Giving a manuals RAG assistant the steps around a match
MLA-C02Data Preparation for ML and AIMediumQuestion 16 of 30
A manufacturing company is building a RAG assistant with Amazon Bedrock Knowledge Bases over 3,000 equipment service manuals. Each manual section runs several pages. Technicians ask narrow questions, such as one torque value, but a correct answer usually needs the procedure steps around that value in the same section. With default chunking, retrieval finds the right sentence, but the model often answers without the steps around it.
Which chunking configuration will meet these requirements?
Show answer and explanation
BUse hierarchical chunking with small child chunks for search and larger parent chunks
Hierarchical chunking meets the requirements. The knowledge base searches the small child chunks, which embed precisely, and then replaces each match with its larger parent chunk, so the model receives the surrounding procedure as well as the matching sentence. Fixed-size chunks of 20 tokens with no overlap make the problem worse, because each result carries even less of the section. No chunking embeds a whole manual as one vector, which blurs a narrow question against hundreds of pages and sends far more text than the answer needs. Semantic chunking with a zero buffer and a low breakpoint threshold splits the text more often, which produces more and smaller chunks, not more context.