AWS Certified Generative AI Developer - Professional
Independent: not affiliated with AWS. Free to try; the study bundle adds a study guide and two practice exams.
5 original questions, none of them in the practice exams.
Pick an answer. Show answer explains every option and links its AWS source.
5 questions / Explanations link AWS sources / Last updated Oct 26, 2026
Question 1
Multiple choice
Domain 1 / Task 1.4 Vector store solutions
A property management company stores building inspection reports in an Amazon Bedrock knowledge base. Each report has a metadata file with city, inspection_year and property_type attributes. Agents type free-text questions such as 'Which Denver apartment buildings inspected in 2025 had roof damage?', and retrieval returns chunks from other cities and years. The application has no filter controls, and the team does not want to write code that extracts constraints from each question. Which change meets the requirement?
Show answerHide answer
Answer: D
Why
With implicit metadata filtering, Amazon Bedrock Knowledge Bases generates a retrieval filter from each query and a metadata schema. The implicitFilterConfiguration takes the metadata attribute schemas and the ARN of the model that writes the filter, so the city and year in a question become a filter with no application code.
Why not the others
AHybrid search combines vector search with raw-text search; it is a search strategy, not a constraint on the city or inspection_year attributes, so chunks from other cities and years can still be returned.
BRaising numberOfResults only raises the maximum number of chunks returned (five by default); it does not constrain which cities or years the chunks come from, so more of them can be off target.
CHierarchical chunking changes which text is returned for a match by replacing child chunks with their parent chunk; it does not filter results by metadata.
A development team is building a chat feature on Amazon Bedrock. The feature must support tool use, and the team must be able to switch between foundation models from different providers without rewriting the request and response handling for each model. Which approach meets these requirements?
Show answerHide answer
Answer: A
Why
The Converse API provides one request and response shape that works with all Amazon Bedrock models that support messages, so the same code runs against different models. It supports tool use, and parameters unique to a model go in the additionalModelRequestFields structure.
Why not the others
BInvokeModel uses each model's native request and response format, so switching models means rewriting the integration.
CThis adds endpoints to operate and a custom normalization layer to maintain, while Amazon Bedrock already provides the common interface.
DBatch inference generates responses asynchronously from input files in Amazon S3 and does not serve an interactive chat.
A platform team created an Amazon Bedrock guardrail with denied topics and a sensitive information filter and published it as version 2. Application teams call Converse and InvokeModel from their own IAM roles, and an audit found requests that omitted the guardrail. The platform team must make any inference request that does not include guardrail version 2 fail, without changing the applications. Which approach meets the requirement?
Show answerHide answer
Answer: B
Why
The bedrock:GuardrailIdentifier condition key lets an IAM policy deny any inference request that does not name the guardrail. AWS's example pairs an Allow on the guardrail ARN and version with an explicit Deny on any other value, and it covers Converse, ConverseStream, InvokeModel and InvokeModelWithResponseStream.
Why not the others
AThis changes every application, which the requirement rules out, and a request that omits the guardrail still succeeds whenever a team forgets it or a review misses it.
CInvocation logs and a weekly review can only look back at requests after they have run; they do not make a request that omits the guardrail fail.
Dbedrock:ApplyGuardrail lets a role use a guardrail; an Allow neither requires nor attaches one, so a request that omits the guardrail still runs without it.
An application lets analysts upload a 90-page supplier agreement and then ask a series of questions about it. Each question is sent to a model on Amazon Bedrock through the Converse API with the full agreement included in the prompt. Input token costs are high and each answer is slow to start. Which change reduces both cost and latency with the smallest change to the application?
Show answerHide answer
Answer: C
Why
Prompt caching reuses a repeated prompt prefix, so the agreement is processed once and later questions read it from the cache at the cache-read rate, with lower latency. The agreement goes before the cachePoint and the question after it, because a change to the prefix is a cache miss.
Why not the others
ATemperature controls randomness and maxTokens caps the length of the output; neither changes the input tokens that are re-read with every question.
BProvisioned Throughput reserves capacity against throttling, which the stem does not describe; it does not reuse the repeated agreement, so every question still reprocesses the full document.
DBatch inference is asynchronous and does not serve interactive questions, and AWS states that prompt caching is not supported with the batch inference API.
A company is comparing two candidate models for a summarization feature. One is an Amazon Bedrock model; the other runs on the company's own GPU servers and must stay there. The team wants both models scored by the same judge model on the same 400 prompts, with an explanation for each score, using Amazon Bedrock model evaluation jobs. How should the team evaluate the external model?
Show answerHide answer
Answer: D
Why
A judge-model evaluation job accepts your own inference responses in the prompt dataset, in the modelResponses field, and skips the invoke step, so a model outside Amazon Bedrock is scored by the same judge and metrics, with an explanation for each prompt. The dataset allows up to 1,000 prompts, so 400 fit.
Why not the others
ACustom Model Import copies the weights into Amazon Bedrock, which the requirement that the model stay on the company's servers rules out, and the job accepts the model's responses directly.
BA SageMaker AI endpoint you host yourself is not a supported generator model type for the job, and hosting there would move the model off the company's servers.
CA human-based job returns ratings from a work team, not scores and explanations from the judge model, so the two models would not be scored on the same basis.
The study bundle has two practice exams with 150 original questions, an answer key and an explanation for every answer, plus the 67-page study guide. Practice exam explanations are shorter than the ones on this page: a short paragraph per question, with no source links.
Study bundle
$19
150 practice questions / Study guide / Answer keys
Every bundle includes 6 months of access: download the PDFs, and any new editions, until 6 months after your purchase date. Your receipt has a license key. Enter it at overqual.com/download for the three PDFs, stamped with your email and order number. They are yours to keep.