Foundation Model Integration, Data Management, and Compliance — AIP-C01 Practice Quiz

Question 1

A financial advisory startup builds a customer support assistant using Amazon Bedrock Agents. Users interact across multiple sessions over several days, and the business requires the agent to remember key facts from earlier conversations (such as a customer's stated risk tolerance) so that follow-up sessions feel continuous. The team already uses session state for within-conversation context, but that context is lost once a session ends. Which approach most directly satisfies the requirement with the least custom development?

Accepted Answer

Enable Amazon Bedrock Agents memory retention, which summarizes and persists session information across multiple sessions for a configured retention period

Answer

Increase the foundation model's context window and pass the full transcript of every prior session in each new request

Answer

Store each session's raw messages in an S3 bucket and re-ingest them into the agent's knowledge base after every conversation

Answer

Rely on the agent's default session state, which automatically carries over across separate user sessions

Question 2

A financial services company runs a customer-facing chatbot on Amazon Bedrock using Anthropic Claude. During market-open hours the traffic spikes predictably to about 12x baseline for roughly 90 minutes, and users experience throttling errors (ThrottlingException) because the on-demand account-level quota is exceeded. Outside these windows, traffic is low and steady. The team wants to eliminate throttling during the spike while minimizing cost the rest of the day. What is the most appropriate approach?

Accepted Answer

Purchase Provisioned Throughput with a no-commitment (hourly) term, activate the model units before the market-open window, and release them after the spike subsides while serving off-peak traffic on-demand.

Answer

Purchase Provisioned Throughput with a monthly commitment term sized to the peak spike, and route all traffic through it at all times.

Answer

Implement client-side exponential backoff with jitter and rely solely on on-demand throughput, accepting retries during the spike.

Answer

Switch to a smaller model during the market-open window to fit within the existing on-demand quota, then switch back afterward.

Question 3

A financial services company is building a customer support assistant with Amazon Bedrock Agents. When a customer asks to cancel a pending wire transfer, the agent must call the bank's internal transfer-cancellation microservice. For security and compliance reasons, this microservice runs inside a private on-premises data center that is not reachable from AWS Lambda or any AWS-hosted compute, and the security team refuses to expose it via any public or VPC-connected endpoint that Bedrock could invoke. The team still wants the agent to orchestrate the conversation, determine when cancellation is needed, and extract the correct parameters. How should the developer configure the Bedrock Agent action group to satisfy these constraints?

Accepted Answer

Configure the action group to use return of control (RETURN_CONTROL executor), so Bedrock returns the identified action and extracted parameters to the calling application, which then invokes the on-premises microservice itself.

Answer

Configure the action group with a Lambda function executor and use AWS PrivateLink to reach the on-premises microservice, since Bedrock requires a Lambda for all action groups.

Answer

Attach an OpenAPI schema pointing directly at the on-premises microservice URL so the agent can call it over the internet after whitelisting Bedrock's IP ranges.

Answer

Use a Bedrock Knowledge Base to store the cancellation API definition and let the agent retrieve and execute it at runtime through RAG.

Question 4

A media analytics company needs to generate summaries for a backlog of 4 million archived news articles using Amazon Bedrock. There is no user-facing latency requirement — the job can run overnight or over several days. The team wants to minimize cost per token while processing this large, static dataset. Which approach best meets these requirements?

Accepted Answer

Use the Bedrock batch inference (CreateModelInvocationJob) feature, providing input records in Amazon S3 and retrieving results from S3 after the job completes.

Answer

Invoke the InvokeModel API in a parallelized loop with hundreds of concurrent threads to process articles as fast as possible.

Answer

Purchase Provisioned Throughput with a long-term commitment and stream each article through the model in real time.

Answer

Use the InvokeModelWithResponseStream API so results return incrementally, reducing the total tokens billed.

