Chunking

Chunking is the practice of splitting documents, code, or data into smaller retrieval units before they are embedded and indexed, so a retriever can return focused passages instead of whole files.

Chunk size and boundaries are a balance: chunks too large dilute the relevant signal and waste context, while chunks too small lose the surrounding meaning needed to answer. Respecting natural structure — headings, paragraphs, functions — and attaching metadata usually beats fixed-length splitting.

For builders it is the most underrated lever in RAG quality: the same corpus and Embeddings model can succeed or fail purely on how it was chunked. The frequent mistake is one global chunk size applied to heterogeneous content, where a single strategy never fits every document type.

Topic: RAG Related: RAG Embeddings Codebase Indexing

Recent Updates

• 2026-06-29: How Exam AI uses embeddings, staged generation, and human review to build an internal exam platform (Все статьи подряд / Искусственный интеллект / Хабр) · habr.com — Exam AI FastAPI sqlalchemy · PostgreSQL · pgvector · Pydantic TaskIQ · Redis · MinIO · React · TypeScript · Vite TanStack Query Zustand xState Orval Tailwind DaisyUI · Keycloak · Docker · Traefik · Habr
• 2026-06-30: YuMoney describes an internal RAG assistant for corporate knowledge (Все статьи подряд / Искусственный интеллект / Хабр) · habr.com — YuMoney Bitbucket Ilya FRIDA
• 2026-07-01: VK describes Discovery AI neurosearch architecture for large-scale content search (Все статьи подряд / Искусственный интеллект / Хабр) · habr.com — AI VK VK Mail Dzen VK Video Evgeny Astafurov
• 2026-07-03: Study Finds Cluster-Based Chunking Did Not Beat Simpler RAG Splits on Academic Theses (cs.CL updates on arXiv.org) · arxiv.org
• 2026-07-04: A Russian RAG splitter that returns boundary indexes instead of rewritten text (Все статьи подряд / Искусственный интеллект / Хабр) · habr.com — Dify · Unsloth · DeepSeek · Milvus · Qdrant T-lite-it-2.1 context-aware-splitter-1b · TinyLlama · LLaMA-2 · Qwen2.5 · DeepSeek V4 Flash
• 2026-07-05: Fine-tuning Qwen3-4B-Instruct-2507 for Karachay-Balkar with dialect augmentation and a new tokenizer (Все статьи подряд / Искусственный интеллект / Хабр) · habr.com — Hugging Face UNESCO · Yandex · Qwen3-4B-Instruct-2507 TSjB/QM-4B · Claude · Gemini NLLB-200
• 2026-07-06: Contextual Retrieval Adds Document Context Before Indexing RAG Chunks (Все статьи подряд / Искусственный интеллект / Хабр) · habr.com — Anthropic · Habr ACME Corporation
• 2026-07-06: Building a threat-intelligence database for LLM agents and SOC (Все статьи подряд / Искусственный интеллект / Хабр) · habr.com — BAAI MinIO · FastAPI · PostgreSQL · PyMuPDF Tesseract · Redis · Celery · LangChain STIX TAXII MITRE ATT&CK · pgvector · BGE-M3
• 2026-07-07: Candidate-Constrained RAG for LongEval-RAG uses rule-based chunking and sentence-level selection (cs.CL updates on arXiv.org) · arxiv.org — arXiv LongEval-RAG · MiniLM BERTScore
• 2026-07-07: HiLS Attention learns chunk selection end to end for ultra-long context (cs.CL updates on arXiv.org) · arxiv.org — arXiv
• 2026-07-07: Building a Local RAG for Procurement Documents with Qdrant and LlamaIndex (Все статьи подряд / Искусственный интеллект / Хабр) · habr.com — Qdrant · LlamaIndex · Gradio · Pydantic pydantic_settings · PyMuPDF Vladimir · Qwen3-14B intfloat/multilingual-e5-base cross-encoder/ms-marco-MiniLM-L-6-v2
• 2026-07-08: Token-Efficient Retrieval for Legal Document Analysis (cs.CL updates on arXiv.org) · arxiv.org — GTE
• 2026-07-09: RAG study shows retrieval quality matters for public health QA (cs.CL updates on arXiv.org) · arxiv.org — UK government · arXiv
• 2026-07-11: Semantic search in the browser with a tiny lookup-table embedding model (Hacker News) · bart.degoe.de — Lunr.js Transformers.js sentence-transformers model2vec potion Hugo · PyTorch · WebAssembly · ONNX · all-MiniLM-L6-v2 potion-base-8M
• 2026-07-15: Habr experiment: benchmarking Karpathy-style LLM-wiki against simple vector RAG (Все статьи подряд / Искусственный интеллект / Хабр) · habr.com — Andrej Karpathy Gemma4-31B GPT-5.1
• 2026-07-17: How to Structure Pages for AI Retrieval: Chunking, Schema.org, robots.txt, and Measurement (Все статьи подряд / Искусственный интеллект / Хабр) · habr.com — OpenAI · Perplexity · Google
• 2026-07-21: A Go-based MCP for local code search and browser automation (Все статьи подряд / Искусственный интеллект / Хабр) · habr.com — Windsurf Cascade · Chrome · Ollama · Playwright
• 2026-07-27: Benchmarking Fine-tuning and Retrieval Strategies for a Multimodal Language Model on the NRC Reactor Operator Licensing Examination (cs.CL updates on arXiv.org) · arxiv.org — Gemma-4-31B-IT · Gemini

FAQ

What is Chunking?

Chunking is the practice of splitting documents, code, or data into retrieval units for RAG and indexing. GROUNDING tracks chunk size, boundaries, metadata, context windows, and when chunking harms answer quality.

Which topic does Chunking belong to?

On the GROUNDING radar, Chunking is grouped under the RAG topic.

Which concepts are related to Chunking?

Related concepts tracked by the radar include RAG, Embeddings, Codebase Indexing.