DBX-CEA Databricks Context Engineer Associate
- Vendor
- Databricks
- Level
- Associate
- Questions
- 45
- Time
- 90 min
- Domains
- 7
- In the bank
- 285
Practice 285 questionsTake a timed mock examDay-by-day study plan
Guides for this exam
- 8 Mistakes That Repeat Across Every Databricks Exam, and How to Spot Each One in the Question · 한국어
- DBX-CEA Study Guide: Treat the Context Window as a Budget and Every Question Gets Easier · 한국어
The exam at a glance
| Level | Associate |
|---|---|
| Questions | 45 |
| Time | 90 minutes (about 2 minutes per question) |
| Passing mark | Not published by the vendor |
Domains and weights
| Domain | Weight | Practice questions here |
|---|---|---|
| D1 · Foundations of Context Engineering | 16% | 46 |
| D2 · System Prompt and Instruction Design | 9% | 26 |
| D3 · Knowledge Retrieval and Genie Configuration | 20% | 57 |
| D4 · Memory Architecture with Lakebase and MLflow | 18% | 52 |
| D5 · Tool Design, MCP, and Agent Context | 13% | 37 |
| D6 · Context Compression and Compaction | 11% | 31 |
| D7 · Multi-Agent and Long-Horizon Task Design | 13% | 36 |
51 of the 285 practice questions (18%) ask for more than one answer.
D1 · Foundations of Context Engineering 16%
Foundation questions ask what belongs in an agent's context window and why it fails when overloaded. Look for duplicated, stale or irrelevant tokens in the trace. Strong options give the model less but better context; weaker ones add everything in case it helps.
What to watch for
- Inject only the memories and chunks relevant to the current request, not the whole store.
- Duplicate chunks from several retrievers waste context; deduplicate before passing them on.
- Raw tool output such as web pages can be summarized or kept outside the planning context.
Tasks in the official exam guide (7)
- 1.1 Match a described agent failure to the context management technique that addresses it
- 1.2 Identify proactive strategies (minimal tool sets, just-in-time retrieval, tool result scoping) that reduce context pressure
- 1.3 Diagnose context failure modes (poisoning, distraction, confusion, clash) from an agent trace
- 1.4 Select the right tool in the Databricks stack (Unity Catalog, Lakebase, MCP, MLflow 3) for a scenario
- 1.5 Find context elements that consume disproportionate attention budget and choose the change that improves focus
- 1.6 Select the reasoning mode (standard, extended or reduced thinking) from token budget and context impact
- 1.7 Identify where context length degrades retrieval or reasoning and choose the intervention that restores performance
D2 · System Prompt and Instruction Design 9%
Instruction design questions improve system prompts and Genie instructions. The deciding detail is whether an instruction is specific enough to test. Strong options state the exact format, scope or default and compare versions on the same evaluation data; weaker ones keep vague wording.
What to watch for
- Replace vague rules like "format dates correctly" with an exact format and an example.
- Compare prompt versions on the same evaluation set; different data makes the scores incomparable.
- Genie instructions should define default time ranges and say what is out of scope.
Tasks in the official exam guide (4)
- 2.1 Select and validate instructions, sample questions and trusted SQL assets for a production-ready Genie space
- 2.2 Choose few-shot examples to include or exclude given a token budget and marginal contribution
- 2.3 Select the targeted system prompt revision that fixes a failure pattern at the least token and maintenance cost
- 2.4 Decide from experiment tracking results whether a higher-token system prompt is justified
D3 · Knowledge Retrieval and Genie Configuration 20%
Retrieval and Genie questions ask why an agent gets the wrong knowledge and how to fix it at the source. Check what was indexed, what values Genie knows and whether retrieval was used. Strong options fix the corpus, metadata or instructions; weaker ones only tune the model.
What to watch for
- Keep scratch or draft folders out of the index so retrieval cites approved sources.
- Map business words to coded values, for example "open" to status P, in Genie instructions or synonyms.
- Reranking narrows many candidates to the few the model needs.
Tasks in the official exam guide (9)
- 3.1 Identify missing or poorly specified Unity Catalog metadata that causes an accuracy gap and pick the highest-impact fix
- 3.2 Select Unity Catalog objects (tables, views, parameterized queries, SQL functions) to curate into a Genie space
- 3.3 Diagnose the AI Search configuration behind retrieval quality problems and pick the remediation
- 3.4 Design a RAG pipeline over a Unity Catalog-governed document corpus that injects chunks into agent context
- 3.5 Select a chunking strategy from document structure, embedding model context length and query types
- 3.6 Select the context elements an agent needs to scope and execute a task
- 3.7 Choose between pre-inference retrieval and just-in-time agentic retrieval for a use case
- 3.8 Identify the retrieval failure mode from MLflow eval logs and UC metadata and pick the governance action that resolves it
- 3.9 Design a governance strategy that limits agent retrieval to authoritative sources before deployment
D4 · Memory Architecture with Lakebase and MLflow 18%
Memory questions separate session history from long-term memory and ask how much to recall. Read what must persist across conversations and what must be audited. Strong options store durable facts in long-term memory and recall a few relevant ones; weaker ones inject too many or set thresholds that miss them.
