GenAI Associate Exam Glossary - 73 Terms

Search the terminology pack for Databricks Certified Generative AI Engineer Associate. Use these definitions with the study guide and practice questions.

Download App Study Guide Free Practice Exam

A

Agent Bricks: A Databricks feature set for solving problems with specialized agent patterns, including Knowledge Assistant, Multiagent Supervisor, and Information Extraction.
Agent Framework: A Databricks framework used to deploy and track LLMs or agents, including with AI Gateway.
AI Gateway: A Databricks capability that includes Inference Tables, Usage Tables, and rate limiting for tracking deployed LLMs or agents.
ai_query(): A Databricks function used to identify and handle batch inference workloads appropriately.
aliases: Named pointers used to promote prompts after they pass validation while preserving version history and enabling rollback. In the question, they are used with MLflow versions.

C

chunking: A data preparation strategy that splits documents into smaller pieces to fit document structure and model constraints for retrieval and generation.
CI/CD: Continuous integration and continuous delivery/deployment practices used here for updating indexes, promoting prompts, and testing components.
context length: The maximum token length an embedding model can handle; the exam includes choosing among context length options based on cost, latency, and quality.
custom Scorers: Databricks scoring tools used to evaluate agents and LLMs.

D

Databricks App: An application built on Databricks that can provide an interactive user-facing interface. In the question, it is used for customer support agents asking questions and receiving grounded answers.
Databricks Certified Generative AI Engineer Associate: A Databricks certification exam that assesses the ability to design and implement LLM-enabled solutions using Databricks, including RAG applications, LLM chains, model selection, governance, deployment, and monitoring.
Databricks Secrets: A Databricks feature for securely storing sensitive values such as API keys. In the question, it is used to store the external MCP server API key.
Delta Lake: A Databricks storage layer used here as the destination for writing chunked text tables in Unity Catalog.

E

embedding dimension: The size of the vector representation produced by an embedding model.
embedding model: A model that converts text into vector representations; the exam covers selecting one based on context length, source documents, queries, and optimization strategy.
embeddings: Vector representations of items used for similarity search. The question uses the number of embeddings as one of the factors for configuring vector search.
evaluation: The process of assessing model or application outputs, including qualitative assessment, scoring, and comparison of metrics.
external MCP server: An MCP server hosted outside Databricks that requires connection details and, in the question, an API key stored in Databricks Secrets.

F

feature store: A store for structured features used in the sample question as an alternative data source for transaction-based information.
fine tune: To train a custom model further on task-specific data. In the question, it refers to fine-tuning a custom embedding model for vector search.
fine-tune: To update a model using a tuning dataset so it better reflects desired or current information.
Foundation Model APIs: Databricks APIs used to serve LLM applications leveraging foundation models.

G

ground truth: Reference-correct answers required by some evaluation judges.
guardrails: Controls used to prevent negative outcomes from LLM applications, including protection against malicious inputs and unsafe outputs.

H

Hugging Face Transformers: A related online tool/service used for working with transformer-based models in generative AI applications.
hybrid search: A search setup that combines vector-based retrieval with other retrieval methods, typically keyword-based search, to improve results. In the question, it can be enabled together with reranking.

I

Inference Tables: Tables used to track inference activity for deployed models or agents.
Information Extraction: An Agent Bricks option used to extract structured information from content.

K

Knowledge Assistant: An Agent Bricks option used to solve problems by providing knowledge-based assistance.

L

LangChain: A tool used in generative AI applications for building chains and related workflows.
latency: The time delay before a search or model response is returned. In the question, it is the most critical metric for the search workload.
legal/licensing requirements: Legal and licensing constraints on data sources that must be followed to avoid legal risk in GenAI applications.
LLM: A large language model used in generative AI applications; the exam expects knowledge of current LLMs, their capabilities, and how to select them for tasks.
LLM chains: Multi-step application flows that combine an LLM with other components such as tools, retrievers, or prompts to produce an output.
LLM-as-a-judge: An evaluation approach where a language model scores or judges model outputs instead of, or in addition to, human raters. In the question, it is proposed as a way to rescore responses.

