Chunking is the quietest lever in the RAG pipeline and one of the most consequential: split too big and retrieval returns walls of irrelevant text, too small and each passage loses the context that made it answerable.
Size and overlap
Start around 300–500 tokens per chunk with a small overlap (10–15%) so a sentence split across a boundary still appears whole in one chunk. Tune from there against your own questions — there's no universal number, only what retrieves well for your content.
Split on structure, not character counts
Fixed-length splits cut tables and lists in half. Split on document structure instead — headings, sections, paragraphs — so each chunk is a coherent unit. This is exactly why converting to clean Markdown first pays off: headings become natural, reliable boundaries.
Chunking feeds everything downstream
Good chunks make embeddings for RAG more precise and give hybrid search cleaner units to match against. Bad chunks can't be rescued by a better model. Kit for AI chunks on structure automatically when you ingest a document — Start free and inspect how your content splits.