RAG Chunking Strategy

Name: OpenModels
Creator: OpenModels
License: https://github.com/openmodelsrun/openmodels

IntermediatedataMinimum 32K context

Chooses how to split documents for retrieval so that chunks stay semantically coherent and answerable. Compares fixed-size, recursive, semantic, and structure-aware splitting, sets chunk size and overlap against the embedding model's limits, handles tables, code, and headings, and covers contextual headers plus parent-document retrieval to avoid losing surrounding meaning.

rag chunking retrieval embeddings document-processing vector-search

Use cases

Picking chunk size and overlap for a given embedding model
Splitting structured documents without breaking tables or code
Adding contextual headers so chunks stand alone
Diagnosing poor retrieval caused by bad chunk boundaries

Example prompt

Our RAG pipeline retrieves irrelevant passages from a corpus of technical manuals with headings,
tables, and code samples. Current setup: fixed 1000-character chunks, no overlap.

Recommend a chunking strategy: what splitter to use, chunk size and overlap for our embedding
model, how to keep tables and code blocks intact, and whether to add contextual headers or
parent-document retrieval. Explain how I would A/B test the change against retrieval metrics.

Recommended models

claude-opus-5 gemini-3-6-flash gpt-5-6-sol

Compatible tools

claude-codecursorkiroany

Modalities

Input: text, code

→

Output: text, code

Author

OpenModels Community

@openmodelsrun