AWS Certified AI Practitioner (AIF-C01) free question 6 / 10 · Domain 3: Applications of Foundation Models
A company's internal policies are revised every month. The team wants to build a generative AI app that answers employee questions based on the latest policies. Which build approach matches the AWS page?
- Search the policies with a knowledge base and add the retrieved content to the prompt
- Paste the full text of the policies at the start of the prompt each time and use prompt caching to cut cost
- Fine-tune the model each month on the internal policies and reflect the latest policies in its weights
- Lower the temperature and have the model pick a likely answer from the knowledge it has
Show the answer and explanation
Answer: A
Why this is correct
Amazon Bedrock Knowledge Bases is a capability for using retrieval augmented generation (RAG), a technique that retrieves information from a data store to augment an LLM's response. The page says knowledge bases abstract away the heavy lifting of building the pipeline and provide an out-of-the-box RAG solution, and that adding a knowledge base removes the need to continuously train a model to use private data, which improves cost-effectiveness. When policies change, you sync the data source and the new policies become searchable.
Why the other options are wrong
"Fine-tune the model each month on the internal policies and reflect the latest policies in its weights" … The page says that with a knowledge base you do not need to keep training the model to use your private data, which improves cost-effectiveness. A monthly retraining is exactly the effort a knowledge base removes.
"Lower the temperature and have the model pick a likely answer from the knowledge it has" … Foundation models are often trained offline, so they are unaware of data created after training. Changing the temperature does not bring the latest policies, which live outside the model, into the answer.
"Paste the full text of the policies at the start of the prompt each time and use prompt caching to cut cost" … The RAG page says the retrieved documents should be large enough to contain useful context but small enough to fit within the prompt's maximum sequence length. Pasting the full text every time runs into the length limit and loses the benefit of RAG, which passes only the relevant parts.
Key point
RAG suits uses that answer from data created after the model was trained or from private internal data. A knowledge base can include citations in responses so sources can be checked. When you change the data source, you need to sync it to pick up additions, changes, and deletions.