AIP-C01 glossary of Bedrock, RAG, vector stores, prompts, agents, guardrails, governance, monitoring, and evaluation terms.
Use this glossary when AWS Certified Generative AI Developer - Professional (AIP-C01) terms start to blur together. The goal is exam recognition: know what each term does, when it is the best answer, and when it is only a distractor.
| Term | Exam meaning | Watch for |
|---|---|---|
| Amazon Bedrock | AWS managed service for building GenAI applications with foundation models, Knowledge Bases, Agents, Guardrails, model invocation, and related capabilities. | Production answers still need IAM, networking, logging, evaluation, and cost controls. |
| Foundation model (FM) | A general-purpose model used for generation, reasoning, summarization, extraction, embeddings, or multimodal tasks. | Bigger is not automatically better; choose by capability, latency, cost, safety, and context needs. |
| Amazon Bedrock Knowledge Bases | Managed RAG capability that connects data sources, embeddings, vector stores, and retrieval for grounded responses. | It does not remove the need for source permissions, freshness, metadata, and retrieval evaluation. |
| Amazon Bedrock Agents | Bedrock capability for tool-using workflows where the model can plan and call actions. | Tool permissions, validation, state, failure handling, and tracing are part of the answer. |
| Amazon Bedrock Guardrails | Policy controls for filtering or constraining inputs and outputs. | Guardrails do not replace IAM, KMS, data classification, retrieval filtering, or audit evidence. |
| Amazon Bedrock Prompt Management | Managed prompt templates, versions, and governance workflows. | Strong when the scenario mentions reusable prompts, approval, rollback, or regression control. |
| Amazon Bedrock Prompt Flows | Workflow-style prompt orchestration with sequential or conditional steps. | Strong when the problem is multi-step prompt logic, not just a single prompt tweak. |
| Amazon Bedrock Model Invocation Logs | Logs that help inspect model requests and responses for debugging, audit, and behavior analysis. | Logging must respect privacy, retention, and access-control requirements. |
| Amazon CloudWatch | Metrics, logs, alarms, dashboards, and anomaly detection for operational visibility. | GenAI monitoring also needs quality, safety, token, retrieval, and version signals. |
| AWS Lambda | Serverless compute often used for routing, preprocessing, post-processing, retrieval glue, or tool execution. | Watch timeout, retries, idempotency, permissions, and cold-start impact. |
| AWS Step Functions | Orchestration service for workflows, retries, branching, circuit breakers, and human-in-the-loop paths. | Strong when workflow order, fallback, approval, or partial failure matters. |
| Amazon OpenSearch Service | Search and vector search platform often used for hybrid retrieval and vector indexes. | Tune indexes, metadata filters, ranking, and shard strategy before blaming the FM. |
| Amazon S3 | Object storage for documents, prompt repositories, logs, artifacts, and source data. | Source data still needs classification, lifecycle, encryption, and access controls. |
| AWS KMS | Key management for encryption of data, logs, and storage paths. | Key policy matters as much as the encryption checkbox. |
| AWS IAM | Identity and access management for applications, data, tools, models, and logs. | Separate permissions for model invocation, retrieval, tool calls, and audit access. |
| AWS CloudTrail | Audit logging for API activity and governance evidence. | Useful for traceability, but not a quality evaluation system by itself. |
| Term | Exam meaning | Watch for |
|---|---|---|
| Retrieval Augmented Generation (RAG) | Pattern that retrieves external context and gives it to an FM to ground responses. | Best for current or proprietary facts; not a substitute for permissions or data quality. |
| Embedding | Vector representation of text, images, or other content used for similarity search. | Choose embedding fit by domain, language, dimensionality, retrieval quality, and cost. |
| Vector store | Database or index that stores embeddings and metadata for semantic retrieval. | Design metadata, indexing, encryption, freshness, and evaluation, not just storage. |
| Chunking | Splitting documents into retrievable segments. | Poor chunk size or structure causes missing context, noisy answers, and higher token cost. |
| Hierarchical chunking | Chunking that preserves document structure such as headings, sections, and parent-child context. | Strong when fixed-size chunks lose meaning across sections. |
| Metadata filtering | Using tags such as tenant, date, document type, owner, or sensitivity to narrow retrieval. | Critical for authorization, relevance, and freshness. |
| Hybrid search | Combining lexical keyword search with vector similarity. | Strong when exact identifiers, policy names, SKUs, error codes, or legal terms matter. |
| Reranking | Reordering candidate retrieval results with a scoring model or secondary ranker. | Useful when initial retrieval finds candidates but relevance order is weak. |
| Query expansion | Rewriting or expanding a user query to improve retrieval. | Can help recall but may increase noise if uncontrolled. |
| Query decomposition | Breaking a complex query into smaller retrieval questions. | Useful for multi-part questions or workflows that need separate evidence. |
| Source attribution | Preserving source IDs, citations, and metadata in the response path. | Required when auditability, transparency, or compliance is named. |
| Retrieval freshness | Keeping indexes synchronized with current source data. | Stale answers usually point to ingestion, sync, cache, or index maintenance before model choice. |
| Term | Exam meaning | Watch for |
|---|---|---|
| Prompt template | Reusable prompt structure with variables and instructions. | Needs versioning, approval, testing, and rollback in production. |
| Prompt assembly | Runtime construction of system instructions, user input, retrieved context, tools, and output constraints. | Debug the assembled request, not only the template fragment. |
| System instruction | High-priority instruction that defines behavior and safety rules. | Keep it separate from user-controlled or retrieved text. |
| Context window | Maximum amount of input and output context a model can process. | Larger context can increase latency, cost, and distraction. |
