Term
Chunking
Chunking is the splitting of longer texts into smaller, self-contained pieces — the first step of any RAG pipeline because embeddings and retrieval work at the chunk level.
Chunking — explained
An embedding model compresses a passage of text into a single vector. If the passage is too long, the model averages across too many topics and the semantic signal blurs. Too short, and context is missing — relevant matches get torn apart. Typical sizes land at 200–800 tokens per chunk, with a 10–20 % overlap so sentences sitting on the boundary keep their context. The chosen strategy — fixed token count, sentence/paragraph boundaries, semantic chunking, hierarchical chunking — has a strong effect on retrieval quality.
Example / Practical context
When indexing technical documentation, you might split by H2 section with a fixed-token-count fallback. Each chunk carries metadata (source, heading path, date) and is stored as an embedding in the vector database. At query time, typically 4–10 chunks are returned and merged by the LLM. Tools like LangChain and LlamaIndex ship standard splitters (RecursiveCharacterTextSplitter, MarkdownTextSplitter, and so on).
Distinction from related terms
Chunking is not tokenization — that splits individual words into sub-token units for the model. Chunking works one level up and splits whole documents into retrieval-friendly pieces. Contextual Retrieval adds an LLM-generated context sentence to each chunk so isolated chunks become more understandable on their own.
Entdecke mehr
AI Workflows by Keyword: How We Make Recurring Routines Enforceable
A typed keyword triggers a fixed AI routine — and every single step must be committed before the next one appears. Why that's the actual trick.
GlossarS-RAG
S-RAG (Search-optimized RAG) is boostN.ai's take on Retrieval-Augmented Generation. It extends classic RAG with search-engine principles: relevance ranking, recency decay and typed links between chunks for deterministic follow-up retrieval.
LexikonRAG Tuning and Balancing — Embeddings, Weights, Decay and MCP for Advanced Setups
How to steer retrieval on purpose: embedding choice, hybrid weights, reranker cascades, time decay, authority boost, MMR and MCP as a retrieval tool.