Text Chunking Strategies for RAG and Vector Search

Retrieval-Augmented Generation (RAG) systems rely on vector databases to retrieve relevant document passages. Chunking documents properly ensures that retrieved text chunks retain context without exceeding embedding model limits.

Experiment with chunking parameters using the RAG Text Chunker.

Popular Chunking Strategies

1. Fixed-Size Chunking with Overlap: Splitting text every $N$ characters (e.g. 500 chars) with a 50-char overlap to prevent splitting sentences in half.
2. Recursive Character / Markdown Chunking: Splitting hierarchically by headings (#, ##), paragraphs (\n\n), sentences (. ), and finally words. Keeps semantic sections intact.
3. Document Structure-Aware Chunking: Preserving code blocks, HTML tables, and markdown lists within single chunks.

# Example: Recursive markdown splitting logic
separators = ["\n# ", "\n## ", "\n### ", "\n\n", "\n", " "]

Test chunk size, overlap, and token estimates with the RAG Text Chunker.