splitting large ones. Each chunk keeps metadata such as headings and page numbers, and a contextualise method prepends section headings to chunk text before embedding. Docling integrates with LangChain and LlamaIndex, making it a strong choice for structure-aware RAG ingestion.

Pranjul Rathour

GenAI Engineer · India

GenAI engineer and AI product builder with 2+ years shipping production-grade AI systems: RAG pipelines, fine-tuned LLMs, hybrid retrieval and multi-modal apps across vision, speech and OCR, architected end to end from ingestion to deployment. Leads engineering for SCULT INDIA's 14-member team, founded the 500+ member TechVerse Enclave community and has mentored 200+ students. Three hackathon first prizes: Changethon 2025 (IIT Roorkee), Product Genesis at Vividhotsava 2025 (CSJMU) and BYTEBATTLE (MeetKats).

Open to GenAI roles, hackathon judging, mentorship sessions and guest talks at colleges. For campus sessions the inviting institution covers the visit end to end: travel (a First AC train ticket for journeys up to five hours, a flight for anything longer), accommodation, meals, and pickup and drop; the honorarium and workshop fees are negotiable on WhatsApp or on the call, and requests are welcome from institutions that can cover the visit this way. Email pranjulrathour41@gmail.com.

PortfolioInvite me to your campusHire meLinkedInGitHubXInstagramThreadsFacebookBlueskyDev.toBloggerHashnodeSCULTEmail