Retrieval‑based systems select from a set of predefined responses or templates based on the user’s query.

They are fast and reliable but limited to covered scenarios.

Generative approaches use language models to construct new responses, enabling more flexible and natural interactions but introducing risks of hallucination and longer latency.

In analytics, retrieval might map common questions to canned queries, while generative models parse and compose SQL on the fly.

Many platforms combine both: retrieval handles frequent tasks and generative models cover long‑tail questions.

Using a semantic layer and grounding techniques reduces hallucinations.

MageMetrics hybrid model retrieves known metrics definitions and uses generative models for novel expressions.