M

managed MCP server: An MCP server provided and operated by Databricks for a supported data source. In the question, it is configured by specifying the server type as managed and providing the server identifier.
masking techniques: Data protection techniques used as guardrails to meet performance objectives while limiting exposure of sensitive data.
MCP: Model Context Protocol. In the question, MCP servers are integrated to give an agent access to external and managed data sources.
MCP server: A server that exposes tools or data sources to an agent through the Model Context Protocol. In the question, both managed and external MCP servers are used.
MCP servers: Servers that can be integrated as managed, external, or custom components based on application requirements.
MLflow: A Databricks-supported tool for managing the solution lifecycle, and for developing, evaluating, tracing, and monitoring models and agents.
mlflow.genai.evaluate(): An MLflow GenAI evaluation function used for consistent agent evaluation with aligned judgments. In the question, it is the place where calibrated SME judgments are applied.
model card: Documentation describing a model’s attributes and metadata used to help select the best model for a task.
model hub: A repository or marketplace from which models can be selected for a task based on metadata or model cards.
Model Serving: A Databricks capability for deploying models and solutions through serving endpoints.
model signature: The defined input and output schema for a model, used when assembling a RAG application.
Mosaic AI Agent endpoint: An endpoint that an app can call to interact with a Mosaic AI agent. In the question, it must be called from the backend without exposing long-lived browser tokens.
Mosaic AI Vector Search: A Databricks vector search offering whose key concepts and components are covered by the exam.
Multiagent Supervisor: An Agent Bricks option used to coordinate multiple agents in solving a problem.

P

PAT: Personal access token. In the question, it is a Databricks token that should not be stored in browser JavaScript because it would expose long-lived credentials.
prompt engineering: The practice of designing prompts to elicit desired model behavior, including specific formatting and output changes.
prompt generation: The creation of prompts for generative AI applications.
prompt templates: Reusable prompt structures used to standardize how prompts are authored and managed across environments. In the question, they must be promoted from development to staging to production with version history and rollback support.
prompt version control: The practice of tracking and managing prompt changes across versions and environments.
pyfunc: A model flavor used to code a chain with pre-processing and post-processing.
Python: A programming language the exam expects candidates to use for RAG applications, agent development, and related APIs and libraries.

R

RAG: Retrieval-augmented generation, an application pattern that combines retrieved source context with generation to build performant applications.
rate limiting: A control that limits the rate of requests to a deployed LLM or agent.
re-ranking: A retrieval step that reorders retrieved results to improve the information retrieval process.
reranking: A post-retrieval step that reorders candidate search results to improve relevance. In the question, it is an optional setting for vector search.
rubrics: Defined scoring criteria used to evaluate outputs consistently. In the question, clear rubrics are used to calibrate SMEs and reduce rater inconsistency.

S

Sentencizer: A text-processing component that splits text into sentences. In the question context, it is presented as a possible NLP task category, but the text does not further define it.
SME: Subject matter expert. In the question, SMEs are the domain experts who review sampled answers and need calibration for consistent judgments.
SME feedback: Feedback from subject matter experts used to improve agent performance.
standard vector search: A vector search configuration that is not storage optimized. In the question, it is contrasted with storage optimized vector search.
storage optimized vector search: A vector search configuration optimized for storage efficiency. In the question, it is one of the setup options for handling a very large inventory.
subject matter expert: An expert whose feedback is used to improve agent performance.
summarization: An NLP task category that reduces a longer text to a shorter gist while preserving intent.

T

Text Classification: A natural language processing task category where a model assigns one or more labels to text. In the question context, it is presented as a possible evaluation category for the memo-to-gist application.
text2text generation: A natural language processing task category where a model transforms input text into output text. In the question context, it is one of the candidate task categories for converting a paragraph-long memo into a single-sentence gist.

U

Unity Catalog: A Databricks governance layer used for data governance and for registering models and storing chunked text in tables.
Usage Tables: Tables used to track usage for deployed models or agents.

V

Vector Search: A Databricks tool for semantic similarity search and retrieval, used to create and query vector search indexes and support RAG applications.

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.

Download App Read the full study guide Take the free practice exam