| Temperature | Sampling parameter that affects randomness. | Lower for deterministic extraction; higher only when creativity is useful and safe. |
| Top-k / top-p | Sampling controls that influence token selection. | Tune with evaluation, not trial-and-error impressions. |
| Structured output | Output constrained to JSON, schema, table, fields, or another required format. | Pair with validation and retries where safe. |
| Function calling / tool use | Model-driven selection or invocation of application tools. | Tools need scoped permissions, input validation, tracing, and failure handling. |
| Agent workflow | Multi-step task pattern involving planning, tools, state, and sometimes human review. | Agents are application workflows, not merely smarter prompts. |
| Model routing | Selecting a model dynamically by task, cost, latency, risk, or availability. | Strong for mixed workloads and fallback patterns. |
| Cross-Region Inference | Bedrock pattern that can help route inference across supported Regions for availability or capacity needs. | Use when the selected model and workload support it; preserve controls and observability. |
| Model customization | Fine-tuning, adapter, or other adaptation process for repeated domain behavior gaps. | Requires lifecycle, registry, deployment, evaluation, rollback, and retirement. |
| Term | Exam meaning | Watch for |
|---|---|---|
| Prompt injection | Malicious or accidental input that tries to override instructions or misuse tools. | Treat user and retrieved text as untrusted; validate tools and preserve instruction hierarchy. |
| Jailbreak | Attempt to bypass safety rules or content controls. | Use guardrails, classifiers, adversarial tests, and monitoring. |
| PII detection | Identifying personally identifiable information in source data, prompts, outputs, or logs. | Detect before storage, retrieval, prompt construction, and output where required. |
| Data masking | Replacing sensitive values before processing or output. | Must preserve task utility while reducing exposure. |
| Tenant isolation | Ensuring one tenant cannot retrieve, prompt with, log, or receive another tenant’s data. | Authorization must happen before retrieval and tool execution. |
| Least privilege | Giving only the permissions needed for model, data, tool, and logging paths. | Broad app roles are weak in GenAI scenarios. |
| Private access | Network design that keeps model/data traffic off public paths where required. | Combine VPC endpoints, policies, IAM, KMS, and logging. |
| Data lineage | Record of where source data came from and how it moved through the system. | Important for audit, compliance, and source attribution. |
| Model card | Documentation about model purpose, limitations, evaluation, and governance considerations. | Useful for responsible AI and deployment review; not runtime enforcement by itself. |
| Responsible AI | Principles such as fairness, transparency, safety, privacy, accountability, and human oversight. | Exam answers should operationalize these with controls, evaluation, and monitoring. |
| Term | Exam meaning | Watch for |
|---|---|---|
| Golden dataset | Representative test set used to compare prompts, models, retrieval, or releases. | Strong for regression testing and quality gates. |
| LLM-as-a-judge | Using a model to evaluate model outputs. | Needs calibration, representative tests, and human review for high-risk use. |
| Groundedness | Degree to which an answer is supported by retrieved or provided evidence. | Poor groundedness usually points to retrieval, prompt assembly, or citation controls. |
| Hallucination | Plausible but unsupported or incorrect model response. | Do not fix only with lower temperature; check grounding, evaluation, and refusal behavior. |
| Canary release | Limited rollout to detect regressions before broad deployment. | Pair with alarms, evaluation metrics, and rollback criteria. |
| A/B test | Comparison of versions such as prompt A versus prompt B or model A versus model B. | Use shared metrics and representative traffic. |
| Output diffing | Comparing response changes across versions. | Useful for subtle regression detection after prompt/model/retrieval changes. |
| Token tracking | Measuring input and output token usage by route, model, user, tenant, prompt, or version. | Essential for cost, latency, and anomaly detection. |
| Semantic cache | Cache that reuses answers for similar intent, not only exact prompt matches. | Requires similarity thresholds, freshness, access scope, and safety review. |
| Provisioned throughput | Reserved model capacity pattern for predictable throughput needs. | Strong only when demand and utilization justify it. |
| Backpressure | Slowing or controlling intake when downstream capacity is saturated. | Better than uncontrolled retries or hidden partial failure. |
| Circuit breaker | Pattern that stops repeated calls to a failing dependency and routes to fallback behavior. | Useful for tool, model, or service disruption scenarios. |
| Forensic traceability | Ability to reconstruct prompt version, model route, retrieved sources, tool calls, approvals, and output. | Critical for compliance, incident review, and regression analysis. |
| Pair | How to separate them |
|---|---|
| RAG vs fine-tuning | RAG supplies changing facts; fine-tuning/customization changes repeated model behavior. |
| Guardrail vs IAM | Guardrail constrains content; IAM authorizes identities, data, models, and tools. |
| Prompt template vs prompt assembly | Template is the reusable design; assembly is the actual runtime request sent to the model. |
| Retrieval quality vs answer quality | Retrieval quality is the evidence supplied; answer quality is what the model produces from evidence and instructions. |
| Latency vs throughput | Latency is one request’s time; throughput is how much work the system handles under load. |
| Evaluation vs monitoring | Evaluation compares expected behavior; monitoring watches production behavior and drift. |
| Logging vs auditability | Logging records events; auditability preserves the right fields, access, retention, and trace context. |
| Prevention vs detection | Prevention blocks or reduces a bad event; detection finds or reports that it happened. |
| Managed service vs custom build | Managed services reduce operational work unless the requirement needs unsupported customization. |
Do not memorize terms in isolation. For each term, write one scenario where it is the best answer, one scenario where it is a distractor, and one signal that proves it worked.