Contextual Knowledge
What the model knows from the prompt, right now.
Facts supplied in the context window — retrieved documents, tool results, the user’s message. Editable per request, attributable, and current.
The practical rule for building reliable AI features: prefer supplying knowledge in context over hoping it lives in the weights. Context is inspectable; weights are not.
Its strengths are exactly the inverse of parametric knowledge: attributable, correctable per request, and current. Its weakness is cost and attention — every retrieved fact spends window space, so the discipline is retrieval quality rather than retrieval volume.
When the context and the memory disagree
A retrieved document contradicts what the model recalls from training.
Say so explicitly in the system prompt: supplied context takes precedence, and note the conflict when it is material. Otherwise the model arbitrates silently, and you find out which way it went only when a user complains.
- blogEffective context engineering for AI agents ↗
Defines context engineering as curating everything supplied at inference, names context rot, and covers just-in-time retrieval, compaction and keeping the window high-signal.
- paperRetrieval-augmented generation for knowledge-intensive NLP tasks ↗
The original RAG formulation: combine a parametric model with a non-parametric dense index, so answers can be grounded in retrieved text.
- docsGrounding overview ↗
Grounding responses against search results or your own corpora, returning support citations that tie individual claims back to sources — the API shape of provenance.
Parametric Knowledge
What the model stores inside its weights.
Context
Everything the model can currently see.
RAG · Retrieval-Augmented Generation
Look things up, then put what you found in the prompt.
Context Engineering
Deciding what goes in the window, in what order, at all.