Oracle Agentic AI Foundations flashcards

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

Agent reasoning loop

The repeated cycle where a model plans, selects a tool, observes the result, and updates its plan until a stop condition or iteration budget is reached.

ReAct pattern

An agent design that alternates explicit reasoning traces with tool actions and observations, improving grounding and making each decision step auditable in logs.

Orchestrator-worker topology

A supervisor agent decomposes the goal and delegates subtasks to specialized worker agents, then assembles and validates their outputs before returning a final answer.

Agent memory tiers

Short-term context window, persisted scratchpad state, and long-term external stores retrieved on demand; each tier has different cost, latency, and retention behavior.

Idempotent tool design

Tools accept an idempotency key so retried calls do not duplicate side effects, which is essential because agent loops may repeat actions after timeouts.

Human-in-the-loop gate

A mandatory approval checkpoint inserted before irreversible or high-value actions, pausing the agent run until an authorized human confirms or rejects the step.

OCI Generative AI service

Oracle's managed inference API for hosted foundation models, used as the reasoning backend so agents do not host or operate model weights themselves.

OCI Generative AI Agents

A managed retrieval-augmented generation service that connects data sources, builds a knowledge base with a vector store, and exposes a grounded agent endpoint.

Autonomous Database AI Vector Search

Oracle Database 23ai capability for storing embeddings and running similarity queries with SQL, letting agents retrieve enterprise data without a separate vector database.

Workload identity for agents

A distinct OCI IAM principal per agent or tool, separate from human users, so permissions can be scoped, audited, and revoked independently.

Model routing by task

Assigning cheap classification or extraction steps to small models and complex planning steps to larger models, reducing cost and latency without hurting quality.

Private endpoint connectivity

Reaching Oracle databases and object storage through a virtual cloud network private path instead of public internet egress, limiting exposure of agent traffic.

Tool schema design

Typed parameters, enums, explicit units, and narrow scope in a tool definition; schema quality is the strongest lever on valid tool-call generation.

Tool description as router

Because the model selects tools from description text, each description must state when to use the tool and when another tool is preferable.

Structured tool errors

Returning error code, message, and retryable flag to the model so it can adapt its plan instead of assuming the call succeeded.

Agent loop budget caps

Hard limits on steps, tokens, wall-clock time, and spend per run, preventing runaway loops and unbounded cost in production agents.

Circuit breaker for tools

A wrapper that stops calling a failing downstream service after a threshold of errors, letting the agent degrade gracefully instead of retrying endlessly.

Tool output as untrusted data

Retrieved text and API responses must never be treated as instructions, because embedded adversarial content can hijack the agent's next action.

RAG pipeline stages

Ingest, parse and chunk, embed, index, retrieve, rerank, then ground the prompt with cited passages; each stage is separately measurable and tunable.

Chunking strategy tradeoff

Smaller chunks sharpen retrieval precision but fragment context; structure-aware splits with overlap usually outperform naive fixed-length chunking on enterprise documents.

Hybrid retrieval with reranking

Combining dense vector similarity with keyword search, then reranking the merged candidate set, typically improves recall and precision over vector-only retrieval.

Embedding model consistency

Index and query must use the same embedding model and version; swapping models requires re-embedding the entire corpus or results become meaningless.

Metadata-filtered retrieval

Applying tenant, date, or classification filters during search enforces access control and relevance, preventing cross-tenant leakage and stale answers.

Grounded refusal

When retrieval returns no passage above the relevance threshold, the agent states it lacks evidence instead of answering from parametric memory.

Retrieval evaluation metrics

Recall@k and context precision measure the retriever separately from the generator, localizing whether failures come from search or from prompting.

Prompt injection defense

Separating instructions from data, validating outputs, and allow-listing tool calls to stop adversarial content in retrieved documents from redirecting agent behavior.

Confused deputy risk

An agent using a privileged service account on behalf of a user can exceed that user's rights; propagate the user's authorization context through every tool call.

Guardrails layer

Pre-execution checks such as input filtering, PII redaction, toxicity screening, and output schema validation that run before any tool action is committed.

Least-privilege tool identity

Each tool's credential is scoped to the minimum actions it needs, read-only by default, so a compromised agent cannot perform broad destructive operations.

NIST AI RMF functions

Govern, Map, Measure, Manage: a voluntary framework structuring AI risk identification, assessment, and treatment across an organization's AI lifecycle.

Agent run trace

A per-run record of prompt version, model, tool arguments, step latencies, token counts, and outcome, forming the basis for debugging and audit.

Golden dataset regression

A fixed set of representative tasks with expected outcomes, run in CI to block prompt, model, or tool changes that degrade success rate.