Choosing foundation models for a use case: AIP-C01 study notes and practice questions
What to know about Choosing foundation models for a use case for the AIP-C01 exam, from the Foundation Model Integration, Data Management, and Compliance domain (31% of the exam guide), then 1 free question to test it. Part of the AIP-C01 study guide.
AIP-C01 gives document sizes, input types, languages, budget and Region rules, then asks which model fits. Most wrong options are a real model that fails one stated requirement, so screen on hard requirements first.
Screen on hard requirements
Context window: the model must fit the longest realistic input plus the output. A large window does not guarantee accurate use of long context, so retrieval may still be the better design.
Modality: image plus text questions need a multimodal model. Audio and video usually need a transcription or Bedrock Data Automation step first.
Tool use and structured output: agents and API-calling features need tool use support. JSON schema output is available for some models only.
Region and lifecycle: confirm the model is offered in your Region, directly or through an inference profile, and that its status is Active. Models in Legacy status have an end date, so a new build should not start on one. See the model lifecycle page.
Then compare quality, latency and cost
Public leaderboards measure general skill. The exam expects you to build an evaluation set from your own prompts, score it with automatic metrics, LLM-as-a-judge or human review, and compare finalists on quality per dollar and latency. A smaller model that clears the quality bar is usually the right answer over the largest model.
Language and output quality
Check every language your users write in. Quality varies by language, so test each one on your evaluation set rather than assuming the English result carries over.
How to choose
Scenario wording
Choose
“200-page manuals must fit in one prompt”
A large context window model, then test whether retrieval is cheaper
“Answer questions about charts in uploaded images”
A multimodal model that accepts image input
“Customers write in several languages”
Per-language quality tests on your own data
“Must run in one Region by policy”
A model offered in that Region, or an approved inference profile
“Agent calls tools with validated arguments”
A model with tool use support and structured output
“Current model is marked Legacy”
Migrate to the named successor model
Common traps
Choosing the top leaderboard model without testing it on your data.
Using the largest model for every task, when a smaller model meets the bar at a fraction of the price.
Ignoring Region availability, since a model you can call in one Region may be missing in another.
Assuming a big context window removes the need for RAG, when cost and accuracy still matter.
Starting new workloads on Legacy models that have a scheduled end date.
A legal publisher runs a research tool that sends each user's question, together with the court opinions that match it, to a foundation model on Amazon Bedrock.
Most opinions are 2,000 to 3,000 words long, and short questions should come back fast and at a low price. The large-context models the team tested cost several times more per token than its small default model.
A few full case files run to about 300,000 tokens, and an answer can depend on any part of such a file. The team will not split these files by hand.
Which TWO approaches meet these requirements? (Choose two.)
Choose two answers.
AA. Use one large-context model for every request, so that any case file fits in a single prompt and the application code stays simple.
BB. Check the token count of each request, send short questions to a low-cost model, and send the few very long case files to a model whose context window holds about 300,000 tokens.
CC. Fine-tune a small model on the case files so that it memorizes their contents, and then send only the user's question at request time, with no case text attached.
DD. Load the case files into a knowledge base that chunks them automatically, and retrieve only the passages that match each question for the low-cost model.
EE. Truncate every case file to its first 4,000 tokens before sending it to a low-cost model, because the answer to most questions appears near the start of an opinion.
Show the answer and explanation
Answers: B, D
A. Wrong: A large-context model can hold the longest file, but the team measured a much higher price per token for it, and every short question would pay that price. That breaks the requirement for fast, low-cost answers to the common requests.
C. Wrong: Fine-tuning does not reliably store the full text of long opinions, so answers may be inaccurate or incomplete. Each new case file would also require another training job before it could be used.
D. Right: The knowledge base chunks the files during ingestion, so no one splits them by hand, and retrieval can reach a passage from any part of a long file. The small model sees only a few passages per request, which keeps latency and cost low. Learn more: Chunking strategies for knowledge bases (opens in a new tab)
E. Wrong: Truncation discards the later parts of long opinions, where holdings and dissents often appear, so answers would be wrong for many questions. It meets the cost goal but breaks the requirement that an answer can come from any part of the file.
Explanation
Routing by request length keeps routine traffic on the cheap, fast model while a long-context model handles the rare huge files. A knowledge base also works, because it chunks the files automatically and retrieval can reach content anywhere in a file. Fine-tuning and truncation both lose content, and a single large-context model is too costly for everyday questions.
Want more on Choosing foundation models for a use case? The 5 full-length practice sets have 5 more Choosing foundation models for a use case questions, 375 in all, none of them in the free exam. One payment of $14.99, no subscription. See the practice sets →