Foundation Model (FM)
A pretrained model capable of performing general-purpose AI tasks and being adapted to different applications.
Domain 1: Foundation Model Integration, Data Management, and Compliance
31%FM, Bedrock, RAG, embeddings, vector stores, retrieval, prompts, and resilience terms
Amazon Bedrock
AWS managed service for building generative AI applications using foundation models and related capabilities.
Amazon SageMaker AI
AWS service for building, training, customizing, deploying, and managing machine learning models.
RAG
Retrieval-Augmented Generation. A pattern where relevant external information is retrieved and supplied to the model as context before generation.
Embedding
A numerical vector representation of content that captures semantic relationships.
Vector Store
A system used to store and retrieve vector representations efficiently.
Chunking
Breaking documents into smaller retrievable units.
Metadata
Additional information associated with a document or chunk, such as author, date, domain, or classification.
Vector Search
Searching for content based on vector similarity.
Keyword Search
Searching based on lexical/exact terms.
Hybrid Search
Combining keyword and vector retrieval.
Reranking
Reordering retrieved candidates according to a stronger relevance evaluation.
Query Expansion
Adding related terms or concepts to improve retrieval.
Query Decomposition
Breaking a complex query into smaller subqueries.
Prompt Template
A reusable prompt structure with variable inputs.
Prompt Management
Managing reusable prompts, versions, and controlled prompt usage.
Guardrails
Controls designed to help enforce responsible and safe model interactions.
Prompt Flow
A workflow that orchestrates multiple prompt/model-related steps, including sequential and conditional processing.
Model Router
An abstraction layer that decides which model should handle a request.
Circuit Breaker
A resilience pattern that temporarily stops requests to a failing dependency.
Graceful Degradation
Providing a reduced but usable service when the preferred AI capability is unavailable.
LoRA
Low-Rank Adaptation, a parameter-efficient model adaptation technique.
Model Registry
A system for managing model versions and deployment lifecycle.
Data Validation
Checking that input data satisfies defined quality and schema requirements.
Data Normalization
Transforming inconsistent data into a consistent representation.
Multimodal
Handling multiple data modalities such as text, image, and audio.
POC
Proof of Concept. A limited implementation used to validate technical feasibility and business value.
AWS Well-Architected Framework
AWS guidance for evaluating architectures across important design principles and operational concerns.
Generative AI Lens
A Well-Architected Lens focused on considerations specific to generative AI workloads.
Domain 2: Implementation and Integration
26%Agents, MCP, deployment, enterprise integration, FM APIs, streaming, resilience, and application tooling
Strands Agents
Framework for building agentic applications where an agent can plan, use tools, and complete multi-step tasks.
AWS Agent Squad
Approach for coordinating multiple specialized agents that collaborate on different parts of a task.
MCP (Model Context Protocol)
Standardized way for an agent or model to interact with tools and external capabilities.
State Management
Tracking information about the current workflow or task, such as step, tool results, and next action.
Memory
Information retained across interactions or tasks, such as preferences and historical context.
ReAct Pattern
Agent pattern that alternates between reasoning and acting with tools until a stopping condition is met.
Human-in-the-Loop (HITL)
Design pattern where humans approve, review, or intervene in high-risk or consequential agent actions.
Circuit Breaker
Resilience pattern that temporarily stops calls to an unhealthy dependency instead of retrying endlessly.
Bedrock Provisioned Throughput
Reserved Bedrock model capacity used when predictable throughput and latency are required.
Model Cascading
Routing simpler requests to smaller models and escalating complex requests to larger or specialized models.
GenAI Gateway
Abstraction layer that centralizes model access, routing, auth, and policy for enterprise GenAI applications.
Streaming Response
Returning model output incrementally as tokens are generated, often via WebSockets or Server-Sent Events.
Exponential Backoff
Retry strategy that increases wait time between attempts to avoid overwhelming a failing dependency.
Model Routing
Selecting which foundation model handles a request based on rules, request characteristics, or metrics.
AWS Outposts
AWS infrastructure deployed closer to or in on-premises environments for low-latency or data-residency needs.
AWS Wavelength
AWS edge infrastructure for ultra-low-latency workloads, often associated with 5G applications.
Amazon Q Developer
AWS developer assistant used to improve productivity for coding, diagnosis, and suggested fixes.
AWS X-Ray
Distributed tracing service used to follow a request across service boundaries and locate latency or failures.
Prompt Flows
Bedrock capability for visual or declarative orchestration of multi-step model workflows.
Least Privilege
IAM principle of granting only the minimum permissions required for an agent, tool, or service to operate.
Domain 3: AI Safety, Security, and Governance
20%Guardrails, PII, PrivateLink, Model Cards, invocation logging, Clarify, and responsible AI
Amazon Bedrock Guardrails
Runtime safety controls that filter inputs and outputs using content filters, denied topics, word filters, PII actions, prompt-attack filters, and contextual grounding checks.
Input Checkpoint (Gate 1)
Guardrails evaluation of the incoming prompt before the foundation model is invoked.
