Databricks GENAI-ASSOC Glossary: RAG, Agents, and Evaluation Terms

April 13, 2026

Databricks GENAI-ASSOC glossary of RAG, agents, evaluation, guardrails, and deployment terms.

On this page

Use this glossary when retrieval, embeddings, evaluation, and deployment terms start to blur together. Keep it beside the cheat sheet and resources instead of using it as a substitute for scenario review.

High-yield terms

Term	Short meaning	Why it matters on GENAI-ASSOC
Agent Bricks	Databricks packaged agent-building components for specific reasoning patterns	Newer current-guide topic
Agent Framework	Databricks framework layer for building and operating agents	Newer current-guide topic
RAG	Retrieval-augmented generation, where retrieved context is supplied to the model before generation	Core system pattern for the exam
Chunking	Splitting source content into retrieval-sized pieces before indexing	Major driver of retrieval quality and cost
Embedding	Numeric vector representation used to compare semantic similarity	Core retrieval representation
Vector search	Retrieval over embeddings to find semantically similar content	Central Databricks tool and concept
Reranking	Secondary scoring step that refines the retrieved result ordering	Important retrieval-quality improvement step
Grounding	Tying model output to retrieved or trusted source material	Helps reduce unsupported answers
Hallucination	Confident-sounding model output that is unsupported or incorrect	Core failure mode
Context window	Maximum input token space available to the model	Important constraint in retrieval and prompt design
Prompt template	Structured prompt pattern reused across requests	Core chain-building concept
Evaluation set	Fixed collection of test prompts and expected behaviors used for quality checks	Central evaluation concept
Guardrail	Safety, policy, or formatting control applied around model behavior	Runtime safety and policy concept
Latency budget	Maximum acceptable end-to-end response time for the application	Deployment and serving trade-off concept
Model serving	Databricks endpoint-based serving for models and chains	Core deployment concept
Foundation Model APIs	Databricks-hosted API path for using supported foundation models	Important serving-path concept
MLflow	Lifecycle tooling for experiments, models, and deployment packaging	Key Databricks platform concept
Unity Catalog	Governance layer for data, permissions, and lineage	Core governance concept
Inference logging	Capturing serving requests or responses for review and monitoring	Important evaluation and monitoring concept
Inference table	Structured Databricks table used to review live inference behavior	Important monitoring surface
Agent Monitoring	Databricks monitoring surface for deployed agent behavior	Current monitoring topic
AI Gateway	Databricks gateway layer with usage, logging, and rate-control features for LLM or agent access	Current deployment and monitoring topic
MCP server	Model Context Protocol server that exposes tools, data, or prompts to agents	Newer current-guide integration topic
Masking	Hiding or redacting sensitive content before or during use	Governance and safety concept
Model card	Metadata and documentation describing a model’s behavior and limits	Important model-selection aid
Pyfunc model	MLflow packaging format that can wrap model logic with pre- and post-processing	Important deployment concept
Prompt version control	Managed tracking of prompt changes across environments or releases	Current deployment-lifecycle topic

Commonly confused pairs

Pair	Keep this distinction clear
chunking vs embedding	splitting documents versus converting them into vectors
retrieval vs reranking	initial candidate fetch versus refined ordering
grounding vs fine-tuning	using source context at request time versus changing model behavior through training
hallucination vs retrieval miss	unsupported generation versus failure to fetch the right context
guardrail vs evaluation	runtime control versus quality-measurement process
vector search vs model serving	retrieving relevant context versus serving the chain or model
MLflow vs Unity Catalog	lifecycle tooling versus governed data or object layer
context window vs chunk size	model input capacity versus size of indexed content units
Agent Framework vs MLflow	agent runtime framework versus lifecycle and packaging tooling
inference logging vs inference tables	capture of live traffic versus tabular monitoring surface
AI Gateway vs model serving	gateway and traffic controls versus the serving endpoint itself
MCP server vs tool inside a chain	standardized tool surface versus the tool logic you expose through it

If three terms blur together

Cluster	Fast separation
chunking / embeddings / vector search	split the content, represent it numerically, then retrieve it
retrieval / reranking / generation	fetch candidates, refine them, then produce the answer
grounding / hallucination / evaluation	improve factual support, detect unsupported output, and measure quality
MLflow / model serving / Unity Catalog	lifecycle tooling, serving path, or governance boundary
guardrails / masking / permissions	runtime safety, content protection, or access control
inference logging / inference tables / Agent Monitoring	live request capture, tabular evidence, or monitoring surface
Agent Bricks / Agent Framework / MCP	packaged agent component, agent-building framework, or tool integration surface

One-sentence memory hooks

If the problem is wrong source chunks, think chunking and retrieval quality before prompt tweaks.
If the problem is unsupported answers, think grounding and evaluation, not only bigger models.
If the problem is runtime safety, think guardrails and masking.
If the problem is serving the chain, think model serving.
If the problem is tracking and packaging the lifecycle, think MLflow.
If the problem is tool integration, think MCP.
If the problem is live traffic control and usage review, think AI Gateway.

Operational clusters worth keeping straight

Cluster	What it usually signals on the exam
chunking / embeddings / vector search	retrieval-system design questions
prompt templates / chain components / model choice	application-design questions
evaluation sets / inference logging / monitoring metrics	evaluation and monitoring questions
guardrails / masking / licensing / permissions	governance and safety questions
MLflow / model serving / Unity Catalog	Databricks platform-integration questions
Agent Bricks / Agent Framework / MCP	current Databricks agent-building questions

If the confusion is really about…

Topic family	Best page to revisit
retrieval and deployment patterns	Cheat Sheet
current Databricks facts and docs	Resources
pacing and review order	Study Plan
overall exam framing	Guide root

Revised on Sunday, May 10, 2026

Resources