ExamPark
Course menuDEA-C01

Exams / DEA-C01 / Study guide / LLMs and vectors in data pipelines

LLMs and vectors in data pipelines: DEA-C01 study notes and practice questions

What to know about LLMs and vectors in data pipelines for the DEA-C01 exam, from the Data Ingestion and Transformation domain (34% of the exam guide), then 1 free question to test it. Part of the DEA-C01 study guide.

Last verified against current AWS documentation ·How we write our questions

What to know for the exam

These questions ask how pipelines feed generative AI applications, or use models to process data.

Preparing data for retrieval (RAG)

  1. Collect and clean documents (remove duplicates, extract text from PDFs with Amazon Textract if needed, strip boilerplate).
  2. Chunk them into passages (fixed size, by section, or semantic).
  3. Embed each chunk with an embeddings model (such as Amazon Titan Text Embeddings in Amazon Bedrock).
  4. Store vectors with metadata in a vector store.
  5. Keep it current: re-sync when documents change.

Amazon Bedrock Knowledge Bases does steps 2 to 5 for you from S3 and other sources, with scheduled or on-demand sync.

Vector stores on AWS

Store Notes
Amazon OpenSearch Service / Serverless Vector search plus keyword search
Aurora PostgreSQL / RDS for PostgreSQL with pgvector Vectors next to relational data
Amazon S3 Vectors Low-cost, large-scale vector storage
Amazon Neptune Analytics, DocumentDB, MemoryDB Graph, document and in-memory options

Using models in pipelines

  • Call Amazon Bedrock from Glue, Lambda or Step Functions to classify, summarize, extract entities or generate descriptions.
  • Batch inference in Bedrock processes large datasets offline at lower cost.
  • Redshift and Athena can call ML models from SQL in some setups.

Governance

  • Remove or mask PII before embedding (Macie, Comprehend PII detection, Glue sensitive data detection).
  • Keep metadata (source, date, access level) with each chunk so retrieval can filter by permissions.

Common traps

  • Embedding whole documents instead of chunks.
  • Never re-syncing the knowledge base, so answers go stale.
  • Embedding sensitive data that users shouldn’t retrieve.

Try the 1 practice question

Practice LLMs and vectors in data pipelinesTake the free practice exam

Question 1

Harder ·Link to this question (question 1)

An insurance company keeps about 40,000 policy documents as PDF and Word files in Amazon S3. Claims agents will use an assistant built on Amazon Bedrock that answers questions from these documents, and documents are added, changed or removed every day.

The data engineering team must chunk the documents, create embeddings, store them in a vector store and bring the index up to date each night. It wants the least custom code, and unchanged documents must not be processed again.

Which solution meets these requirements?

Choose one answer.

  1. A. Write an AWS Glue job that splits every document, calls an embeddings model for each chunk, and overwrites a DynamoDB table each night.
  2. B. Create a Bedrock knowledge base with the S3 bucket as its data source and a vector store, and start a data source sync each night.
  3. C. Fine-tune a foundation model on the documents each week, so the assistant answers from what the model learned rather than from a vector store.
  4. D. Have a Lambda function read the full text of the relevant files from S3 at question time and place all of it in the model prompt.
Show the answer and explanation

Answer: B

  • A. Wrong: The team would write and maintain the chunking and embedding code, and the job reprocesses every document each night. DynamoDB also does not offer vector similarity search for the assistant's lookups.
  • B. Right: Correct. A knowledge base parses, chunks and embeds the S3 documents and writes the vectors to its vector store. A sync is incremental, so it processes only documents that were added, modified or deleted since the last sync. Learn more: Sync your data with your Amazon Bedrock knowledge base (opens in a new tab)
  • C. Wrong: Fine-tuning is model training, which is outside this data pipeline's scope, and a weekly run leaves daily changes out of the answers. It also retrains on every document each time.
  • D. Wrong: Without an index, the function cannot tell which documents are relevant, and whole documents quickly exceed prompt limits and raise cost. The option skips the chunking and embedding the team was asked to build.

Explanation

Retrieval for a generative AI assistant needs documents chunked, embedded and stored in a vector store, then kept current. A Bedrock knowledge base does all of this from an S3 data source, and each sync handles only added, modified or deleted documents. Custom embedding jobs, fine-tuning and whole-document prompts add code, miss updates or skip the vector index.

Retrieval augmented generation with a knowledge base Sync: documents from S3 and other connectors are split into chunks, turned into embeddings, and stored in a vector store. Query: a user question retrieves the best matching chunks from the vector store, and the model writes an answer from them with citations. Sync Documents S3, connectors Chunk + embed embeddings Vector store managed or own Query User question from the app Retrieve best chunks Model answers with citations
The nightly sync chunks and embeds new policy documents, so agents' questions retrieve current text.

Study

Want more on LLMs and vectors in data pipelines? The 5 full-length practice sets have 5 more LLMs and vectors in data pipelines questions, 325 in all, none of them in the free exam. One payment of $14.99, no subscription. See the practice sets

Study LLMs and vectors in data pipelines

More Data Ingestion and Transformation topics