I85PREP
⌘ K
Databricks Context Engineer Associate study materials
Databricks Context Engineer Associate study resources

Databricks Context Engineer Associate practice questions

One question at a time. Build your Databricks Context Engineer Associate understanding.

First 50 of 200 questions. Each one lists the choices, the correct answer, and an explanation. The checker at the top of this page saves progress in this browser when it is available.

Continue in the practice tool above to load more questions, 50 at a time.

Question 1Context Engineering Foundations and the Databricks Platform

In the Databricks context engineering workflow, what is the primary purpose of a retrieval-augmented generation (RAG) pipeline when building an application on the platform?

  1. To reduce the token cost of the system prompt by compressing it automatically
  2. To replace the need for a vector store by caching all documents in memory
  3. To ground model responses in retrieved, domain-specific documents rather than relying only on parametric knowledge Correct answer.
  4. To fine-tune the base model weights on proprietary documents before deployment

Explanation

RAG retrieves relevant chunks at query time and injects them into the prompt so the model answers from grounded evidence. It does not change model weights, which is the role of fine-tuning. This grounding improves factual accuracy and lets the application cite current, private data.

Question 2Context Engineering Foundations and the Databricks Platform

A team wants the lowest-latency way to serve a foundation model endpoint inside Databricks for a chat application. Which platform component is designed for this purpose?

  1. Databricks SQL warehouses
  2. Delta Live Tables pipelines
  3. Mosaic AI Model Serving endpoints Correct answer.
  4. Unity Catalog metastores

Explanation

Model Serving endpoints expose foundation models and custom models as REST APIs with managed scaling and low-latency inference. SQL warehouses are for analytical queries, not model inference. Delta Live Tables and metastores handle data pipelines and governance, not serving.

Question 3Context Engineering Foundations and the Databricks Platform

Which Databricks feature provides a centralized governance layer for catalogs, tables, models, and other data assets with fine-grained access control?

  1. Databricks Repos
  2. Delta Live Tables
  3. Unity Catalog Correct answer.
  4. Auto Loader

Explanation

Unity Catalog is the unified governance layer that manages catalogs, schemas, tables, volumes, and models with granular privileges. Delta Live Tables is a pipeline framework, Repos manages source code, and Auto Loader ingests files. Governance of assets across the workspace is the defining role of Unity Catalog.

Question 4Context Engineering Foundations and the Databricks Platform

When chunking documents for a RAG application, why is chunk overlap commonly used?

  1. To preserve context that would otherwise be split across chunk boundaries Correct answer.
  2. To reduce the total number of embeddings that must be stored
  3. To guarantee that every chunk fits within the embedding model's maximum input length
  4. To eliminate the need for a reranking step after retrieval

Explanation

Overlap repeats text at chunk edges so sentences or ideas cut by a boundary remain recoverable in at least one chunk. It increases, not decreases, stored embeddings. It does not guarantee length limits or remove the value of reranking.

Question 5Context Engineering Foundations and the Databricks Platform

A developer wants to store and query vector embeddings natively inside a Delta table so that similarity search stays close to the governed data. Which Databricks capability supports this?

  1. Photon vectorized query engine
  2. Vector Search indexes built on Delta tables Correct answer.
  3. Lakehouse Federation connectors
  4. Delta Sharing

Explanation

Databricks Vector Search creates indexes over Delta tables and supports similarity search while keeping data governed by Unity Catalog. Photon accelerates SQL execution, federation queries external systems, and Delta Sharing exchanges data across organizations. None of those provide native embedding similarity indexes.

Question 6Context Engineering Foundations and the Databricks Platform

What is the main role of a system prompt in an LLM-powered Databricks application?

  1. To store the user's conversation history for later retrieval
  2. To configure the cluster size used by the serving endpoint
  3. To embed the retrieved documents into the vector index
  4. To set persistent instructions and behavioral constraints for the model Correct answer.

