Task 2.5: Implement application integration patterns and development tools.

• AA knowledge base that ingests the invoices so that users can ask questions about each one

  Incorrect

  A knowledge base answers questions over documents. It does not produce a structured record of 14 normalized fields per invoice.

• BAn Amazon Bedrock Data Automation project with a custom blueprint for the invoice fields and rules

  Correct

  Blueprints list the fields to extract with their data types and natural-language context that specifies normalization and validation, across documents and images of varying layouts.

• CAmazon Comprehend custom entity recognition trained on labeled invoices from each supplier

  Incorrect

  Custom entity recognition finds entities in text but needs labeled training data and does not derive values such as whether a late fee applies.

• DAmazon Textract AnalyzeExpense with a Lambda function that maps its output and computes the derived fields

  Incorrect

  AnalyzeExpense extracts standard invoice fields well, but the normalization and derived fields would be custom code, which the business team wants to avoid.

Intelligent document processing with custom, normalized and derived fields is what Bedrock Data Automation blueprints are designed for: fields, types and natural-language instructions in one artifact, applied to many layouts.

• AAn S3 event notification on each object that invokes a Lambda function to process the file

  Incorrect

  Events fire only when objects are written, not for an existing archive, and there is no concurrency cap or visual per-item tracking.

• BAn AWS Step Functions Map state in Inline mode that processes all objects in a single execution

  Incorrect

  Inline mode runs inside one execution with limits on concurrency and history size that do not suit hundreds of thousands of items.

• CAn AWS Step Functions Distributed Map state over the S3 objects with MaxConcurrency set to 100

  Correct

  A Distributed Map processes large sets of S3 objects as child workflow executions, MaxConcurrency limits how many run in parallel, and each item's success or failure is tracked separately.

• DAn AWS Step Functions Distributed Map state without a concurrency limit so that the run finishes fastest

  Incorrect

  Without a limit the map can run up to 10,000 child executions at once, far above the model quota, which leads to throttling.

Document processing at archive scale is a Distributed Map job. Its MaxConcurrency setting is how you respect a downstream quota such as model concurrency, and child executions give per-item status and retries.

• AExport the logs to Amazon S3 every night and build an AWS Glue job and an Amazon Athena table for the analysis

  Incorrect

  This works, but it is a new pipeline with a day of delay, when Logs Insights can answer the question directly.

• BCreate a CloudTrail trail with data events for Amazon Bedrock and search it for InvokeModel calls

  Incorrect

  CloudTrail records API calls for auditing, not the model's responses and stop reasons.

• CQuery the invocation log group with CloudWatch Logs Insights, grouping max_tokens stops by template version

  Correct

  Invocation logs carry the request metadata and the response, and Logs Insights can filter and aggregate those fields in place without a new pipeline.

• DTurn on AWS X-Ray active tracing for the Lambda function and inspect the trace map for slow segments

  Incorrect

  X-Ray shows where time is spent across services, but it does not aggregate model stop reasons by prompt template version.

Prompt troubleshooting starts with the invocation logs. Tagging requests with requestMetadata, such as template version, makes them filterable, and CloudWatch Logs Insights turns the logs into answers without moving data.

• ACall CountTokens with the Converse-format messages and summarize the oldest turns when the count passes a threshold

  Correct

  CountTokens returns the token count the model would use for the same input, at no charge, so the function can record it and shrink the conversation before it is too large.

• BSet maxTokens to the size of the model's context window so that every request is guaranteed to fit

  Incorrect

  maxTokens limits the generated output, not the input, and a very large value reserves more quota at the start of each request.

• CAdd an API Gateway request validator with a JSON schema that rejects bodies above a fixed size in bytes

  Incorrect

  Request validation checks the request against a schema, which is useful, but bytes do not map to tokens, and rejecting the request does not summarize the conversation.

• DEstimate the size as characters divided by four and drop the newest messages until the estimate fits the window

  Incorrect

  Tokenization is model-specific, so a character ratio is only a rough guess, and dropping the newest messages removes the user's current question.

Token limits are managed in the API layer. CountTokens gives the model-specific input count for free, which supports both cost records and context management such as summarizing older turns before the request is sent.

• AA Lambda function that reads the OAuth secret from Secrets Manager and calls the CRM

  Incorrect

  A function fits calls that need custom logic. It is the code the team does not want to maintain.

