Data Lineage and Source Citation

When an AI system makes a decision or writes an answer, people eventually ask where it came from. Which data trained this model? Who approved it, and for what use? Which document backs up this answer? Being able to reply is called provenance, and it matters for audits, for fixing problems and for user trust. The exam expects you to know the AWS tools that record it: the AWS Glue Data Catalog for data, SageMaker AI lineage tracking and Model Cards for models, and citations for generated answers.

Lineage and Provenance

Data lineage is the record of where data came from and what was done to it along the way: which sources fed a dataset, which jobs transformed it, and which models were trained on the result. Provenance is the broader idea of knowing an item's origin. Together they let a team answer an auditor's question with evidence instead of memory, trace a bad prediction back to a faulty data source, and retrain a model the same way later. Lineage is easiest to keep when tools record it automatically as work happens, rather than relying on people to write it down afterwards.

Cataloging Data

You can't trace data you can't find. The AWS Glue Data Catalog is a central store of metadata about your datasets: where each one lives, its schema and its properties. Crawlers scan sources such as Amazon S3, Amazon RDS and Amazon Redshift and fill in the catalog automatically, so it stays current as data changes. The catalog also keeps lineage information about the transformations applied to data, and it works with AWS Lake Formation for fine-grained access control. Services such as Amazon Athena and SageMaker AI can use the same catalog, so everyone works from one description of the data.

Tracking and Documenting Models

SageMaker AI ML Lineage Tracking records the steps of a machine learning workflow, from data preparation to deployment. It tracks training jobs, the datasets they used, pipelines, endpoints and the models themselves, so you can look up exactly which artifacts produced a model and reproduce it.

Amazon SageMaker Model Cards are the human-readable side. A model card documents a model in one place: its intended uses and the uses it isn't meant for, a risk rating (unknown, low, medium or high), training details and metrics, evaluation results, and notes such as ethical considerations. Every edit other than an approval-status change creates a new version, giving an unchanging history, and a card can be exported as a PDF for auditors. Model Cards also connect to the SageMaker Model Registry, where model versions are tracked.

Citing Sources in Answers

Generated answers need provenance too. When an application uses Retrieval Augmented Generation (RAG) through Amazon Bedrock Knowledge Bases, the response comes back with citations: each part of the answer is linked to the source chunks it drew on, with the location of the original document, such as its S3 path. An application can show these as footnotes so users can check the answer against the source. Citations don't make an answer correct on their own, but they make it checkable, which is what a regulated business usually needs.

On the Exam

Match the question to the record it needs:

A common trap is choosing a guardrail's contextual grounding check when users must see the source. The check can block unsupported answers, but it doesn't show the user where an answer came from.