Parametric and contextual knowledge
A model answers from two very different sources of knowledge, and telling them apart explains much of how the platform is designed.
- Parametric knowledge is what the model absorbed during training and stored in its parameters. It is broad and always available, but it is frozen at the model's knowledge cutoff, cannot be inspected, and is reconstructed from compressed patterns rather than looked up, which is why it can be subtly wrong.
- Contextual knowledge is what the model reads directly from its context window at the moment of the request: the documents, data and conversation you put in front of it. It is current, specific and traceable to a source, but limited to what fits in the window.
📝Why the distinction matters
Parametric knowledge is what the model remembers; contextual knowledge is what it can see right now. When the two disagree, grounding the model in trustworthy contextual knowledge is what keeps answers accurate and verifiable.
Comparative view
| Aspect | Parametric knowledge | Contextual knowledge |
|---|---|---|
| Source | Baked into the model's weights during training | Supplied in the context window at request time |
| Freshness | Frozen at the knowledge cutoff | As current as the information you provide |
| Scope | Broad, general | Narrow, specific to the task |
| Traceable to a source | No | Yes |
| How to change it | Retrain or fine-tune | Change what you put in the context |
Why it matters in Rational AI
The whole point of grounding a model in your Knowledge is to shift answers from parametric to contextual. RAG retrieves the most relevant resources and places them in the context window, so the model answers from your current, citable material instead of relying on what it happened to memorise during training.