Explanation

The system prompt conveys durable instructions such as persona, tone, and guardrails that shape every response. It is not a storage location for history, an indexing mechanism, or a cluster configuration. Separating these concerns keeps prompt design clean and predictable.

Question 7Context Engineering Foundations and the Databricks Platform

Which statement best describes the relationship between a Databricks workspace and a Unity Catalog metastore?

  1. A metastore exists only inside a single cluster and disappears when it terminates
  2. A metastore can be attached to multiple workspaces to share governed assets Correct answer.
  3. A workspace can attach to multiple metastores simultaneously within one region
  4. Each workspace must contain its own isolated metastore with no sharing

Explanation

A Unity Catalog metastore is a regional governance container that can be attached to multiple workspaces, enabling shared catalogs and access policies. It is not per-workspace or per-cluster. A workspace attaches to one metastore at a time.

Question 8Context Engineering Foundations and the Databricks Platform

In a RAG pipeline, what does an embedding model produce for each document chunk?

  1. A relevance score comparing the chunk to every possible query
  2. A token count used to bill the serving endpoint
  3. A numeric vector that captures the chunk's semantic meaning Correct answer.
  4. A compressed summary of the chunk in natural language

Explanation

Embedding models map text into dense numeric vectors positioned so that semantically similar texts are close together. These vectors are stored in an index and compared with the query embedding at retrieval time. They are not summaries, precomputed relevance scores, or billing metrics.

Question 9Context Engineering Foundations and the Databricks Platform

A team notices the retrieval step of their RAG application returns many loosely related chunks, lowering answer quality. Which change most directly improves retrieval precision?

  1. Increasing the model's temperature setting
  2. Adding a reranking step after initial vector retrieval Correct answer.
  3. Shortening the system prompt to save tokens
  4. Reducing the number of documents in the source corpus

Explanation

A reranker scores the initially retrieved candidates against the query and reorders them, so only the most relevant chunks reach the prompt. Temperature affects generation randomness, not retrieval. Corpus size and system prompt length do not directly sharpen relevance ranking.

Question 10Context Engineering Foundations and the Databricks Platform

Which Databricks approach lets you point a foundation model endpoint at a custom fine-tuned model while keeping the same serving interface used for base models?

  1. Registering the model in Unity Catalog and deploying it to a Model Serving endpoint Correct answer.
  2. Creating a SQL warehouse and calling the model from a dashboard
  3. Mounting the model files to DBFS and loading them in a notebook
  4. Publishing the model to a Delta Live Tables pipeline

Explanation

Models logged and registered in Unity Catalog can be deployed to Model Serving endpoints, exposing a consistent REST interface for inference. SQL warehouses, DBFS mounts, and pipeline frameworks are not serving mechanisms. Registering the model also keeps it governed.

Question 11Context Engineering Foundations and the Databricks Platform

A developer needs to keep a Vector Search index synchronized as new documents arrive in a Delta table. Which design choice supports near-real-time freshness?

  1. A batch job that re-embeds the entire corpus hourly
  2. A snapshot copy of the table written to a new schema on each update
  3. A Delta Sync index that automatically tracks changes to the source table Correct answer.
  4. A manually triggered full rebuild of the index each night

Explanation

Delta Sync indexes automatically update as the underlying Delta table changes, keeping the index fresh without manual rebuilds. Full rebuilds and full re-embeds are expensive and slow. Copying snapshots does not update the index.

Question 12Context Engineering Foundations and the Databricks Platform

When designing prompts for a RAG application, why should retrieved context be clearly delimited and labeled in the prompt?

  1. It guarantees the model will never hallucinate
  2. It reduces the number of tokens the model consumes
  3. It helps the model distinguish source material from user instructions Correct answer.
  4. It allows the embedding model to index the prompt at runtime

Explanation

Clear delimiters and labels separate trusted context from user input, helping the model follow instructions and resist injection. This practice improves reliability but does not guarantee zero hallucination. It also does not reduce tokens or trigger indexing.

Question 13Context Engineering Foundations and the Databricks Platform

A team wants to evaluate whether their RAG application's answers are faithful to retrieved documents. Which evaluation approach is most appropriate?

