Context Engineering for RAG : The Four Typed Inputs Behind Every RAG Answer

Quick take

Context engineering for Retrieval-Augmented Generation (RAG) is about assembling four specific types of inputs into a single query for a large language model. Named by Tobi Lütke and Andrej Karpathy in 2025, this approach breaks down a document’s data into typed pieces that combine to improve RAG responses. These include the core document content, a curated corpus, active conversation context, and tool-based extensions, each adding layers of relevant information before the LLM call.

Why it matters

Operators building RAG systems often treat context as a simple chunk of text. This practice presses them to rethink input design. Understanding the four typed inputs clarifies how to package the right signals for better accuracy and relevance. This splits what was once a fuzzy, one-size-fits-all context into purposeful components, reducing wasted tokens and improving answer quality. Instead of blindly dumping large corpora, teams can prioritize dynamic conversation context or inject tool outputs meaningfully.

The method forces a more disciplined way to integrate external knowledge, conversation state, and executable tools. It exposes previously hidden complexity in crafting RAG inputs and raises the bar for how builders structure their pipelines. For businesses, this means more predictable control over RAG answer quality and fewer bad outputs triggered by noisy or irrelevant context. It also shifts how RAG projects estimate cost and latency by focusing on what specific context types to prioritize during model calls.

AI builders and operators who grasp this approach will improve their RAG workflows by understanding that context is not just bigger or more data, but better typed and better combined. This narrows the guesswork in prompt design and token usage while unlocking easier ways to add new knowledge sources.

AI Quick Briefs Editorial Desk