AWS GenAI Developer Pro flashcards

33 cards. Read the term and the definition here. The deck at the top of this page is optional.

Converse API

A Bedrock runtime API that accepts a model-agnostic list of messages and returns a uniform response, letting one code path call multiple foundation models.

Tool use flow

The model returns a toolUse block naming a function and arguments; the application executes it, returns a toolResult block, and re-invokes the model for a final answer.

Provisioned throughput

Pre-purchased Bedrock capacity measured in model units that guarantees invocation throughput and avoids on-demand throttling for steady high-volume workloads.

Cross-Region Inference

An inference profile that automatically routes requests to multiple regions, increasing available throughput beyond a single region's quota.

Response streaming

InvokeModelWithResponseStream returns model output as incremental chunks over an event stream, improving perceived latency for chat interfaces.

Exponential backoff with jitter

A retry strategy that increases wait time between attempts and adds randomness, preventing synchronized retry storms during Bedrock throttling errors.

Knowledge base ingestion

The pipeline that syncs a data source, splits documents into chunks, generates embeddings, and writes vectors to the configured vector store.

Chunk overlap

Duplicate text carried between adjacent chunks so information spanning a boundary remains retrievable; too much overlap inflates index size and cost.

Hybrid search

Combining keyword matching with vector similarity to improve recall for exact terms, identifiers, and rare words that embeddings represent poorly.

Metadata filtering

Restricting vector search candidates by attributes such as tenant ID or document date before similarity ranking, reducing irrelevant results and enforcing isolation.

RetrieveAndGenerate

A Bedrock Knowledge Base API that performs retrieval, builds the augmented prompt, invokes the model, and returns the answer with source citations.

Reranking

A second-stage scoring step that reorders initially retrieved passages by relevance to the query, improving precision before passages enter the prompt.

Fine-tuning data format

JSONL records containing prompt and completion fields stored in S3, optionally split into training and validation sets to detect overfitting.

Continued pre-training

Customization using large unlabeled domain text to adapt a base model's vocabulary and style before any task-specific fine-tuning occurs.

Distillation

Training a smaller student model on high-quality outputs from a larger teacher model to reduce latency and inference cost while retaining accuracy.

Custom model invocation

A fine-tuned Bedrock model is invoked through its custom model ID and requires provisioned throughput; on-demand invocation is not supported.

LLM-as-a-judge

Using a capable model with a scoring rubric to evaluate outputs at scale; results should be calibrated against human-labeled samples.

Golden dataset

A fixed, curated set of inputs with expected outputs used to compare model versions and gate promotion on metric regressions.

IAM model scoping

Restricting bedrock:InvokeModel permissions to specific model or inference profile ARNs so callers can only use approved foundation models.

KMS customer managed key

An encryption key you control whose policy must grant the Bedrock service principal access to encrypt knowledge base vectors and custom model artifacts.

PrivateLink endpoint

A VPC interface endpoint that keeps Bedrock and vector store traffic on the AWS network, removing exposure to the public internet.

Prompt injection

Malicious instructions embedded in user input or retrieved documents that attempt to override system instructions or trigger unauthorized tool calls.

Tenant isolation in RAG

Enforcing per-tenant separation using metadata filters on retrieval, scoped IAM credentials, and separate indexes where strict isolation is required.

Invocation logging

An optional Bedrock setting that writes request and response bodies to CloudWatch Logs or S3 for auditing and debugging, requiring explicit enablement.

InvocationLatency metric

A CloudWatch metric reporting model response time; alarm thresholds detect degradation caused by larger prompts or upstream throttling.

Token cost formula

Total cost equals input tokens multiplied by input price plus output tokens multiplied by output price, with output usually priced higher.

Semantic caching

Storing prior answers keyed by embedding similarity so near-duplicate queries return cached responses without another model invocation.

Batch inference

An asynchronous Bedrock job that processes large offline prompt sets at lower cost than synchronous on-demand invocation.

Content filters

Guardrail settings that block or flag categories such as hate, insults, sexual content, and violence at configurable strength levels.

Contextual grounding check

A guardrail that scores whether an answer is supported by retrieved source text and relevant to the question, blocking low-scoring responses.

Denied topics

Guardrail definitions of subjects the application must not discuss, described with natural-language definitions and optional example phrases.

Automated reasoning checks

A guardrail feature that validates model claims against formal policy rules and identifies which specific rule a response violates.

CloudTrail model audit

Recording Bedrock control-plane and invocation API activity in CloudTrail trails to support governance, investigation, and compliance reporting.