Retrieval augmented generation over an indexed internal: Which technique best fits this constraint?
Retrieval augmented generation grounds answers in the latest internal documents at query time without retraining, satisfying the no-weight-change and always-current constraints.
The question
A bank wants a general-purpose model to answer staff questions using its constantly-changing internal policy documents, but it will not retrain or alter the base model's weights and must reflect the latest documents at query time. Which technique best fits this constraint?
Preparing for AIGP? Take the free 5-min readiness quiz →
- Fine-tuning the base model on a labeled snapshot of the documentsA plausible improvement method, but fine-tuning alters weights and bakes in a snapshot, violating both stated constraints.
- Deploying an agentic architecture that chains multiple tool-using stepsA real technique for complex tasks, yet agentic orchestration does not by itself supply the latest documents at query time.
- Retrieval augmented generation over an indexed internal document store ✓RAG injects current documents into the prompt at query time without changing the base model, matching the no-retrain, always-current constraint.
- Moving the model to edge devices to lower inference latency for staffA valid siting decision, but edge deployment addresses latency and location, not grounding answers in current documents.
The trap
Choosing fine-tuning to inject domain knowledge when the requirement is up-to-date, retrieval-grounded answers without altering the model. How to remember it
Retrieval augmented generation grounds answers in the latest internal documents at query time without retraining, satisfying the no-weight-change and always-current constraints.
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
- Open-source versus proprietary access to the model: Which model-type distinction is being weighed? →
- Performing or reviewing an impact assessment: Before go-live, which activity most directly evaluates the →
- The data-use and training-rights clause governing customer: Which term in the vendor agreement should the →
- 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.