Output Checkpoint (Gate 2)
Guardrails evaluation of the model completion before it is returned to the client.
Denied Topics
Natural-language topic definitions that Guardrails blocks across supported models without changing prompts or weights.
Contextual Grounding Checks
Guardrails capability that compares a candidate completion to retrieved source reference passages to detect ungrounded hallucinations.
PII ANONYMIZE
PII handling mode that replaces sensitive values with placeholders such as [NAME] or [EMAIL] while allowing the request to continue.
PII BLOCK
PII handling mode that rejects the request when configured sensitive entities are detected.
AWS PrivateLink / Interface VPC Endpoint
Private network path for Bedrock API calls so traffic stays on the AWS backbone instead of the public internet.
KMS Customer Managed Key (CMK)
Customer-controlled encryption key used for data at rest across Bedrock-related stores such as S3, Knowledge Bases, and OpenSearch.
Amazon Macie
Service that discovers sensitive data such as PII in S3 before indexing into RAG corpora.
SageMaker Model Cards
Standardized documentation of model provenance, training data, metrics, and intended usage boundaries for audits.
Bedrock Model Invocation Logging
Opt-in logging of Bedrock request payloads, completions, and metadata to CloudWatch Logs or encrypted S3.
AWS Glue Data Catalog
Central metadata catalog for training and RAG datasets used for lineage and schema tracking.
AWS Lake Formation
Fine-grained access control over data lakes, often paired with Glue Data Catalog for AI dataset governance.
SageMaker Clarify
Bias and fairness evaluation and continuous monitoring for model endpoints.
Bedrock Agent Tracing
Step-by-step visibility into agent reasoning paths such as Rationale, Action, and Observation.
Prompt Attack / Jailbreak Filter
Guardrails filter that detects attempts to override instructions or exfiltrate system prompts.
Domain 4: Operational Efficiency and Optimization for GenAI Applications
12%Token cost, cascading, Prompt Caching, Batch Inference, streaming, and GenAI observability
Prompt Compression
Reducing prompt fluff and retaining only critical semantic content to lower input token cost.
Model Cascading
Routing simple queries to cheap/fast models and escalating complex queries to higher-capability models.
Bedrock Batch Inference
Asynchronous Bedrock execution for non-real-time jobs, typically ~50% cheaper than on-demand calls.
Bedrock Prompt Caching
Caches pre-computed KV activation states for long static prompt segments to cut repeated input token charges and latency.
Semantic Caching
Returning a stored answer when a new query is highly similar to a prior one, avoiding a new FM invocation.
Time-To-First-Token (TTFT)
Latency until the first generated token reaches the user; improved by streaming.
ConverseStream / InvokeModelWithResponseStream
Bedrock streaming APIs that emit tokens incrementally for WebSocket or SSE delivery.
Temperature
Decoding parameter controlling randomness; near 0.0 for deterministic math/code, higher for creative tasks.
Top-P / Top-K
Decoding controls that limit sampling to a nucleus of high-probability tokens.
Metadata Pre-Filtering
Restricting vector search candidates by metadata before distance calculations to improve retrieval latency.
HNSW / ANN Index
Approximate nearest-neighbor indexing used to accelerate vector similarity search.
InputTokenCount / OutputTokenCount
CloudWatch/Bedrock metrics used to track token consumption for cost and capacity analysis.
Bedrock Provisioned Throughput
Reserved Bedrock capacity for sustained high-volume traffic with strict latency SLAs.
Domain 5: Testing, Validation, and Troubleshooting
11%LLM-as-a-Judge, quality gates, truncation diagnostics, Hybrid Search, and agent IAM troubleshooting
ROUGE / BLEU
Lexical n-gram overlap metrics; useful for basic summary matching but weak for factual correctness.
BERTScore
Embedding-based semantic similarity between generated text and reference answers.
LLM-as-a-Judge
Using a stronger foundation model to score candidate responses against defined quality rubrics.
Faithfulness / Groundedness
Whether every claim in a completion is supported by retrieved context.
Answer Relevance
Whether the completion directly addresses the user prompt without tangential bloat.
Context Precision & Recall
Whether retrieval pulled needed facts without excess noise.
Position Swapping
Evaluating A vs B and B vs A, then averaging, to neutralize LLM judge position bias.
Amazon Bedrock Model Evaluation
Managed evaluation jobs with Automated or Human Evaluation workflows for model/prompt comparison.
Quality Gate
CI/CD check that blocks promotion when faithfulness, truncation, or other scores miss thresholds.
stopReason: max_tokens
Completion stopped because the configured output token limit was hit; fix by raising maxTokens.
ValidationException (input overflow)
Pre-generation failure when prompt+context exceed the model context window; fix by chunking/compressing input.
Hybrid Search
Combining BM25 keyword matching with dense vector search, often plus reranking, for exact codes and semantic matches.
StartIngestionJob
API/job that syncs Bedrock Knowledge Bases with source data (e.g., S3) after uploads or updates.
lambda:InvokeFunction (Agents)
IAM permission the Bedrock Agent execution role needs to call tool Lambdas; missing it yields AccessDeniedException.