← All questions
RAGMedium· HyDE· Retrieval· Hallucination

Your hypothetical document invents a surgeon’s name and...

HyDE documents are not content for the user; they are probes for the vector space. By generating multiple independent samples and averaging them, the specific hallucinations (which vary across samples) are minimized, while the consistent domain vocabulary (which is shared) remains dominant. This argument holds unless the model is fundamentally wrong about the topic, in which case the error is amplified.

Code playground (Python)