Question 5

A European healthcare analytics company runs a generative AI application on Amazon Bedrock using Anthropic Claude models. Under GDPR and internal policy, all prompt and completion data containing patient information must never be processed or stored outside the EU. During a load test, the team enabled cross-region inference profiles to improve throughput and reduce throttling. Their compliance officer flags a concern about where inference requests are being routed. Which action best ensures the solution meets the data residency requirement while still addressing throughput needs?

Accepted Answer

Use an EU-scoped cross-region inference profile so requests are only routed among EU Regions, and request a service quota increase for on-demand model throughput in those Regions.

Answer

Keep the cross-region inference profile enabled, since Bedrock automatically encrypts all data in transit and at rest, which satisfies GDPR residency obligations.

Answer

Disable cross-region inference and instead route overflow traffic to a US Region during peak load, applying client-side PII masking before each request.

Answer

Switch to provisioned throughput in a single EU Region only, because provisioned throughput is the only Bedrock feature that guarantees data never leaves a Region.

Question 6

A data engineering team is preparing a proprietary corpus of 2 million customer support transcripts to fine-tune a foundation model on Amazon Bedrock. During exploratory analysis they discover that roughly 30% of the transcripts are near-duplicates (repeated boilerplate greetings, canned responses, and copy-pasted templates), and a subset contains full credit card numbers embedded in free-text fields. The team wants the fine-tuned model to generalize well and wants to minimize compliance risk before the customization job runs. Which data preparation approach best addresses both concerns?

Accepted Answer

Deduplicate the transcripts before training and apply PII detection/redaction (e.g., Amazon Comprehend or Macie-driven workflows) to remove or mask credit card numbers, then use the cleaned corpus for the customization job.

Answer

Run the fine-tuning job on the raw dataset and rely on Bedrock Guardrails at inference time to block credit card numbers in outputs, since deduplication is handled automatically by the training service.

Answer

Keep all near-duplicates because higher record volume always improves fine-tuning quality, and encrypt the S3 bucket with SSE-KMS so the credit card numbers are protected at rest during training.

Answer

Split the dataset so credit card records are used only for validation rather than training, and leave the duplicates in the training split to preserve realistic frequency distributions.

Question 7

A European healthcare company must build a clinical document summarization application using Amazon Bedrock. Regulatory requirements mandate that all patient data, including inference requests and responses, must never leave the EU. The team wants to use Anthropic Claude models but discovers their preferred model version is not yet available in any EU region. Compliance has explicitly rejected any cross-region inference that could route data outside the EU. Which approach best satisfies both the functional and compliance requirements?

Accepted Answer

Select a foundation model that is available within an EU Bedrock region (such as an EU-hosted Claude or alternative model), even if it is not the exact preferred version, and deploy the workload entirely in that region

Answer

Use the model in a US region and rely on Bedrock Guardrails to redact PII before the request leaves the EU, ensuring no identifiable data crosses the boundary

Answer

Enable cross-region inference profiles restricted to the EU-US corridor so Bedrock automatically balances load while keeping data in the EU whenever possible

Answer

Deploy the preferred model version by importing it via Bedrock Custom Model Import into a US region and connect to it over a VPC endpoint with PrivateLink

Question 8

A retail company is building a semantic product search system using Amazon Bedrock. Their engineering team generates embeddings for 2 million product descriptions using Amazon Titan Text Embeddings and stores them in Amazon OpenSearch Serverless. During testing, they observe that search relevance is inconsistent: some semantically similar products score poorly while unrelated products with longer descriptions rank higher. The team configured the OpenSearch k-NN index to use the L2 (Euclidean) distance metric. What is the MOST likely cause and appropriate fix?

Accepted Answer

The embeddings are not normalized, so L2 distance is biased by vector magnitude; switch the k-NN index to use the cosine similarity metric or normalize vectors before indexing.

Answer

The Titan embedding model dimension is too small; upgrade to a higher-dimensional embedding model to improve relevance.

Answer

OpenSearch Serverless does not support k-NN search reliably; migrate the vector store to Amazon Aurora with pgvector.