  1. Tracking the number of workspace users
  2. Measuring endpoint latency percentiles
  3. Counting the number of chunks stored in the index
  4. Scoring answers for groundedness against the retrieved context Correct answer.

Explanation

Groundedness or faithfulness scoring checks whether claims in an answer are supported by the retrieved context, which is central to RAG quality. Latency, index size, and user counts measure operational or scale metrics, not answer faithfulness.

Question 14Context Engineering Foundations and the Databricks Platform

Which Unity Catalog object is the appropriate place to store non-tabular files, such as PDFs used as RAG source documents, with governed access?

  1. A managed volume Correct answer.
  2. A job definition
  3. A cluster library
  4. A notebook revision

Explanation

Volumes are Unity Catalog objects designed for governed storage of non-tabular files like documents and images. Notebook revisions, cluster libraries, and job definitions are code and orchestration artifacts, not governed file storage.

Question 15Context Engineering Foundations and the Databricks Platform

A developer configures a Model Serving endpoint and observes that bursts of traffic cause request failures. Which configuration most directly addresses this?

  1. Reducing the number of retrieved chunks per query
  2. Increasing the endpoint's provisioned concurrency or scaling limits Correct answer.
  3. Lowering the model temperature
  4. Changing the workspace's default catalog

Explanation

Provisioned concurrency and scaling settings determine how many concurrent requests an endpoint can handle, directly affecting burst capacity. Temperature, chunk counts, and catalog defaults do not control serving throughput.

Question 16Context Engineering Foundations and the Databricks Platform

What distinguishes few-shot prompting from zero-shot prompting in an LLM application?

  1. Few-shot requires fine-tuning; zero-shot does not
  2. Few-shot only works with embedding models, not generative models
  3. Zero-shot always produces higher accuracy than few-shot
  4. Few-shot includes example input-output pairs in the prompt; zero-shot provides none Correct answer.

Explanation

Few-shot prompting supplies a small number of demonstrations to guide format and behavior, while zero-shot relies only on instructions. Neither requires fine-tuning, and few-shot often improves accuracy over zero-shot rather than the reverse.

Question 17Context Engineering Foundations and the Databricks Platform

A data engineer wants to ingest continuously arriving JSON files into a Delta table as the source for a RAG pipeline. Which Databricks feature is designed for incremental file ingestion?

  1. Databricks Assistant
  2. Auto Loader Correct answer.
  3. Photon
  4. Delta Sharing

Explanation

Auto Loader incrementally and efficiently ingests new files as they land in cloud storage, tracking which files have been processed. Delta Sharing exchanges data across organizations, Photon accelerates SQL, and the Assistant helps author code. Only Auto Loader targets incremental file ingestion.

Question 18Context Engineering Foundations and the Databricks Platform

When selecting a chunk size for a RAG corpus of long technical manuals, which trade-off should guide the decision?

  1. Chunks must always equal the embedding model's maximum input length
  2. Larger chunks capture more context but dilute relevance and consume more prompt tokens Correct answer.
  3. Chunk size has no effect on retrieval quality once embeddings exist
  4. Smaller chunks always improve answer quality regardless of content

Explanation

Chunk size balances context completeness against retrieval precision and token budget. Larger chunks retain more surrounding context but can mix topics and crowd the prompt. There is no universal best size, and chunks need not equal the model maximum.

Question 19Context Engineering Foundations and the Databricks Platform

A team wants to log prompts and responses from their RAG application for later analysis and quality monitoring within Databricks. Which combination best supports this?

  1. Storing responses only in the serving endpoint's memory
  2. Printing outputs to notebook cells and reviewing them manually
  3. Emitting responses to cluster logs and relying on log rotation
  4. Writing inference records to a Delta table and querying them with SQL Correct answer.

Explanation

Persisting inference records to a Delta table makes prompts and responses queryable, auditable, and available for evaluation dashboards. Notebook prints and endpoint memory are ephemeral, and cluster logs are not a durable analytical store.

