Task 5.2: Troubleshoot GenAI applications.

• AA guardrail intervened in the output, so review the guardrail's content filter strengths

  Incorrect

  A guardrail intervention reports a different stop reason and replaces the output with the blocked message.

• BThe output reached the maxTokens limit, so raise it or ask for a shorter summary

  Correct

  A max_tokens stop reason means the model hit the output limit set in inferenceConfig, so the limit or the requested length must change.

• CThe input exceeded the context window, so split the contract into smaller chunks

  Incorrect

  The input is within the context window, and an oversized input is rejected with a validation error rather than cut off silently.

• DA stop sequence in the prompt ended generation, so remove all stop sequences

  Incorrect

  A stop sequence ends generation with a stop_sequence reason, not max_tokens.

Check stopReason first when output looks cut off. max_tokens means the output budget ran out; end_turn means the model finished; other values point to stop sequences, tool use or guardrail interventions.

• ARequest a higher cross-Region requests-per-minute quota for the model in us-east-1

  Incorrect

  Quota problems return throttling errors, not access denied errors.

• BEnable the destination Regions in the account settings so that requests can be routed to them

  Incorrect

  Cross-Region inference can route to Regions that are not manually enabled, so Region opt-in is not the cause.

• CReplace the inference profile with the global profile, which needs only the profile permission

  Incorrect

  The global profile also needs model permissions, and it would route requests worldwide, which changes the data residency posture.

• DAllow bedrock:InvokeModel on the model ARN in every destination Region of the profile

  Correct

  Requests through a geographic profile can be routed to any destination Region in it, so the role needs permission for the foundation model in each of those Regions as well as for the profile.

Intermittent AccessDeniedException after moving to cross-Region inference usually means the policy covers the source Region only. Grant the inference profile plus the foundation model in all of the profile's destination Regions, optionally conditioned on the profile ARN.

• AIncrease the SDK's maximum retry attempts so that a reset connection is retried more often

  Incorrect

  Retries eventually succeed, but each first call still waits for the dead connection to time out, so the delay remains.

• BPurchase Provisioned Throughput so that capacity stays warm during quiet periods

  Incorrect

  Bedrock reports no errors or throttles. The dropped connection is on the network path, not a capacity problem.

• CReplace the interface VPC endpoint with a NAT gateway so that connections use the public endpoint

  Incorrect

  NAT gateways have the same 350-second idle timeout, so the problem would remain.

• DEnable TCP keep-alive in the SDK and set the OS keep-alive time below 350 seconds

  Correct

  Interface endpoints, NAT gateways and Network Load Balancers drop idle TCP connections after 350 seconds. Keep-alive needs both the SDK setting and an OS interval shorter than that timeout.

Long-idle pooled connections are silently dropped by NAT gateways, VPC interface endpoints and NLBs after 350 seconds. Turning on SDK keep-alive alone is not enough, because the Linux default keep-alive interval is two hours; lower the kernel setting as well.

• AIncrease the number of results per query so that relevant documents are more likely to appear

  Incorrect

  More results from incompatible vectors are still poorly ranked, so this treats the symptom without fixing the mismatch.

• BSwitch the index to hybrid search so that keyword matching compensates for the new model

  Incorrect

  Keyword matching can mask part of the problem, but the vector half of the search would stay broken.

• CRe-embed the documents with the new model into a new index and query that index

  Correct

  Query and document vectors must come from the same embedding model to be comparable. Vectors from different models live in different spaces, even when the dimensions match.

• DAdd a reranker model after retrieval so that the best documents are moved to the top

  Incorrect

  A reranker can only reorder what retrieval returned, and mismatched vectors return the wrong candidates.

Embedding spaces are model-specific. Changing the query model without re-embedding the corpus breaks similarity search. In Bedrock knowledge bases, queries are converted with the same embedding model used during ingestion, and changing the embedding approach means creating a new knowledge base.

• ASend the iPhone photos as document content blocks instead of image blocks

  Incorrect

  Document blocks accept formats such as PDF, CSV, Word, Excel, HTML, text and Markdown, not photos, and the model must receive images.

