Databricks Context Engineer Associate Practice Exam

• A. Context poisoning: an incorrect delivery claim was saved and reused as fact
• B. Context distraction: accumulated history pulls the agent toward past replies
• C. Context confusion: surplus tool definitions steer the agent to the wrong output
• D. Context clash: two retrieved policy versions disagree in the same window Correct answer

Option D is correct because the trace shows two retrieved versions of the same policy (v1 says 2 business days, v2 says 5) sitting in the window together, so the agent answers differently at T9 and T14. That is context clash: conflicting information in one context. Option A would need a wrong claim saved and reused as fact, option B needs history steering behaviour, and option C needs surplus tool definitions, and the trace shows none of these.

• A. Extended thinking with a large thinking-token budget
• B. Standard mode with a chain-of-thought instruction added to the prompt
• C. Extended thinking on a 10% sample of messages and standard mode on the rest
• D. Reduced thinking with a low thinking-token cap Correct answer

Option D is correct because the task is a simple 12-way label on 2 million messages a day with a 400 ms limit and minimal token spend. Thinking tokens count as output tokens, so a low thinking cap keeps both latency and cost down. Option A and option B add reasoning tokens to every message, and option C still pays that cost on a sample while giving the other 90% no benefit.

• A. Superfluous tools or documents steer the output off course
• B. Two contradictory pieces of information sit in one window
• C. A long history is over-weighted against the current request
• D. An incorrect claim persists and is reused in later steps Correct answer

Option D is correct because context poisoning is a wrong claim that enters the window and is then reused by later steps. Superfluous material is confusion, contradictory items are clash, and over-weighting a long history is distraction.

• A. The escalate example and the greeting example
• B. The issue_refund example and the escalate example Correct answer
• C. The simple search_orders example and the issue_refund example
• D. The two search_orders examples with different filters

Option B is correct because few-shot examples should target the observed failures: the agent never calls escalate and malforms issue_refund arguments. One well-formed issue_refund example and one escalate example fix both within the 400-token room. search_orders already succeeds, so the search examples in options C and D waste space, and the greeting in option A teaches nothing about either failing tool.

• A. Notebook source code, job schedules and alert rules
• B. Foundation model weights, prompt templates and vector indexes
• C. Text instructions, sample questions and trusted SQL assets Correct answer
• D. Table comments, column lineage graphs and access grants

Option C is correct because a Genie space is configured with text instructions, sample questions and trusted SQL assets such as example queries. Notebook code, model weights, vector indexes and lineage graphs are not part of a space's configuration.

• A. A parameterized example query holding the verified net revenue SQL Correct answer
• B. A table comment that states the net revenue formula in prose
• C. A longer text instruction that repeats the join rules in more detail
• D. Ten more sample questions that mention refunds in different words

Option A is correct because a parameterized example query holding the verified SQL gives Genie the exact joins and filters to reuse, so the refunds join is no longer omitted. Prose in a comment, longer instructions or more sample questions had already failed or stay open to interpretation.

• A. Split each sentence into its own chunk
• B. Split on section headings, then split any section above 512 tokens Correct answer
• C. Split at fixed 4,000-token windows with no overlap
• D. Embed each manual as one chunk and let the index rank passages

Option B is correct because the manuals have numbered sections under headings and most questions are answered by one section, so heading-based chunks keep each answer's context together. Splitting any section above 512 tokens keeps every chunk within the embedding model's input limit. Sentence chunks lose context, 4,000-token windows exceed the 512-token limit, and a whole manual cannot be embedded as one input.

• A. A scheduled job copies the most-used documents into the system prompt each morning
• B. The agent reads the whole corpus into its context window at session start
• C. The agent calls a retrieval tool mid-run and loads results when a step needs them Correct answer
• D. An embedding search runs before the model call and injects top-k chunks into the prompt

Option C is correct because just-in-time agentic retrieval means the agent decides during the run to call a retrieval tool and loads results only when a step needs them. Option D describes pre-inference retrieval, where a search runs before the model call and injects top-k chunks. Options A and B load content up front, on a schedule or at session start, rather than on demand.

• A. Add all 31 tables across two spaces, splitting them 16 and 15
• B. Add the 25 largest tables and tell Genie in an instruction to ignore staging copies
• C. Add the 18 domain tables and leave out the staging copies and logs Correct answer
• D. Add the 27 tables that are not logs and tell Genie in an instruction to prefer gold copies

Option C is correct because the 18 domain tables answer the business questions, and leaving out the staging copies and logs removes the duplicate sources that make answers vary. The 50-table limit is not the constraint here, since even all 31 tables fit, so the choice rests on curation. Options B and D keep the staging copies and rely on instructions, which do not reliably stop Genie from choosing them, and option A splits the domain across two spaces without fixing the ambiguity.

• A. In an in-context scratchpad the agent updates after each clause
• B. In a Delta table keyed by task ID, read back at the start of each run Correct answer
• C. In the system prompt, rewritten by a deployment each night
• D. In the first job's conversation history, passed in full to later jobs

Option B is correct because progress that must survive restarts across three days and be inspected by humans needs durable, external state. A Delta table keyed by task ID can be read back at the start of each run and queried by reviewers. An in-context scratchpad (option A) is lost when a run ends, a system prompt rewritten nightly (option C) is not a store reviewers can query, and passing a full conversation history (option D) bloats context and is hard to inspect.

