What is the optimal chunking strategy for RAG systems processing technical documentation?

Optimal chunking for technical docs uses hierarchical semantic chunking: first split by document structure (chapters, sections, subsections), then apply recursive character splitting within sections with 512-1024 token chunks and 50-100 token overlap. Key strategies: semantic chunking uses embedding similarity to find natural break points (split when cosine similarity between consecutive sentences drops below a threshold); document-aware chunking respects markdown headers, code blocks, and table boundaries; and sliding window with dynamic sizing adjusts chunk length based on content density. For code documentation, keep function definitions with their docstrings as atomic units. Include metadata with each chunk: source document, section path, page number, and neighboring chunk IDs for context expansion. Benchmark your chunking against retrieval accuracy on a held-out test set, as optimal chunk size varies significantly by domain: legal documents prefer larger chunks (1500+ tokens) while API docs work better with smaller, focused chunks (256-512 tokens).