Foundation Model Integration, Data Management, and Compliance31% · 6 topics

Q1. Executives want to decide within 2 weeks whether generative AI can automate inquiry handling. Which approach is most appropriate?

1. Build full production infrastructure first, and only then check quality and cost
2. Build a PoC using several Bedrock models on representative inquiry data, and measure quality, latency and cost to decide
3. Only interview internal staff and judge feasibility on paper without trying any model
4. Pick the most famous model and connect it to the company-wide desk without a PoC

Q2. Multiple teams are each building generative AI apps, and design quality varies. Which is the most appropriate way to align designs?

1. Review designs with the Well-Architected Generative AI Lens and standardize shared components (prompt templates, guardrails, logging)
2. Have all teams copy the code of the first app and build the same architecture even when requirements differ
3. Write no design documents and just have each team explain verbally at a weekly meeting
4. Let each team design freely and rebuild only the apps that have problems later

Q1. You want an FM to read an entire contract of over 100 pages at once to find contradictory clauses. What should you check first when choosing the FM?

1. Whether the model provider has an office in Japan
2. Whether the color and font of the model's output can be changed to match the app's screen design
3. Whether the document, instructions and output fit within the model's maximum context length
4. Whether the model's name is easy to remember and explain to internal users

Q2. You are building an app where users send a photo of equipment together with a question, and it answers likely causes of failure. Which implementation is most appropriate?

1. Choose a model without image input and paste the image data as a string into the prompt
2. Write only the image file name in the text and let the model guess the content from the file name
3. Print the image, have staff transcribe its content into text, and send only that text to a text-only model
4. Choose a model that supports image input and send both the image and text in a Converse API message

Q1. For financial analysis, you want answers to hard questions that require many steps of calculation and deliberation. Somewhat longer response time is acceptable. What is the most appropriate model choice?

1. Use a model with reasoning capability and set a limit on how much it thinks
2. Use the reasoning model for every task, including simple questions
3. Fine-tune a large model on financial data and use it
4. Use a fast, cheap model with a chain-of-thought instruction

Q2. The model you want is not offered in the Tokyo Region, but it is available through a Japan geography cross-Region inference profile. What should you check before using it?

1. That using the profile automatically stores all data in the Tokyo Region
2. The weather in the destination Regions and the size of the data center buildings
3. Only whether the profile name matches internal naming rules
4. Whether processing in the Regions requests are routed to is allowed by company data-location rules

Q1. You want to clean up inconsistently written meeting notes before embedding. Which approach is most appropriate?

1. Ingest the notes as is and use smaller chunks to improve retrieval
2. Switch to a larger multilingual embedding model to absorb the variations
3. Enable hybrid search so keyword matches cover the variations
4. Normalize dates, amounts and product names with Lambda or Bedrock into a fixed format before ingesting

Q2. You want to ground answers only in “proposals from the last year for manufacturing customers.” Which design is most appropriate?

1. Build a separate app per industry and have sales reps choose the right one
2. Add industry, creation date and author metadata to proposals and use metadata filters in retrieval
3. Retrieve from all proposals and ask the model to ignore old ones
4. Write the industry in proposal file names and instruct the model in the prompt to pick correctly by file name

Q1. Vector search documents exceed 100 million, and search latency is increasing. What should you review first in OpenSearch Service?

1. Greatly increase vector dimensions to raise only search accuracy
2. Put all documents in one small shard and minimize instance count for cost
3. Review shard count and size, replicas, instance types and how indexes are split by domain
4. Rebuild the index on every search so searches are always up to date

Q2. You also want to use internal Confluence and SharePoint documents as RAG sources. Which approach requires the least effort?

1. Build an Amazon Kendra index and have the agent search it
2. Crawl both sites' URLs with a web crawler data source
3. Export documents to S3 with Glue jobs and use S3 as the data source
4. Connect Confluence and SharePoint with Bedrock knowledge base data source connectors

Q1. After switching the model to another provider's via AppConfig, failures to parse output JSON increased. What is the most appropriate next step?

1. Prepare per-model prompt templates, enforce format with structured output, and evaluate before switching
2. Switch gradually in AppConfig, raising the share while watching the failure rate
3. Add a fallback that resends failed requests to the previous model
4. Set temperature to 0 and keep the same prompt on the new model

Q2. Every time a prompt changes, you want to confirm quality has not dropped. Which mechanism is most appropriate?

1. Compare before and after each change in a CI pipeline with an evaluation dataset of expected outputs
2. Keep versions in Prompt Management so you can roll back when problems appear
3. Run an A/B test in production and keep the version users rate higher
4. Have the editor try a few questions in the playground and ship if fine