• A. Top-k 4 with a minimum score of 0.80 Correct answer
• B. Top-k 5 with a minimum score of 0.60
• C. Top-k 8 with a minimum score of 0.50
• D. Top-k 2 with a minimum score of 0.50

Option A is correct because only the first three scores (0.91, 0.88, 0.84) are relevant and the next is 0.62. A minimum score of 0.80 admits exactly those three, 600 tokens, within the 800-token budget, and the top-k of 4 is a cap that is not reached. Option B admits the unrelated 0.62 memory, and option C admits 0.62, 0.58 and 0.55, so both inject unrelated memories (option C also exceeds the budget), and option D with top-k 2 drops the third relevant memory.

• A. Return the memory for a given memory_id with the id as the AI Search query
• B. Find notes that restate a preference in other words with an AI Search query Correct answer
• C. Load the row for a known memory_id with a SQL query on the key column Correct answer
• D. Rank a user's memories by relevance to the request with SQL ORDER BY created_at
• E. Find notes on a topic given in new words with a SQL LIKE filter on keywords

Options B and C are correct because AI Search is suited to semantic matching, such as finding notes that restate a preference in different words, while a known memory_id is an exact key lookup that a SQL query on the key column does best in Lakebase. Option A uses a similarity search for an exact key, option D orders by created_at which ranks by recency, not relevance, and option E uses a keyword LIKE filter which misses notes phrased in new words.

• A. Merge both tools into one get_customer_records tool with a record_type input
• B. Rename get_invoice to get_billing_documents and keep both descriptions as written
• C. Remove billing and payment details from the get_customer_orders description Correct answer
• D. Add a system prompt rule telling the agent to call get_invoice for billing questions

Option C is correct because the get_customer_orders description claims billing and payment details, which overlaps with get_invoice and pulls invoice requests to the wrong tool. Since the tool descriptions are reused by other agents, the smallest fix is to narrow that one description to what the tool really returns. Merging tools (option A) and renaming (option B) are larger changes that leave or add overlap, and a system prompt rule (option D) fixes only this agent and leaves the misleading description in place.

• A. Split the tools between two agents by the alphabetical order of their names
• B. Replace tool schemas with one summary per MCP server for the whole run
• C. List tool names with one-line summaries and load a full schema on demand Correct answer
• D. Load the full schema of each MCP tool when the session starts

Option C is correct because progressive disclosure for MCP tools means the agent first sees tool names with short summaries and loads the full schema only when it needs a particular tool. This keeps context small. Option D loads every schema at the start, option B removes per-tool detail for the whole run so the agent cannot call tools reliably, and option A just splits tools between agents without disclosing anything progressively.

• A. Load a tool's schema after the model selects that tool Correct answer
• B. Cache the tool list in a Delta table and send it with each call
• C. Show one summary per server first, then list the tools of the chosen server Correct answer
• D. Copy the four servers' tool descriptions into the system prompt
• E. Connect the Databricks SQL server and drop the other three

Options A and C are correct because both defer detail: loading a schema only after the model selects the tool, and showing one summary per server before listing its tools, cut the 10,000-token baseline while all four servers stay reachable. Connecting only the Databricks SQL server and dropping the other three (E) loses coverage, and caching the list in Delta (B) or copying descriptions into the prompt (D) still sends the same tokens.

• A. Error messages from retries that later succeeded
• B. Verbatim tool outputs from earlier file reads
• C. Paths of the files the agent has already edited
• D. Rejected options and the reasons for rejecting them Correct answer

Option D is correct because the summary kept the files edited and the current task but lost why schema B was rejected (the latency constraint), so the agent proposed it again. Rejected options and their reasons are exactly the kind of decision rationale that compaction should preserve. Options A, B and C describe information that is either still present (file paths) or safe to drop (verbose outputs, resolved retry errors).

• A. It keeps cost the same and improves fidelity by summarizing at each turn
• B. It raises token cost and keeps more of the original wording of tool results
• C. It raises fidelity and raises the number of compaction calls
• D. It lowers token cost and risks dropping subtle details later steps need Correct answer

Option D is correct because aggressive compaction summarizes heavily, which lowers token cost but risks dropping subtle details that later steps need. Conservative compaction keeps more detail and costs more tokens. Options A, B and C each claim that aggressive compaction raises fidelity or keeps more original wording, which is the opposite of what heavier summarizing does.

• A. The user's wording of the original request
• B. Verbose output of the earlier CREATE statements
• C. Superseded drafts of the schema definition
• D. Completed steps and the outcome of each Correct answer
• E. Identifiers the user supplied, such as the account ID Correct answer

Options D and E are correct because asking for the account ID again points to a lost identifier, and re-creating the schema points to lost completed steps. The user's wording, verbose CREATE output and superseded drafts can be dropped without causing these symptoms.

• A. Add the fiscal calendar and approved tables to each task message Correct answer
• B. Have each sub-agent ask the user for the missing details mid-run
• C. Move the sub-agents to a model with a 200,000-token window
• D. Send each sub-agent the orchestrator's full trace as well

Option A is correct because sub-agents only know what the orchestrator puts in their task message, so the fiscal calendar and approved tables the user gave earlier must be passed along in that message. This adds a few lines without growing their windows. Option D sends the full trace and bloats their context, option C adds window size but not the missing information, and option B interrupts the user from inside sub-agents.

• A. The extract-to-classify boundary compresses away detail classify needs Correct answer
• B. The report agent receives unfiltered outputs from three upstream agents
• C. The classify and validate agents repeat the same checks on each clause
• D. The extract agent accumulates more context than one window can hold

Option A is correct because the extract agent squeezes a 40,000-token contract into a 150-word summary, so classify never sees clauses that exist only in the full text. That is a boundary placed where too much detail is compressed away. The other agents run well within their windows, which rules out option D, and the stem gives no sign of unfiltered upstream outputs (option B) or repeated checks (option C).