Question 20Context Engineering Foundations and the Databricks Platform

Which statement accurately describes how Unity Catalog privileges flow through the catalog hierarchy?

  1. Privileges granted on a table automatically apply to unrelated catalogs
  2. Privileges granted on a catalog can be inherited by schemas and tables within it Correct answer.
  3. Privileges are assigned only at the workspace level and ignore catalogs
  4. Privileges must be granted separately on every table with no inheritance

Explanation

Unity Catalog uses a three-level namespace of catalog, schema, and table, and privileges granted at a higher level can be inherited by objects beneath it. This inheritance reduces administrative overhead. Privileges do not cross into unrelated catalogs.

Question 21Context Engineering Foundations and the Databricks Platform

An application must answer questions over data that changes frequently, and stale answers are costly. Which architecture decision best supports freshness?

  1. Caching all answers in the system prompt at deployment
  2. Relying solely on the base model's pretrained knowledge
  3. Retrieving from a continuously synchronized index over the live Delta table Correct answer.
  4. Embedding the corpus once and never updating the index

Explanation

A continuously synchronized index keeps retrieved context aligned with the current table state, so answers reflect recent changes. One-time embedding, static prompt caching, and pretrained knowledge all go stale as the underlying data evolves.

Question 22Context Engineering Foundations and the Databricks Platform

A RAG system retrieves relevant chunks yet still produces answers that ignore them and rely on the model's prior beliefs. Which prompt-engineering adjustment most directly addresses this?

  1. Raising the temperature to encourage creativity
  2. Increasing the number of few-shot examples unrelated to the task
  3. Removing the retrieved context to shorten the prompt
  4. Instructing the model to answer only from the provided context and to state when the context is insufficient Correct answer.

Explanation

Explicit grounding instructions direct the model to treat retrieved context as the source of truth and to admit insufficient evidence instead of guessing. Unrelated examples, higher temperature, and removing context all worsen adherence to retrieved evidence.

Question 23Context Engineering Foundations and the Databricks Platform

A team compares two retrieval configurations: one uses pure vector similarity, the other combines keyword and vector search. For queries containing exact product codes, why might the hybrid approach be preferable?

  1. Keyword search alone always outperforms vector search on all query types
  2. Keyword matching captures exact identifiers that dense vectors may blur, while vectors capture semantic intent Correct answer.
  3. Hybrid search removes the need for an embedding model entirely
  4. Vector search cannot store more than one embedding per document

Explanation

Dense vectors represent meaning but can lose exact lexical identifiers such as product codes, which keyword matching preserves. Combining both signals covers exact-match and semantic cases. Hybrid search still uses embeddings, and neither method dominates universally.

Question 24Context Engineering Foundations and the Databricks Platform

A developer must choose between fine-tuning and RAG for teaching a model a proprietary response style that depends on many internal examples. Which reasoning is soundest?

  1. Neither approach can influence response style
  2. RAG permanently changes model weights to encode the style
  3. Fine-tuning is required whenever any private document is involved
  4. Fine-tuning can adapt style and format from many examples, while RAG supplies factual context at query time Correct answer.

Explanation

Fine-tuning adjusts model behavior using many examples, making it suited to consistent style and format, whereas RAG injects factual context per query without changing weights. RAG does not alter weights, and private documents can be handled by retrieval alone. Both approaches can influence style.

Question 25Context Engineering Foundations and the Databricks Platform

A team observes that their RAG application performs well on benchmark questions but poorly on real user queries. Which evaluation practice best diagnoses the gap?

  1. Relying only on the benchmark scores already collected
  2. Building an evaluation set from representative real user queries and scoring retrieval and answer quality on it Correct answer.
  3. Increasing the number of chunks returned for every query
  4. Lowering the endpoint's provisioned concurrency

Explanation

Evaluation sets should reflect real query distribution, including phrasing and edge cases, to expose weaknesses hidden by curated benchmarks. Returning more chunks may add noise, and reusing benchmark scores cannot reveal the mismatch. Concurrency is unrelated to quality.

