Chunk isolation is the practice of separating retrieved text into tightly controlled pieces before it is passed to an AI model. In a RAG system, documents are often split into chunks for search and ranking, but unsafe chunking can let hidden instructions sit beside trusted facts and influence the model’s behavior.
This matters because indirect prompt injection often works by burying malicious directives inside ordinary-looking content. Strong chunk isolation helps reduce instruction leakage across context by keeping untrusted material separate, labeling its source, and limiting how much of it the model can see at once. Defenders use it with provenance tracking, prompt boundaries, and output filtering. In real attacks, a poisoned web page or uploaded file may only trigger when its chunk is retrieved and mixed with the user query, so isolation is a key control for lowering the chance that retrieved data can steer answers or tool actions.