Changing the chat endpoint moves only one part of a retrieval system. A document can also reach an embedding API, a reranker, a vector database and a tracing service. Each can have a different owner and retention policy.
Follow one fictional paragraph
Index a test document containing an invented return policy. Ask a question that requires that paragraph. Record each destination that receives the paragraph or a derived representation. If you change embedding models, validate retrieval quality and vector compatibility before rebuilding the production index.
Open the TrustedRouter migration guide · Can AI outputs and tool results expose private data?