Task 1.3: Implement data validation and processing pipelines for FM consumption.

• AWrite a Lambda function that is triggered by each new S3 object and checks every record before the knowledge base ingests it

  Incorrect

  A Lambda function can validate files, but it is custom code without declarative rules or a quality score, which the engineers want to avoid.

• BAdd a DQDL ruleset with AWS Glue Data Quality to the ETL job, fail it without loading the target, and publish the results

  Correct

  Glue Data Quality evaluates DQDL rules inside the ETL job, reports a data quality score, can fail the job without loading data, and publishes metrics to CloudWatch and EventBridge for alerting.

• CCall Amazon Comprehend on each description during the job and drop records whose entity list is empty

  Incorrect

  Comprehend extracts entities from text, which is useful for enrichment, but it is not a rule engine for empty fields or duplicate keys and adds per-record cost.

• DRun Amazon Macie discovery jobs on the landing bucket and stop the ingestion pipeline whenever Macie reports new findings

  Incorrect

  Macie discovers sensitive data such as PII in S3. It does not check completeness or uniqueness of business fields.

Validation belongs before the data reaches the vector store. Glue Data Quality gives declarative DQDL rules, rule recommendations, a score per run, configurable job failure behavior and CloudWatch and EventBridge integration, so bad supplier files never reach the knowledge base.

• AAn Amazon Transcribe batch job that runs on the recording every five minutes during the call

  Incorrect

  Batch transcription works on stored files, so suggestions would trail the conversation by minutes instead of seconds.

• BAn Amazon Bedrock Data Automation project that processes each call recording after the call ends

  Incorrect

  Data Automation is a strong choice for extracting insights from stored audio, but processing after the call ends cannot help the agent during the conversation.

• CAmazon Transcribe streaming transcription with speaker partitioning and PII redaction enabled

  Correct

  Streaming transcription returns text while the call is in progress, speaker partitioning labels each speaker, and streaming PII redaction replaces identified PII with a [PII] tag before the text reaches the model.

• DAmazon Comprehend DetectPiiEntities on the raw audio stream before it is sent to the model

  Incorrect

  Comprehend analyzes text, not audio. The speech must be transcribed first, and speaker labels would still be missing.

Real-time requirements rule out post-call and batch processing. Transcribe streaming covers all three needs in one step: low-latency text, speaker labels and redaction of PII before the transcript is sent to the model.

• AWrite JSONL lines with a recordId and a Converse-format modelInput, and set the job's model invocation type to Converse

  Correct

  Each line of a batch input file has a recordId and a modelInput. With the Converse invocation type, modelInput uses the Converse request body, and the recordId lets results be matched even though output order is not guaranteed.

• BWrite JSONL lines with only the provider-specific InvokeModel body and rely on line numbers to match the output

  Incorrect

  The InvokeModel format is allowed, but it would not reuse the Converse code, and line numbers are unreliable because the output order can differ from the input order.

• CWrite one JSON array that holds all Converse request bodies in exactly the same order as the transcripts in the database

  Incorrect

  Batch inference expects JSONL with one record per line, and output order is not guaranteed to match input order, so position cannot be used for matching.

• DWrite CSV rows with a transcript column and a prompt column, and let batch inference build the request bodies

  Incorrect

  Batch inference does not accept CSV input. Records must be JSON objects in JSONL files with a modelInput field.

Batch records are JSONL objects with recordId and modelInput, and the job's invocation type decides whether modelInput follows the InvokeModel body of the model or the Converse request format. Always carry your own recordId, because results are not returned in input order.

• AA Lambda function that converts dates to ISO 8601 and units to one system with fixed rules before the prompt is built

  Correct

  Known, fixed conversion rules are best applied deterministically in code. A Lambda function gives the same output for the same input, can be audited, and costs far less than a model call.

• BAn Amazon Comprehend DetectEntities call that tags each date and quantity in the report

  Incorrect

  Entity detection finds and labels dates and quantities, but it does not convert them to a single format or unit.

• CA higher temperature setting so that the model considers more ways to read each date and unit

  Incorrect

  A higher temperature makes the output more random, which works against both accuracy and reproducibility.

• DA first Amazon Bedrock call that rewrites each report into a consistent format before the separate summarization call

  Incorrect

  A model can reformat text and is useful when the rules are fuzzy, but its output is not guaranteed to be identical across runs and it adds the cost of a second inference.

