Parametric knowledge
Patterns and facts encoded in model parameters during training rather than supplied in context.
Illustrative conversation
- Where did that framework fact come from if no document was retrieved?
- It likely came from parametric knowledge learned during training, so verify it if accuracy or recency matters.
Definition
Parametric knowledge is information a model has absorbed into its learned weights during training. It lets the model answer without retrieving an external document, but it can be incomplete, outdated, difficult to attribute, or recalled incorrectly. Retrieval-augmented systems combine this internal capability with non-parametric, sourceable knowledge supplied in the current context.
Reviewed