They concern different uses and should be discussed separately in the commercial schedule. A generic permission for “AI” is too imprecise to price or operate confidently.
RAG combines a generative model with retrieved information; the foundational research describes combining model parameters with external memory. That technical distinction helps explain why ongoing retrieval access can have a different commercial purpose from training. Lewis et al. — Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
| Use proposed | Question the owner should resolve | Suggested commercial boundary |
|---|---|---|
| Broad training | Which models and training runs are covered? | Specify corpus version, permitted models and further training |
| Fine-tuning | Which specialised task and recipients? | Separate adaptation from general model development |
| Retrieval / grounding | What may be indexed, cached, displayed and cited? | Define freshness, attribution, extracts and user entitlements |
| Evaluation | Is the content only used to measure performance? | Protect test answers and prohibit unapproved training reuse |
| Derived datasets / synthetic outputs | May the buyer create substitutes or distribute transformations? | Define allowed derivatives and downstream recipients |
These are proposed negotiation categories, not statements that each is a distinct statutory right. Have counsel translate the business intent into the applicable grant and restrictions. Include embeddings, caches, subcontractors and affiliates in the discussion; omitting them leaves the operational scope unclear. Price any optional expansion when it is requested, rather than assuming it has no additional value.