A recent article in Towards Data Science argues that most "hallucinations" in Retrieval-Augmented Generation (RAG) systems are actually extraction errors, not fabrications from the model's internal knowledge. The author, [Author Name], proposes that when a RAG model provides an incorrect answer, the issue typically stems from the preceding stages of the extraction chain—parsing, question formulation, retrieval, or the generation contract itself—rather than the model inventing information.

The traditional RAG approach often treats the LLM as an oracle that outputs a raw answer string. However, the article advocates for a paradigm shift where the LLM acts as a function, and the output is a typed Pydantic object. This "typed generation contract" includes structured data, citations, fidelity flags, and self-assessment fields, enforced by a validator before the user sees the result.

Seven patterns are outlined to ensure the generation process adheres to this contract. These patterns emphasize extracting typed values, ensuring completeness through structure rather than self-rating, and using two booleans instead of a single confidence float to represent answer status. A key takeaway is that the LLM should extract, and external logic should compute, preserving an audit trail. The article also highlights the importance of using one prompt per answer shape, dispatched at runtime, to maintain traceability and avoid the complexity of monolithic prompts. This approach aims to make RAG systems more reliable and auditable.