Question

A developer is building a Bedrock Knowledge Base to power a RAG assistant over 5,000 highly structured technical maintenance manuals. Each manual uses consistent Markdown headings (##) for procedures, and users frequently ask questions that must return a complete, self-contained procedure without mixing content from unrelated sections. Early tests with a fixed 300-token chunking strategy return fragmented answers that cut procedures in half or blend adjacent procedures. Which chunking approach should the developer adopt to improve retrieval quality?

Accepted Answer

Switch to hierarchical chunking so parent chunks provide broad context while smaller child chunks are used for retrieval matching, keeping related procedure content grouped.

Answer

Increase the fixed chunk size to 1,000 tokens with 20% overlap so more content fits in each chunk regardless of document structure.

Answer

Disable chunking entirely (no chunking) so each manual is embedded as a single vector, preserving all procedures together.

Answer

Use semantic chunking based on embedding similarity thresholds to split text wherever topic drift is detected across sentences.