Answer

The product descriptions were not chunked; split each description into smaller chunks before generating embeddings.

Question 9

A healthcare startup uses Amazon Bedrock with Anthropic Claude to power a patient-facing chatbot. Compliance requires that any Social Security numbers or phone numbers entered by users are never sent to the model or stored in logs, and that the model refuses to produce medical dosage recommendations. The team wants a managed, model-agnostic mechanism that enforces these rules on both input prompts and model output without writing custom regex-based filtering code in the application layer. Which approach best meets these requirements?

Accepted Answer

Configure an Amazon Bedrock Guardrail with sensitive information filters (PII masking/blocking) plus denied topics, and attach the guardrail to the InvokeModel calls.

Answer

Enable Amazon Bedrock model invocation logging to CloudWatch and use a Lambda subscription filter to scrub PII after each request completes.

Answer

Use Amazon Comprehend PII detection in the application before calling Bedrock, and rely on the model's system prompt to refuse dosage questions.

Answer

Fine-tune the Claude model on a redacted dataset so it learns to ignore SSNs and phone numbers and never produce dosage recommendations.

Question 10

A software company wants to build a RAG-based internal support assistant using Amazon Bedrock Knowledge Bases. Their engineering documentation lives in Atlassian Confluence, and content is updated frequently by multiple teams. The team wants to minimize custom code for ingestion, keep the knowledge base current with Confluence changes, and avoid building and maintaining an ETL pipeline to export documents into Amazon S3. What is the MOST operationally efficient approach?

Accepted Answer

Configure a native Confluence data source connector in the Bedrock Knowledge Base and schedule periodic syncs to ingest content directly.

Answer

Build an AWS Glue job that exports Confluence pages nightly to an S3 bucket, then configure the S3 bucket as the Knowledge Base data source.

Answer

Use a Lambda function triggered on a schedule to call the Confluence API, chunk documents, generate embeddings with Titan, and write vectors directly to OpenSearch Serverless.

Answer

Export Confluence content manually to PDF files, upload them to S3, and run a full ingestion job whenever documentation changes.

Question 11

A media company uses an Amazon Bedrock Knowledge Base backed by an Amazon S3 data source containing 50,000 documents. During a scheduled ingestion job, the sync completes but the team notices that roughly 300 documents were not indexed. They need a reliable way to identify exactly which documents failed and why, without re-processing the entire corpus. What is the most effective approach?

Accepted Answer

Review the ingestion job's data source sync status and inspect the per-document failure details and reasons emitted to the configured Amazon CloudWatch/S3 logs, then correct and re-sync only the affected documents.

Answer

Delete the entire Knowledge Base and recreate it from scratch, since partial ingestion failures cannot be diagnosed at the document level.

Answer

Increase the provisioned throughput of the embedding model so all 50,000 documents are processed together in a single batch.

Answer

Enable hybrid search on the vector store so that documents that failed dense embedding are still retrievable through keyword matching.

Question 12

A developer builds a Bedrock Knowledge Base using Amazon Titan Text Embeddings V2 configured to output 1024-dimensional vectors, stored in an Amazon OpenSearch Serverless vector index. After several months, the team decides to switch the embedding model to Cohere Embed English, which produces 1024-dimensional vectors but uses a different embedding space and similarity distribution. What must the team do to ensure retrieval quality remains correct after the switch?

Accepted Answer

Re-ingest and re-embed the entire source corpus with the new model into a new (or recreated) vector index, since query and document embeddings must come from the same model to be comparable.

Answer

Simply update the embedding model in the Knowledge Base configuration; because both models output 1024 dimensions, the existing index and vectors remain fully compatible.

Answer

Keep the existing document vectors but only re-embed new queries with Cohere, relying on OpenSearch to normalize the vector spaces automatically.

Answer

Add a second vector field to the same index for Cohere embeddings and let Bedrock blend results from both embedding models during retrieval.

