NCP-AAI Exam Prep
NCP-AAI Exam Glossary - 40 Terms
Search the terminology pack for NVIDIA Certified Professional: Agentic AI. Use these definitions with the study guide and practice questions.
A
- agent coordination
- The process by which multiple agents synchronize actions, share information, and work toward a common goal.
- automated retraining
- A pipeline that updates or fine-tunes models automatically based on new data, feedback, or evaluation results.
C
- case-based reasoning
- A problem-solving approach that stores past successful cases and retrieves similar examples to guide new decisions.
- chain-of-thought prompting
- A prompting method that encourages a model to generate intermediate reasoning steps before producing an answer.
- chunking
- The process of splitting large documents or inputs into smaller segments for processing, indexing, or summarization.
- context management
- Techniques for selecting, storing, and presenting relevant prior information to a model during interaction.
- context-aware suggestions
- Recommendations or prompts generated based on the current conversation state, user intent, or surrounding context.
- continuous improvement loop
- A recurring process that collects user feedback, analyzes it, and applies changes to improve system performance over time.
- conversational UI
- A user interface that supports natural multi-turn dialogue between a user and a system or agent.
- cross-encoder
- A model that jointly encodes a query and candidate document to produce a high-quality relevance score for reranking.
D
- distributed architecture
- A system design where computation and storage are spread across multiple machines to improve scale and resilience.
E
- expert reviewers
- Human evaluators with domain knowledge who assess model outputs for quality, correctness, or usefulness.
F
- faithfulness
- An evaluation criterion measuring whether generated content accurately reflects source information without hallucination.
- feedback analysis
- The process of reviewing and interpreting collected feedback to identify issues, trends, and improvement opportunities.
G
- goal decomposition with monitoring
- A planning framework that breaks broad objectives into measurable sub-goals and tracks progress with regular feedback.
- gRPC
- A high-performance remote procedure call framework commonly used for efficient service-to-service communication.
H
- hierarchical summarization
- A summarization approach that summarizes smaller chunks first and then summarizes the summaries to reduce cost while preserving information.
- horizontal scalability
- The ability to increase system capacity by adding more machines or nodes rather than upgrading a single machine.
I
- inline feedback buttons
- UI controls embedded in the workflow that let users quickly submit feedback during interaction.
J
- JSON-over-HTTP
- A communication pattern that sends JSON payloads over standard HTTP, often used for interoperable web APIs.
L
- language interoperability
- The ability for components written in different programming languages to communicate correctly using shared protocols.
- latency percentiles
- Performance metrics such as p50, p95, and p99 that show response time distribution across requests.
M
- Milvus
- An open-source distributed vector database designed for large-scale similarity search and horizontal scalability.
- minimum thresholds
- Required baseline values for metrics that must be met before a system is considered acceptable.
P
- Protocol Buffers
- A compact binary serialization format used with gRPC that supports efficient messaging and schema evolution.
R
- ReAct agent
- An agent architecture that alternates between reasoning steps and action steps to solve tasks iteratively.
- reasoning trace logging
- Recording intermediate reasoning steps, decisions, or actions to support debugging and system improvement.
- reranking
- A retrieval technique that reorders initially retrieved results using a stronger scoring model to improve precision.
- ROUGE
- A summarization evaluation metric that measures overlap between generated text and reference summaries.
S
- schema evolution
- The ability to change a data schema over time while maintaining compatibility between older and newer clients.
- shared memory workspace
- A central memory store accessible by multiple agents so they can coordinate through shared state and results.
- sliding window with summarization
- A context management strategy that keeps recent turns verbatim while summarizing older interactions to control token usage.
- stratified dataset
- A dataset constructed to include representative proportions of different case types, including common and edge cases.
- structured user feedback
- Feedback collected in a consistent format, such as ratings or labeled evaluations, so it can be compared and analyzed systematically.
T
- task decomposition
- A reasoning and planning method that breaks a complex goal into smaller executable sub-tasks.
V
- vector database
- A database optimized for storing and searching high-dimensional vector embeddings for retrieval tasks.
- vector database-backed long-term memory
- A long-term memory mechanism that stores embeddings externally so user preferences and past information can be retrieved across sessions.
- version metadata
- Descriptive information such as version number and effective date used to manage document revisions and validity.
- version-aware storage
- A storage design that keeps documents along with version metadata so different document states can be tracked over time.
W
- weighted scoring function
- A formula that combines multiple evaluation metrics using assigned weights to reflect business priorities.
About These Definitions
These definitions are loaded from the shared release pack. Use them with the study guide and practice questions to connect vocabulary to exam scenarios.