Improving input quality does not always need a model. When the rules are fixed and results must be reproducible and cheap, deterministic preprocessing in Lambda is the right tool; a model-based rewrite fits fuzzy cleanup where some variation is acceptable.

• AAnalyzeExpense, built for invoices and receipts

  Incorrect

  AnalyzeExpense targets invoices and receipts, not loan applications.

• BDetectDocumentText, which returns lines and words

  Incorrect

  Text detection returns lines and words without linking labels to their values.

• CAnalyzeDocument with only the TABLES feature type

  Incorrect

  Tables extracts table structure, not labeled form fields.

• DAnalyzeDocument with the FORMS feature type

  Correct

  Textract extracts form data as key-value pairs when you request the FORMS feature type.

Choose the Textract feature by document structure: FORMS for labeled fields, TABLES for grids, and expense analysis for invoices and receipts.

• ABuild an Amazon Lex V2 bot that processes each recording

  Incorrect

  Lex builds conversational voice and text interfaces for live interactions. It is not a bulk transcription service for archived recordings.

• BRun Amazon Transcribe batch jobs with speaker diarization enabled

  Correct

  Transcribe processes media files stored in Amazon S3 as batch jobs, and speaker diarization labels which speaker said each part of the transcript.

• CRun Amazon Transcribe Medical batch jobs on the recordings

  Incorrect

  Transcribe Medical is designed for medical speech such as physician-dictated notes. Customer support calls are not its domain.

• DSend each recording to Amazon Nova 2 Sonic for transcription

  Incorrect

  Nova 2 Sonic is a speech-to-speech model for real-time spoken conversations. It does not produce batch transcripts of archived files.

Stored recordings call for batch transcription, and the record's requirements decide the features: diarization when the transcript must say who spoke.

• AData Wrangler in SageMaker Studio Classic

  Incorrect

  Studio Classic is not available for new onboarding, and Data Wrangler has moved into Canvas.

• BAmazon Comprehend topic modeling

  Incorrect

  Topic modeling is not available to new customers and does not clean tabular data.

• CData preparation in SageMaker Canvas

  Correct

  Data Wrangler capabilities are now in SageMaker Canvas, which supports visual and natural language data preparation.

• DA SageMaker AI real-time endpoint

  Incorrect

  Endpoints serve model inference, not data preparation.

For low-code data preparation, use the current Canvas experience instead of the Studio Classic tools that are closed to new onboarding.

• AS3 event notifications that invoke a Lambda function

  Correct

  Amazon S3 can invoke a Lambda function when objects are created, which suits short, per-file processing.

• BA nightly AWS Glue crawler on the bucket

  Incorrect

  A crawler catalogs data. It does not convert documents to text.

• CAn EC2 instance that lists the bucket every minute

  Incorrect

  Polling on a server adds cost and management for irregular uploads.

• DA SageMaker AI endpoint that the partner calls directly

  Incorrect

  Partners would need to change their process, and an endpoint is not an event trigger.

Event-driven processing with S3 notifications and Lambda fits short, irregular jobs without servers to manage.

• AS3 event notifications that invoke a Lambda function to check each new object and delete bad records

  Incorrect

  S3 can invoke a Lambda function when objects are created, which suits per-file processing. Here the bad records would already sit in the ingestion prefix, and deleting them loses them for review.

• BA Firehose data transformation Lambda function that returns rejected records as ProcessingFailed

  Correct

  Firehose treats records returned as ProcessingFailed as unsuccessfully processed and delivers them to the processing-failed folder with error details, while valid records continue to the destination. The function can log the rule each rejected record broke to CloudWatch Logs.

• CFirehose dynamic partitioning on the app ID so that each app writes to its own prefix

  Incorrect

  Dynamic partitioning groups streaming data by keys within the records. It organizes records but does not check them.

• DFirehose record format conversion to Apache Parquet with a schema from an AWS Glue table

  Incorrect

  Format conversion turns JSON into columnar files for analytics. It is not a rules check, and Parquet is not among the file formats a knowledge base ingests.

Validate streaming data in the stream itself: a transformation function can pass, drop or fail each record, and Firehose keeps failed records with their error details.