Question

A financial services company is building a customer-support assistant using Amazon Bedrock Knowledge Bases. The source data consists of conversational chat transcripts where a single customer issue often spans many short back-and-forth messages, and the meaning of any one message depends heavily on surrounding context. Early testing with the default fixed-size chunking (300 tokens) shows the retriever frequently returns fragments that split a single question-and-answer exchange, causing incomplete and confusing responses. The team wants Bedrock to group text into chunks based on topical/meaning boundaries rather than arbitrary token counts, without writing a custom ingestion pipeline. Which chunking configuration should they select when creating the data source?

Accepted Answer

Semantic chunking, which uses an embedding model to group adjacent sentences by meaning and split at points of low semantic similarity

Answer

Fixed-size chunking with a larger token size (e.g., 1000 tokens) and 20% overlap to reduce split exchanges

Answer

Hierarchical chunking, which creates parent and child chunks to preserve document structure across headings

Answer

No chunking, ingesting each transcript file as a single chunk so no exchange is ever split