• BDownscale the HEIC photos before sending them, keeping their original format

  Incorrect

  The small photos fail too, so size is not the cause. The format value itself is rejected.

• CConvert HEIC photos to JPEG or PNG and set the matching image format

  Correct

  The image block's format accepts only png, jpeg, gif and webp. Converting the photos and declaring a supported format fixes the validation error.

• DRetry the failed calls with exponential backoff and jitter in the SDK client

  Incorrect

  A validation error means the input breaks a constraint of the API. Retrying the same request fails the same way.

Validate request inputs against the API's allowed values before calling the model. A ValidationException that appears for one input type points to a request constraint, not to capacity or transient faults.

• AStart a new sync of the data source so that the metadata attributes are indexed again

  Incorrect

  The metadata is already ingested. Filtering requires a supported index engine, which a resync does not change.

• BCreate a faiss-engine vector index, attach it to a new knowledge base and ingest again

  Correct

  Metadata filtering on an OpenSearch Serverless vector store needs the index to use the faiss engine. With nmslib, AWS directs you to create a new index with faiss, or let Bedrock create one, and a new knowledge base.

• CCopy the department name from each metadata file into the document text before ingestion

  Incorrect

  Words in the text can influence similarity, but they do not enforce a filter, so results from other departments can still appear.

• DChange the space type of the vector field to cosine similarity and rebuild the index

  Incorrect

  The space type sets how vector distance is computed. It does not make the nmslib engine support metadata filtering.

Some retrieval features depend on how the vector store was configured, not on the query. When a feature silently does nothing, check the knowledge base's documented requirements for that vector store before tuning queries.

• AAdd five more billing examples so that the model learns the billing category's boundaries better

  Incorrect

  More billing examples make the example set even more one-sided, which pushes outputs further toward billing.

• BMove the six examples from the user message into the system prompt in an examples section

  Incorrect

  Examples can sit in either place, so moving them changes their position, not the imbalance that biases the outputs.

• CReplace the examples with ones that cover all nine categories, including common edge cases

  Correct

  Examples should represent the expected distribution of inputs and outputs, from common cases to edge cases, because biased examples lead to biased outputs.

• DRaise the maximum output tokens so that the model can reason before it names the category

  Incorrect

  More output room does not correct the skew that the examples create, and the output must remain a single category name.

Few-shot examples teach the model what the label distribution looks like. When one class dominates the examples, the model over-predicts it, so diagnose by per-class results and rebalance the examples before reaching for heavier fixes.

• ADrop minLength and check the length in code

  Correct

  String constraints such as minLength are not supported, and schemas with unsupported features return a 400 error.

• BRetry the same request with exponential backoff

  Incorrect

  The schema will be rejected again on every retry.

• CSwitch to a higher service tier

  Incorrect

  Service tiers do not change which schema features are supported.

• DRaise maxTokens for the request

  Incorrect

  Output length is unrelated to schema validation.

Structured outputs support a subset of JSON Schema. Move unsupported constraints into application-side validation.

• ATurn on model invocation logging and read the prompts and responses of the two prompt nodes

  Incorrect

  Invocation logs cover model calls only. They do not show the knowledge base node's output or which conditions the condition node evaluated as satisfied.

• BRecord CloudTrail data events for the flow alias and review each InvokeFlow event

  Incorrect

  Data events show who invoked the flow and when, not the inputs and outputs of the nodes inside it.

• CInsert a Lambda function node after every node that writes its output to CloudWatch Logs

  Incorrect

  This would change the flow definition, which is not allowed, and it adds code to maintain for a diagnosis the service already supports.

• DInvoke the flow with the failing inputs and enableTrace set to true, then read each trace

  Correct

  Flow traces return the input and output of each node and which conditions a condition node satisfied, so the node that produced the wrong intermediate result can be found without changing the flow.

Reasoning-path tracing turns a wrong final answer into a specific failing step. Reproduce with known failing inputs and follow the trace node by node before changing prompts or logic.