Question 26Context Engineering Foundations and the Databricks Platform

A developer is building a RAG application on Databricks and needs a place to store the vector embeddings that will be queried at serving time. Which Databricks component is designed to store and index embeddings for low-latency similarity search?

  1. Delta Live Tables
  2. Vector Search Correct answer.
  3. Unity Catalog volumes
  4. MLflow Model Registry

Explanation

Vector Search is the Databricks service that stores embeddings in a vector index and supports approximate nearest-neighbor queries for RAG retrieval. Delta Live Tables is a pipeline framework, the Model Registry tracks model versions, and volumes store files, so none of them provide similarity search over embeddings.

Question 27Context Engineering Foundations and the Databricks Platform

When preparing a corpus for a RAG application on Databricks, a data engineer splits source documents into chunks before generating embeddings. What is the primary reason for chunking rather than embedding entire documents?

  1. Chunking allows retrieval of focused passages that fit within the model's context window and improves answer precision. Correct answer.
  2. Chunking ensures that every document is represented by exactly one vector.
  3. Chunking eliminates the need for a separate embedding model.
  4. Chunking reduces the total storage cost of the embedding index.

Explanation

Embedding whole documents produces vectors that blur many topics together and often exceed the context window when retrieved. Chunking yields smaller, topically focused units that can be retrieved precisely and inserted into the prompt within token limits.

Question 28Context Engineering Foundations and the Databricks Platform

A team is designing a RAG pipeline on Databricks where source documents are updated continuously in a Delta table. They want the Vector Search index to reflect new and changed content without manual intervention. Which approach best meets this requirement?

  1. Create a Vector Search index with Delta Sync enabled so it automatically tracks changes from the source Delta table. Correct answer.
  2. Store embeddings in a Unity Catalog volume and query them with SQL.
  3. Schedule a daily job that drops and recreates the index from scratch.
  4. Use a Model Serving endpoint that recomputes embeddings on every request.

Explanation

Delta Sync indexes in Databricks Vector Search automatically synchronize with the source Delta table, including inserts, updates, and deletes, without requiring manual rebuilds. Recreating the index daily introduces staleness and overhead, while the other options do not provide vector indexing at all.

Question 29Context Engineering Foundations and the Databricks Platform

A developer is configuring a Databricks Model Serving endpoint for a RAG application and wants to control how many tokens the foundation model can generate in a single response. Which parameter in the serving request should be adjusted?

  1. max_tokens Correct answer.
  2. frequency_penalty
  3. top_p
  4. temperature

Explanation

The max_tokens parameter caps the number of tokens generated in the completion, directly limiting response length. Temperature and top_p control randomness, and frequency_penalty discourages repetition, but none of them set a hard limit on output length.

Question 30Context Engineering Foundations and the Databricks Platform

A RAG application on Databricks returns answers that are fluent but sometimes contradict the retrieved documents. An engineer proposes changing the prompt to instruct the model to answer only from the provided context and to state when the context is insufficient. Why is this prompt change an appropriate mitigation?

  1. It guarantees that the retrieval step will return only relevant chunks.
  2. It explicitly constrains the model's generation to be grounded in the supplied passages, lowering the chance of unsupported claims. Correct answer.
  3. It fine-tunes the model weights to prefer retrieved content over prior knowledge.
  4. It reduces the model's context window usage so more documents can be retrieved.

Explanation

Grounding instructions in the prompt tell the model to rely on the retrieved passages and to abstain when they are insufficient, which reduces hallucinated or contradictory statements. Prompt changes do not alter retrieval quality, context-window size, or model weights.

Question 31Data Ingestion and Preparation for Context

When preparing a Delta table for retrieval-augmented generation, which Unity Catalog object is the appropriate place to store the chunked text and its embedding column so that access can be governed per-user?

  1. A local Parquet file in the driver's /tmp directory
  2. A notebook widget holding the embeddings
  3. A cluster-scoped Spark broadcast variable
  4. A managed Delta table in a catalog with grants on the table Correct answer.