Question 13

A financial services company built a Bedrock Knowledge Base over regulatory filings. Analysts complain that queries containing exact regulatory codes (e.g., 'Rule 15c3-3') and specific ticker symbols often return semantically related but incorrect documents, while conceptual questions like 'what are liquidity requirements' work well. The team wants to improve retrieval accuracy for both keyword-exact and conceptual queries without building a separate search system. What should they configure?

Accepted Answer

Change the search type on the Knowledge Base retrieval configuration to HYBRID so both semantic (dense vector) and keyword (sparse) matching are combined

Answer

Switch the embedding model to one with a higher dimension count to capture the exact tokens in regulatory codes

Answer

Reduce the chunk size so each regulatory code appears in its own isolated chunk for more precise vector matches

Answer

Increase the numberOfResults parameter and rely on the foundation model to filter out the incorrect documents during generation

Question 14

A financial services company runs a Bedrock Knowledge Base backed by an Amazon S3 data source and an OpenSearch Serverless vector store. Every night, a compliance team uploads a mix of brand-new policy PDFs, revised versions of existing PDFs (same object keys), and they delete a small number of retired policies from the S3 bucket. The developer needs the knowledge base to reflect all three types of changes with minimal cost and without re-embedding documents that have not changed. Which approach best meets these requirements?

Accepted Answer

Run a Bedrock Knowledge Base ingestion (sync) job on the data source after each nightly upload, relying on the service to add new files, re-embed modified files, and remove vectors for deleted objects.

Answer

Delete and recreate the entire knowledge base each night, then run a full ingestion job so the vector index always matches the current S3 contents.

Answer

Configure an S3 event notification to invoke a Lambda that calls the embedding model directly and writes vectors to OpenSearch, bypassing the knowledge base ingestion pipeline.

Answer

Schedule a nightly ingestion job but exclude deleted files, since the knowledge base cannot remove stale vectors and old policies must be manually purged from OpenSearch.

Question 15

A SaaS company is building a customer support assistant using Amazon Bedrock Knowledge Bases. Each of their 400 enterprise tenants has its own set of support documents that must never be visible to other tenants. The engineering team wants a single, cost-efficient architecture that avoids provisioning hundreds of separate vector indexes while still guaranteeing strict tenant isolation at query time. Which approach best meets these requirements?

Accepted Answer

Create one Bedrock Knowledge Base with a shared vector index, attach a tenant_id metadata attribute to each ingested document, and apply a metadata filter on tenant_id in every Retrieve request.

Answer

Provision a separate Bedrock Knowledge Base and dedicated OpenSearch Serverless collection for each tenant to physically isolate their data.

Answer

Store all tenant documents in one shared index without metadata and rely on a post-retrieval LLM prompt instruction telling the model to only use the current tenant's data.

Answer

Create one Knowledge Base and encrypt each tenant's documents with a different KMS key so only the correct tenant can decrypt retrieved chunks.

Question 16

A pharmaceutical company is building a research assistant on Amazon Bedrock Knowledge Bases. Analysts frequently ask questions that require connecting related entities across many documents — for example, 'Which compounds share a mechanism of action with the drug studied in trial X, and what were their adverse events?' With standard vector-based RAG, the assistant retrieves individually relevant chunks but consistently fails to synthesize answers that depend on multi-hop relationships between entities scattered across the corpus. Which approach should the team adopt to best improve answer quality for these relationship-heavy queries?

Accepted Answer

Configure the Bedrock Knowledge Base to use GraphRAG with Amazon Neptune Analytics as the vector store, so entities and their relationships are captured in a knowledge graph during ingestion

Answer

Increase the number of retrieved chunks (topK) to 100 and enlarge the chunk size so more context is passed to the foundation model in a single prompt

Answer

Switch the embedding model to a higher-dimension model to capture more semantic nuance between the compound documents

Answer

Enable semantic chunking so that each chunk aligns with a complete semantic unit before embedding

Question 17

A financial services company builds a single Amazon Bedrock Knowledge Base backed by Amazon OpenSearch Serverless. It ingests documents from three S3 buckets: one with public marketing content, one with internal HR policies, and one with confidential M&A deal documents restricted to a small deal team. All employees can query the same chatbot application, but each user must only receive answers derived from documents they are authorized to see. The team wants to avoid building and maintaining three separate Knowledge Bases. What is the most effective approach?

