AWA AIF-C01 - Domain 3 - Applications of Foundation Models

A broadly trained model that can be adapted for many different downstream tasks and applications.

Choosing a foundation model based on factors such as task requirements, performance, cost, latency, context window, and supported modalities.

The scale of a model, often associated with its number of parameters. Larger models may provide greater capabilities but generally require more computational resources.

Learned numerical values within a model that are adjusted during training to represent patterns in the training data.

Using a trained foundation model to generate an output from an input prompt or request.

Settings that influence how a model generates its response, such as temperature, top-p, and maximum output tokens.

A setting that controls randomness in generated output. Higher values generally produce more varied responses; lower values generally produce more predictable responses.

A sampling setting that limits token selection to a probability mass containing the most likely candidate tokens.

A limit on how many tokens a model can generate in its response.

The input instructions, context, data, or examples supplied to a foundation model to guide its output.

The practice of designing prompts to obtain more useful, accurate, relevant, and consistent model outputs.

Asking a model to perform a task without providing examples of the desired task in the prompt.

Providing a single example demonstrating the desired task or output format.

Providing multiple examples to demonstrate the desired task, pattern, or output format.

A reusable prompt structure containing fixed instructions and variable fields that can be populated for different requests.

Instructions that establish the model's behavior, role, constraints, or priorities for an interaction.

An attack in which malicious or unintended instructions are inserted into model input to manipulate the model's behavior.

Connecting model responses to reliable, relevant information from external sources rather than relying solely on the model's learned knowledge.

A technique that retrieves relevant information from external data sources and provides it to a model as context when generating a response.

A collection of information that can be searched or retrieved to provide relevant context to an AI application.

A database optimized for storing and searching numerical vector representations, commonly used for semantic similarity searches.

Searching based on the meaning or intent of content rather than requiring an exact keyword match.

Finding stored vectors that are mathematically similar to a query vector, often using distance or similarity measures.

A prompt enriched with retrieved information, additional context, instructions, or other data before being sent to a foundation model.

Additional training of a foundation model using task- or domain-specific data to modify or improve its behavior for a particular use case.

Training a model on examples of instructions and desired responses so it becomes better at following natural-language instructions.

Adapting a foundation model to better meet a specific application's requirements through techniques such as prompting, RAG, or fine-tuning.

Further pre-training a foundation model on additional domain-specific or specialized data to expand or adapt its learned knowledge.

A fully managed AWS service that provides access to foundation models from multiple providers for building generative AI applications.

Bedrock allows applications to select foundation models from different providers through a common managed service interface.

A Bedrock capability that helps applications implement RAG by connecting foundation models with organizational data and retrieving relevant information.

A Bedrock capability that enables foundation models to perform multi-step tasks by using instructions, knowledge sources, and connected actions.

A Bedrock capability for applying configurable safeguards to model inputs and outputs, helping enforce responsible AI requirements.

A capability for creating, storing, managing, and reusing prompts and prompt variations for generative AI applications.

Capabilities for evaluating foundation models using automated or human-based methods against defined criteria.

A capability that allows supported foundation models to be customized using a user's training data for a particular task or domain.

A managed AWS service for building, training, customizing, deploying, and operating machine learning models.

AWS's family of generative AI assistants designed to help users with tasks such as business questions, development, and AWS-related work.

Breaking a complex task into multiple model calls where the output of one step becomes input or context for another step.

Directing requests to different models based on factors such as task complexity, performance requirements, cost, or capability.