Information represented in a model’s learned parameters through training, rather than supplied solely through retrieval or current context.
Parametric knowledge is information represented in a model’s learned parameters, often called weights. Training shapes those parameters, allowing the model to generate responses based on learned patterns. This differs from retrieving a particular document and supplying it as context for a current question.
The information is not permanently fixed in every sense: additional training can change model parameters. However, a normal conversation that supplies a new fact should not be assumed to update the underlying model. Context available during an interaction and knowledge learned through training are different mechanisms.
A model’s generated statement does not provide a dependable inventory of what it learned or a traceable source for every claim. Learned information may be incomplete or outdated. When current facts matter, use appropriate source evidence rather than treating confident recall as proof.
Takeaway: Separate what an assistant appears to know from what it has actually checked. Current business details need an explicit maintenance and verification process.
Example
A model may describe a product from patterns learned during training while a retrieved catalog shows a newer specification. The application should evaluate the current catalog’s relevance rather than assume the model’s familiar-sounding description remains correct.
What to check
Ask whether an answer relies on provided documents, retrieval or unsupported model recall. Do not describe a document upload as automatically retraining the model. If fine-tuning is proposed, clarify exactly what changes and how it will be evaluated.
For more detail, see Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks.