AIP-C01 sample questions with answers

1. A.

   Reranking is available only through the console and the AWS Management Console session is therefore the only context in which a reranker model can be applied to knowledge base results, so the service must replicate the reordering in application code.

2. B.

   The reranker model must be enabled per knowledge base rather than per caller, and the engineer enabled it for the console session only, so the service will succeed once the same model is attached to the knowledge base resource itself.

3. C.

   Reranked requests exceed the maximum request size for the Retrieve operation because the candidate documents are echoed back into the request, and the service is receiving an access denied error as the generic response to an oversized request body.

4. D.

   The service's execution role has permission to retrieve from the knowledge base but not to invoke the reranker model, which is a separate model invocation that requires its own permission.

1. A.

   Replicate the asset-tracking database into Amazon Aurora PostgreSQL with the pgvector extension and let the knowledge base read directly from that replica, so that a webhook-driven write becomes visible to retrieval as soon as the replica applies it.

2. B.

   Expose the asset-tracking REST API through the Web Crawler data source connector, pointing the crawler at the API's JSON endpoints and setting the crawl frequency high enough that changed records are picked up within the freshness window the team has committed to.

3. C.

   Have the webhook write each changed record into an Amazon S3 bucket configured as the knowledge base data source, then start an ingestion job on a short schedule so that the bucket and the knowledge base stay close enough together for the team's freshness target.

4. D.

   Connect a custom data source to the knowledge base and have the webhook handler submit each changed record through IngestKnowledgeBaseDocuments, which indexes it into the vector store in a single action with no sync step and no intermediate copy of the record staged in Amazon S3 beforehand.

Choose 2.

1. A.

   Embed the entire conversation transcript as the retrieval query on every turn, so that all prior context contributes to the query vector and no rewriting logic has to be maintained by the application.

2. B.

   Carry the resolved entities and filters of the thread in server-side session state and reapply them to each turn's retrieval request, so a follow-up inherits the scope the earlier turns established.

3. C.

   Increase the maximum number of source chunks returned for follow-up turns so that the correct passage is statistically more likely to appear somewhere in a larger candidate set even though the query vector itself remains ambiguous.

4. D.

   Lower the generation model's temperature on follow-up turns so that it is less likely to introduce subjects that were not present in the retrieved passages for that turn.

5. E.

   Disable hybrid search on follow-up turns, because lexical matching on very short queries such as why amplifies the noise that a dense vector search would otherwise smooth out across the candidate set.

6. F.

   Rewrite each follow-up into a self-contained standalone query using the conversation history before it is sent to retrieval, so that the text being embedded carries the subject the user omitted.

1. A.

   The prompt is the default system prompt, which includes generic example content as a static template provided by the service; it is not from another customer's data, and it can be overridden by supplying a custom text prompt template.

2. B.

   The examples are retrieved chunks from a data source that was attached to the knowledge base earlier and has since been removed, and they persist because the data deletion policy was left at RETAIN when that data source was deleted from the knowledge base.

3. C.

   The examples were generated by the orchestration step, which invents plausible question and answer pairs to calibrate the search query before retrieval, so they reflect this knowledge base's own corpus even though the subject matter looks unfamiliar to the analyst.

4. D.

   Model invocation logging captures prompts from every knowledge base in the account in a shared log stream, so the unfamiliar examples belong to a different team in the same organisation and the incident should be scoped to that team's knowledge base rather than this one.

1. A.

   Call the Rerank operation directly through the Amazon Bedrock API, passing the user query and the merged list of documents, and use the reordered list the model returns.

2. B.

   Register the two stores as data sources on an Amazon Bedrock knowledge base so that reranking applies to them, since a knowledge base can federate queries to external stores and rerank the combined result set on the team's behalf.

3. C.

   Nothing directly, because reranker models in Amazon Bedrock are available only as a configuration on Retrieve and RetrieveAndGenerate, so the team must first migrate its corpus into a knowledge base before any reranking capability becomes available to it.