What to watch for
- Session history covers one conversation; long-term memory holds facts that matter in later ones.
- A full audit trail is a separate log from the memory the agent uses for personalization.
- Too many recalled memories confuse the model; too strict a threshold misses useful ones.
Tasks in the official exam guide (8)
- 4.1 Identify the mismatch between memory type and information need and select the correct memory strategy
- 4.2 Identify when a Delta-backed state object is required instead of an in-context scratchpad
- 4.3 Choose AI Search or structured query retrieval for memories persisted in Lakebase
- 4.4 Identify the most reliable context configuration from MLflow 3 experiment results across agent runs
- 4.5 Evaluate tradeoffs of static vs dynamic retrieval from Lakebase
- 4.6 Diagnose over-retrieval (context pollution) and under-retrieval risks in a memory system
- 4.7 Configure persistent agent memory across sessions with a Lakebase-backed durable store
- 4.8 Identify the pipeline stage where user intent should be resolved before context retrieval
D5 · Tool Design, MCP, and Agent Context 13%
Tool design questions ask how tools and MCP servers shape an agent's context. Look at tool names, descriptions and output size. Strong options give tools distinct purposes and small, relevant outputs; weaker ones return whole records or let similar tools overlap.
What to watch for
- Return only the fields the agent needs, with filters or pagination, instead of full records.
- Distinct names and descriptions stop an agent from using two tools interchangeably.
- Let the agent look up a tool's schema before calling it instead of guessing arguments.
Tasks in the official exam guide (6)
- 5.1 Apply progressive disclosure to MCP tool access to cut token use versus raw tool dumps
- 5.2 Find the description overlap that makes an agent confuse two MCP tools
- 5.3 Explain how progressive disclosure controls tool information in the context window at each stage
- 5.4 Evaluate which raw tool outputs can be cleared when message history nears capacity
- 5.5 Select the best Unity Catalog-registered tool for a task by functional fit and input/output compatibility
- 5.6 Identify capabilities to package as Agent Skills and the loading strategy that minimizes baseline context cost
D6 · Context Compression and Compaction 11%
Compression questions ask what to keep when history is compacted. Constraints, open goals and identifiers must survive; noise can go. Strong options define what compaction must preserve; weaker ones summarize freely and lose rules or ids.
What to watch for
- Keep user constraints, such as tables not to touch, verbatim through compaction.
- Strip repeated debug metadata from tool results before it fills the history.
- Preserve the current plan and next step so the agent continues after compaction.
Tasks in the official exam guide (5)
- 6.1 Identify the information wrongly discarded in compaction and revise the compaction prompt to preserve it
- 6.2 Tune a compaction prompt: maximize recall first, then improve precision
- 6.3 Decide whether hard-coded trimming heuristics suffice or a more sophisticated compaction is needed
- 6.4 Identify content that is safe to remove during compaction
- 6.5 Evaluate tradeoffs of aggressive vs conservative compaction
D7 · Multi-Agent and Long-Horizon Task Design 13%
Multi-agent questions ask how to split work and pass context between agents. Check what each handoff carries and whether shared rules reach every sub-agent. Strong options pass compact structured results and the user's constraints unchanged; weaker ones paraphrase or send full reasoning.
What to watch for
- Sub-agents should return short structured results, not their full reasoning.
- Pass the user's constraints verbatim to every sub-agent instead of paraphrasing them.
- Give parallel writers shared definitions, such as fiscal or calendar quarters, before they start.
Tasks in the official exam guide (6)
- 7.1 Diagnose multi-agent failures caused by insufficient shared context
- 7.2 Diagnose sub-agents receiving task messages instead of full traces at dispatch and pick the fix without growing their context
- 7.3 Identify the context propagation change that prevents conflicting outputs
- 7.4 Identify the sub-agent output design change that reduces orchestrator context load
- 7.5 Diagnose agent boundary placement failures (handoff compression overhead or context growth)
- 7.6 Identify the long-horizon strategy mismatch behind a failure and pick the replacement strategy
A study order that works
The 16-day study plan for DBX-CEA takes the domains in this order. Each day has a short reading, key terms and a quiz from the bank.
- Days 1–2: D1 · Foundations of Context Engineering
- Day 3: D2 · System Prompt and Instruction Design
- Days 4–6: D3 · Knowledge Retrieval and Genie Configuration
- Days 7–9: D4 · Memory Architecture with Lakebase and MLflow
- Days 10–11: D5 · Tool Design, MCP, and Agent Context
- Days 12–13: D6 · Context Compression and Compaction
- Days 14–15: D7 · Multi-Agent and Long-Horizon Task Design
- Day 16: Final Review: Decision Cheat-Sheet Across the Seven Domains
Recent changes to these questions
- : 1 question reworded, 3 answer choice sets changed, 3 answer keys corrected, 3 explanations rewritten
- : 60 questions added
- : 1 answer choice set changed, 1 answer key corrected, 1 explanation rewritten
- : Exam added to the bank (225 questions added)
About these questions
Every question is original, written from the public exam guide and checked against the vendor's own documentation. None comes from a real exam. Each one has an explanation of why the answer is right and why the other options are wrong. How the questions are written.