• BAn EventBridge rule with an API destination, triggered by an event from the workflow

  Incorrect

  API destinations call HTTPS endpoints as targets of rules, which suits event delivery. The CRM's response does not come back to the workflow, so the note ID is lost.

• CAn HTTP Task with an EventBridge connection for OAuth and a Retry on the 429 status

  Correct

  An HTTP Task calls third-party HTTPS APIs from the workflow, and the API response becomes the state output for the next state. Its EventBridge connection holds the OAuth credentials, and a Retry can match States.Http.StatusCode.429.

• DAn Amazon SQS queue that the CRM reads new summaries from

  Incorrect

  A queue decouples a producer from a consumer that can poll it. The SaaS CRM only offers its REST API.

Workflows can call SaaS APIs directly: HTTP Tasks with connections remove glue code and keep credentials out of the definition.

• AChain extract, compare and draft prompts as separate Step Functions tasks

  Correct

  Prompt chaining splits a task into linked LLM calls that pass intermediate results, and Step Functions can orchestrate them. Each task's output can be checked and saved, and Retry applies to the failed state only.

• BAdd five complete contract and email examples to the single prompt

  Incorrect

  Few-shot examples help a model calibrate its output. It is still one call, with no stage outputs to inspect and no stage-level retry.

• CFine-tune the model on 3,000 past contracts paired with the emails sent

  Incorrect

  Fine-tuning learns a task from labeled examples. It changes the model, not the single opaque call.

• DAsk the model to reason step by step and then write the email in one response

  Incorrect

  Step-by-step reasoning helps multi-step deductions within one response. A failure still means repeating the whole call.

Prompt chaining turns one opaque call into steps you can test, audit and retry separately.

• AEnable Amazon Bedrock model invocation logging for the account

  Incorrect

  Invocation logging collects the full request and response data of model calls, which suits reviewing prompts and outputs. It does not time the other calls in the request.

• BTurn on X-Ray tracing for the API stage, function and SDK clients

  Correct

  X-Ray traces each request through API Gateway and Lambda and shows the calls the application makes to downstream AWS resources, building a service map and per-request segments.

• CTurn on CloudWatch Lambda Insights for the Lambda function

  Incorrect

  Lambda Insights collects system-level metrics such as CPU time, memory and cold starts, which suits resource problems inside the function. It does not break down downstream calls.

• DCreate a CloudTrail trail that records the Bedrock API calls

  Incorrect

  CloudTrail records API activity, such as who called which operation and when, for auditing. It does not show where the time of each request went.

Latency questions across service boundaries need distributed tracing. Logs and metrics tell you that something is slow; traces tell you where.

• AMove the API to an HTTP API with a longer integration timeout

  Incorrect

  HTTP APIs are lighter and cheaper, but their maximum integration timeout is 30 seconds, and usage plans with API keys are a REST API feature.

• BRequest a higher integration timeout for the current edge-optimized API

  Incorrect

  The integration timeout can be raised above 29 seconds for Regional and private REST APIs, not for edge-optimized ones.

• CPut an Application Load Balancer in front of the Lambda function

  Incorrect

  A load balancer waits for a Lambda target until it responds or times out, which suits long calls. It offers no usage plans or API keys.

• DSwitch the API to a Regional endpoint and raise its integration timeout

  Correct

  An edge-optimized API can be changed to Regional, and Regional REST APIs can raise the integration timeout above 29 seconds, which may lower the Region's throttle quota, acceptable at this volume.

Long synchronous calls are possible on REST APIs, but the endpoint type and the throttle trade-off decide whether the timeout can move.

• AStart a new execution for each pack with the original input

  Incorrect

  A new execution gets a new ID and runs every state again, including the eleven model calls.

• BRedrive each failed execution so that it resumes at the upload state

  Correct

  Redrive continues a failed Standard execution from the unsuccessful step, within 14 days, with the same input, definition and execution ARN. Successful steps are not rerun, and the redrive events are appended to the original history.

• CRun the upload state alone with the TestState API and the failed input

  Incorrect

  TestState runs a state definition in isolation for testing. The original executions would stay failed, and their histories would not show the upload.

• DAdd a Catch to the upload state that waits an hour and tries again

  Incorrect

  A Catch sends errors to a fallback state in later executions. It changes the definition and does nothing for the executions that already failed.

To finish a failed Standard execution without repeating its completed work, redrive it from the failed step.