4. D.

   Invoke the reranker model through the standard model invocation API with the merged list serialised into the prompt, because reranker models are ordinary text models that return a ranked list when asked to in natural language rather than being exposed through any dedicated operation.

1. A.

   Build a post-retrieval filter in the application that calls the source repository for each returned passage, checks the caller against the document's permissions, and discards passages the caller may not see before assembling the prompt for the generation model.

2. B.

   Encrypt each document with a different AWS KMS key and grant users decrypt permission only for the documents they may read.

3. C.

   Use a managed knowledge base with access control list awareness enabled, so the source repository's permissions are carried into retrieval and results are scoped to the calling user.

4. D.

   Create one knowledge base per permission group, sync each document into every knowledge base whose group may read it, and route each user's query to the knowledge base that matches the permission group carried in their identity token at sign-in.

1. A.

   Warmup calls are only meaningful on collections whose refresh interval is below one second, and because a Classic collection refreshes far more slowly than that the design's peak-preparation step would have no measurable effect on the first queries of the day.

2. B.

   Classic vector search collections cannot hold more than a few hundred million vectors, so a corpus of this size has to be sharded by the application across several collections with a routing layer in front of them, which is the part the design is missing.

3. C.

   Classic vector search collections support only the HNSW algorithm with Faiss and do not support IVF or IVFQ, and they do not support the warmup, stats or model training API operations.

4. D.

   Product quantization is unavailable in every AWS vector store, so the design must fall back to full 32-bit floating-point vectors.

1. A.

   The data source was configured with a data deletion policy of RETAIN, which keeps the vector embeddings in the vector store when the data source is deleted; a policy of DELETE removes them.

2. B.

   The knowledge base keeps a second copy of every vector for source attribution, and that copy is retained independently of the policy.

3. C.

   Vectors are only removed by an ingestion job, so the team should have run one final sync after emptying the bucket and before deleting the data source, because a deleted data source can no longer be synced and its vectors are therefore orphaned in the store forever.

4. D.

   Retrieval is answering from a cached copy of previous responses, and the team should clear that cache, because vectors belonging to a deleted data source are always removed from the vector store immediately when the data source itself is deleted.

1. A.

   A completed ingestion job does not guarantee immediate queryability: for vector stores other than Aurora it can take a few minutes for newly synced embeddings to become available, so the test needs to poll with a bounded wait rather than assert once.

2. B.

   The assertion is testing the wrong thing, because retrieval always returns the previously indexed version of a document until the next ingestion job completes, which means any pipeline that publishes and tests in one run will be one version behind by design.

3. C.

   The OpenSearch Serverless collection is under-provisioned and is shedding write requests under load, so the pipeline should be changed to retry the ingestion job whenever the retrieval assertion fails, since a failed assertion indicates that some documents were never written.

4. D.

   The knowledge base is applying hierarchical chunking, so the new wording exists only inside child chunks that are replaced by parent chunks at retrieval time, and the assertion should be rewritten to look for the parent chunk's summary text instead of the published wording.

Choose 2.

1. A.

   It allows the customer to choose the search strategy per request from semantic, hybrid or a customer supplied ranking function, which is the main relevance advantage it holds over the customer-managed option where the strategy is fixed when the knowledge base is created.

2. B.

   It includes a built-in managed semantic reranker optimised for accuracy and performance at no extra cost, while a customer-managed knowledge base has no managed reranking.

3. C.

   It supports agentic retrieval, and its search behaviour is described as agentic and semantic hybrid retrieval optimised across ingested file types, rather than a search strategy the customer selects.

4. D.

   It requires the customer to provision and scale the underlying vector and text data stores, while handling only parsing and chunking on the customer's behalf so that relevance tuning remains a customer responsibility.

5. E.

   It supports only the Amazon S3 and custom data source types, which is why teams needing wider connectivity are directed to build a customer-managed knowledge base instead.

6. F.

   It removes the ability to apply a reranker model of the customer's own choosing, because the managed reranker is the only reranking path available once a managed knowledge base has been created.