Most RAG Hallucinations Are Extraction Errors: Seven Patterns for a Typed Generation Contract

Most errors in Retrieval-Augmented Generation (RAG) systems are extraction failures, not hallucinations from parametric memory. To address this, systems should use a typed generation contract instead of returning a simple string answer. This involves treating the LLM as a function that populates a structured object, like a Pydantic schema, which includes the answer, citations, and self-assessment flags. The article details seven patterns for implementing this approach, such as extracting raw values for later computation and using structural cues to determine completeness.