Explanation

Governed context data should live in a Unity Catalog table so that GRANT statements control who can read chunks and embeddings. Driver-local files, widgets, and broadcast variables are not governed catalog objects and do not persist as shared data assets.

Question 32Data Ingestion and Preparation for Context

You need to ingest CSV files that arrive daily into a landing volume and keep only rows that changed since the previous load. Which combination best supports this incremental preparation pattern?

  1. Converting the CSVs to images and storing them in a volume
  2. A full overwrite of the bronze table on every run with no change tracking
  3. Auto Loader with schema inference writing to a bronze table, followed by a MERGE into silver Correct answer.
  4. Reading the CSV with pandas on the driver and printing the dataframe

Explanation

Auto Loader incrementally discovers new files in a volume, and MERGE applies inserts and updates to a downstream table based on a key. Together they avoid reprocessing the entire history on each daily run.

Question 33Data Ingestion and Preparation for Context

Which Delta Lake operation should you use to remove records that are no longer permitted in a context corpus while physically deleting their data from storage to satisfy a data-subject removal request?

  1. DELETE followed by VACUUM after the retention interval Correct answer.
  2. Time travel to an older table version
  3. VACUUM with a retention of zero hours
  4. OPTIMIZE with Z-ORDER on the id column

Explanation

DELETE marks rows as removed logically, and VACUUM later removes unreferenced data files once the retention period has elapsed. Running VACUUM before the retention interval is blocked by default to protect readers and time travel.

Question 34Data Ingestion and Preparation for Context

A source system delivers JSON where some records contain a nested struct for author metadata and others omit it entirely. Which read option helps you avoid failing the load when the struct is missing in some files?

  1. Set recursiveFileLookup to false
  2. Set the mode to DROPMALFORMED to discard every record with a struct
  3. Set mergeSchema to true so evolving structures are reconciled Correct answer.
  4. Set the header option to true

Explanation

mergeSchema allows Spark to combine differing schemas across files, so a struct present in some files and absent in others is handled without a read failure. Dropping malformed records would silently discard valid data, and header applies to delimited text, not JSON structures.

Question 35Data Ingestion and Preparation for Context

Which statement correctly describes the role of a bronze table in a medallion-style pipeline that feeds a retrieval corpus?

  1. It stores aggregated evaluation metrics for the model
  2. It stores only the final embeddings served to the application
  3. It stores raw ingested data with minimal transformation for traceability Correct answer.
  4. It stores the Unity Catalog grant history

Explanation

The bronze layer preserves raw ingested records so downstream silver and gold layers can be rebuilt and audited. Embeddings and serving artifacts belong to later layers, not to bronze.

Question 36Data Ingestion and Preparation for Context

You are chunking long PDFs into passages before embedding. Which preparation step most directly improves retrieval quality for a question-answering use case?

  1. Splitting on semantic boundaries such as headings and paragraphs with modest overlap Correct answer.
  2. Removing all punctuation to reduce token count
  3. Concatenating all pages of a document into one chunk
  4. Splitting on fixed character counts with no overlap

Explanation

Chunking along semantic boundaries keeps related ideas together, and a modest overlap preserves context that spans a boundary. Fixed splits can cut sentences, while a single giant chunk dilutes relevance during similarity search.

Question 37Data Ingestion and Preparation for Context

Which Delta Lake feature lets you enforce that a column used as a chunk identifier is never null when data is written?

  1. A table comment describing the column
  2. A CHECK constraint defined on the table Correct answer.
  3. A change data feed setting
  4. A Z-ORDER index on the column

Explanation

Delta Lake supports CHECK constraints that reject writes violating the predicate, such as a non-null chunk identifier. Z-ORDER affects file layout, comments are documentation, and change data feed records row changes rather than validating them.

Question 38Data Ingestion and Preparation for Context

Which Spark read pattern is appropriate for ingesting only files added since the last run from a cloud storage location into a context pipeline?

  1. