Chunking

An employee asks the knowledge assistant whether a customer's email address counts as confidential. The data classification standard answers in one sentence: "Customer personal data, including names, email addresses and phone numbers, is Confidential at minimum." The retriever finds the right place, ranks it first, and hands the model a chunk that ends with "Customer personal data, including names, email addresses". The words "is Confidential at minimum" open the next chunk, which ranks second. When both reach the prompt, the model has to notice that two separate pieces of evidence form one sentence; when a tighter budget or a re-ranker keeps only the best chunk, the classification is simply gone.