Foundation Models on AWS — AWS AIF-C01

A prompt is the complete input passed to a generative AI model, and can combine a system instruction, retrieved context, few-shot examples, and the user’s actual question, all concatenated before the model sees them. In Amazon Bedrock, each call to a foundation model is built around this prompt payload, and the Converse API gives a consistent format for handling multi-turn conversation history alongside it. The exam distinction to know is prompt versus system prompt: the system prompt sets persistent behavior (tone, role, guardrails), while the user prompt carries the per-turn request.

Chunking breaks source documents into smaller text segments before they are converted into embeddings and stored in a vector database. In a Retrieval-Augmented Generation (RAG) pipeline built with Amazon Bedrock Knowledge Bases, chunking happens at ingestion time and directly shapes retrieval quality at inference time. Note the distinction from embedding: chunking sets the unit of text that gets retrieved, while embedding determines how that unit is represented numerically. Chunk size is a balancing act — larger chunks preserve more surrounding context but can flood the prompt with irrelevant content, whereas smaller chunks sharpen precision but may omit the context needed for a coherent answer.

Attention is a learned mechanism inside transformer models that assigns a weight to every token pair, letting the model decide how much one token should influence another’s representation. In self-attention, each token attends to all others in the same sequence, producing context-aware embeddings rather than fixed-meaning word vectors. On the AIF-C01 exam, distinguish self-attention from cross-attention: self-attention operates within one sequence, while cross-attention connects two, such as an encoder output and a decoder input in translation. Models accessed through Amazon Bedrock rely on attention-based transformers, which is why they handle long-range dependencies that earlier recurrent networks struggled with.

Nondeterminism is the behavior of large language models where identical prompts can produce different outputs across separate invocations. It arises because models sample from a probability distribution over possible next tokens rather than always selecting the highest-probability token. In Amazon Bedrock, foundation models expose inference parameters — most notably temperature and top-p — that control how much randomness is injected during sampling. The key exam distinction is between nondeterminism and hallucination: nondeterminism is an intentional, controllable design property, while hallucination is an unintended factual error. Setting temperature to zero minimizes nondeterminism but does not eliminate hallucination, making them separate concerns when evaluating model reliability.

Interpretability refers to the degree to which a human can understand the internal mechanisms of a model — why specific inputs produce specific outputs. It is a structural property of the model itself: simpler models like linear regression or shallow decision trees are inherently more interpretable than deep neural networks or large language models, whose billions of parameters resist human inspection. On AIF-C01, interpretability falls under responsible AI and governance concerns in Domain 4. The key exam distinction is that interpretability concerns the model’s inner workings, while explainability describes model behavior in human-understandable terms after the fact, often via post-hoc tools. Amazon SageMaker Clarify provides feature-attribution explanations that address explainability without making a black-box model inherently interpretable. High interpretability is often traded against model performance — watch for questions framing this as a transparency-versus-accuracy tension.

The context window is the total number of tokens a model can process in one request, covering both the input prompt and the generated output. In Amazon Bedrock, each foundation model exposes a different maximum context length, so model choice means matching that window to the task, such as summarizing long legal documents. The exam nuance is that input and output share the window: a very long prompt leaves fewer tokens for the response. This differs from RAG, which extends effective memory by retrieving relevant chunks before the call rather than enlarging the window itself.

Amazon Bedrock is a fully managed, serverless service giving developers single-API access to a curated catalog of foundation models from AWS and third parties like Anthropic, Meta, and Mistral — no infrastructure or training required. It’s the exam’s go-to answer for the fastest, lowest-operational-overhead path to generative AI on AWS, since you switch providers by changing one API call. Know its build-on-top features: Knowledge Bases (managed RAG grounding on private data), Agents (multi-step task automation with tool calls), and Guardrails (content filtering and safety). Don’t confuse it with SageMaker, which trains, fine-tunes, and hosts custom models.

Amazon SageMaker JumpStart is a model hub and solution accelerator inside SageMaker Studio that lets you browse, evaluate, and deploy foundation models and pre-built ML solutions without writing infrastructure code. It supports one-click deployment to SageMaker endpoints and fine-tuning of supported models on custom datasets, covering both open-source and select proprietary models. The key exam distinction is that JumpStart is a SageMaker feature, not Amazon Bedrock. Bedrock is the fully managed, serverless API for foundation models; JumpStart gives more control but runs within the SageMaker training and hosting environment.

PartyRock is a no-code, browser-based app builder powered by Amazon Bedrock that lets anyone create generative AI apps by connecting widgets such as text generators, image generators, and chatbots without writing code. It is built for exploration and rapid prototyping, making foundation models accessible to non-developers. On the AIF-C01 exam, the key point is that PartyRock sits above Bedrock: Bedrock provides the underlying foundation model APIs and enterprise controls, while PartyRock is a self-contained playground built on top of it. PartyRock is not a production deployment tool and does not expose Bedrock’s programmatic APIs.

Amazon Q is a family of generative AI assistants built by AWS for two audiences. Amazon Q Business connects to internal knowledge sources — documents, wikis, ticketing systems — so employees get answers grounded in company data without writing model code. Amazon Q Developer aids software engineers with inline code suggestions, debugging help, and AWS resource guidance inside IDEs and the console. The key exam distinction is scope: Amazon Q is a fully managed, end-user-facing product, not an API for building custom AI applications. Candidates confuse it with Amazon Bedrock, which gives developers foundation model access to build their own solutions. Amazon Q is the finished assistant; Bedrock is the platform underneath.