Accepted Answer

Attach metadata attributes (e.g., a classification/access-group field) to documents during ingestion and apply metadata filters at query time based on the authenticated user's entitlements when calling Retrieve or RetrieveAndGenerate.

Answer

Rely on the S3 bucket policies and IAM permissions of the source buckets; Bedrock Knowledge Bases automatically enforce the originating bucket's access controls at retrieval time.

Answer

Encrypt each S3 bucket with a different KMS key so that only users with access to the corresponding key can retrieve the associated vectors from OpenSearch Serverless.

Answer

Use a Bedrock Guardrail with a denied-topics policy to block the model from generating responses that reference confidential M&A or HR content for unauthorized users.

Question 18

A financial services company is building a RAG assistant using Amazon Bedrock Knowledge Bases over lengthy regulatory filings. During testing, they find that small chunks retrieve precisely relevant passages but the foundation model lacks enough surrounding context to generate coherent, complete answers. Larger chunks improve answer coherence but degrade retrieval precision because embeddings become diluted across many topics. They want to preserve precise retrieval while giving the model broader context at generation time. Which Bedrock Knowledge Bases chunking approach best addresses this?

Accepted Answer

Use hierarchical (parent-child) chunking so small child chunks are embedded and searched, while the larger parent chunk is returned to the model for generation.

Answer

Switch to fixed-size chunking with a very large token size and rely on metadata filtering to narrow results.

Answer

Disable chunking entirely (no chunking) so each document is embedded as a single vector for maximum context.

Answer

Use semantic chunking with a low breakpoint threshold to force many small chunks and increase the number of retrieved results.

Question 19

A legal services company is building a Bedrock Knowledge Base from a document repository in Amazon S3. The repository contains a mix of native text PDFs and scanned PDFs (image-based) of older contracts. After the initial ingestion and sync, users report that queries about clauses found only in the scanned contracts return no relevant results, while queries about native text PDFs work correctly. The team confirms the scanned PDFs synced without errors. What is the MOST appropriate way to make the scanned contract content retrievable?

Accepted Answer

Enable the foundation-model-based advanced parsing (parsing configuration) option in the Knowledge Base data source so it extracts text from image-based content, then resync.

Answer

Increase the vector dimension of the embedding model so it can capture the additional visual features in the scanned documents.

Answer

Switch the chunking strategy from fixed-size to semantic chunking so the scanned PDF paragraphs are split correctly.

Answer

Add metadata filtering to the query so scanned PDFs are prioritized in the retrieval results.

Question 20

A financial services team uses a Bedrock Knowledge Base with the RetrieveAndGenerate API to answer customer questions about mutual fund products. Compliance requires that every generated answer must cite the source document sections it used and must decline to answer when the retrieved passages do not contain sufficient information (rather than relying on the model's general knowledge). The default responses sometimes include information not present in the retrieved chunks and omit citations. Which approach best addresses these requirements while keeping the managed RAG workflow?

Accepted Answer

Customize the knowledge base's generation prompt template to instruct the model to answer only from the retrieved search results, decline when insufficient, and reference source passages, and enable citation output in the RetrieveAndGenerate response.

Answer

Switch the knowledge base to a smaller, cheaper foundation model so it has less general knowledge to draw from, reducing the chance of unsupported statements.

Answer

Lower the temperature and top-p inference parameters to zero so the model becomes fully deterministic and never hallucinates content outside the retrieved chunks.

Answer

Increase the number of retrieved results (numberOfResults) to the maximum so the model always has enough context and therefore never needs to decline or add outside information.

More AIP-C01 practice