Retrieval augmented generation: Which option best fits all three requirements?
RAG retrieves the latest documents at query time, keeping content out of the weights and requiring no retraining on updates.
The question
A firm wants a foundation model to answer staff questions using internal policy documents that change weekly, without embedding those documents into model weights, and with minimal re-engineering each time content updates. It is choosing an enhancement technique. Which option best fits all three requirements?
Preparing for AIGP? Take the free 5-min readiness quiz →
- Full fine-tuning on the policy corpus, which bakes the documents into the model's weights so it can recall the content directly during use.Plausible as customization, but fine-tuning embeds content into weights and requires re-training on every weekly update, violating two constraints.
- Retrieval augmented generation, which fetches the current documents at query time so updates need no weight changes and stay outside the model. ✓Correct because RAG grounds answers in the latest documents at inference, avoids baking content into weights, and adapts to weekly updates without retraining.
- Pre-training a new model from scratch on the policy corpus, which yields maximal control over exactly what knowledge the model contains.Plausible for control, but from-scratch pre-training is extremely costly, embeds content into weights, and is impractical for weekly changes.
- An agentic architecture that chains planning steps, which improves multi-step task execution but does not itself keep document knowledge current.Plausible as an enhancement, but agentic chaining addresses task orchestration, not grounding answers in frequently changing source documents.
The trap
Defaulting to fine-tuning to inject proprietary knowledge when the content changes constantly. How to remember it
RAG retrieves the latest documents at query time, keeping content out of the weights and requiring no retraining on updates.
How many of these would you get right?
One of 1581 AIGP questions on Certsqill. Take a free five-minute check and see your score per domain — not one number, but which section to open tonight.
Test your AIGP readiness — freeMore Understanding How to Govern AI Deployment and Use questions
- An open-source model the team can self-host and inspect: Which profile best satisfies all three constraints? →
- The assessment omits evaluation of potential harms: Which gap should the reviewer flag as most critical before →
- The vendor's broad reuse of customer data for training: Given the deployer's own regulatory and safety →
- All 424 Understanding How to Govern AI Deployment and Use questions →
Part of the Certsqill AIGP question bank · Understanding How to Govern AI Deployment and Use ·
Every answer, right and wrong, comes with its own explanation.
