Question

A legal firm builds a RAG application on Amazon Bedrock that retrieves from a knowledge base containing contracts from multiple clients. During testing, when an attorney asks a question about Client A's contracts, the system sometimes returns passages from Client B's documents, creating a confidentiality risk. The retrieval relevance is otherwise good. What is the MOST appropriate way to ensure each query only retrieves documents belonging to the correct client?

Accepted Answer

Apply metadata filtering on the client identifier so that retrieval is scoped to only the requesting client's documents

Answer

Fine-tune the foundation model on each client's contracts separately to prevent cross-client leakage

Answer

Lower the temperature parameter so the model produces more deterministic, client-specific answers

Answer

Reduce the top-k retrieval value to 1 so only a single passage is ever returned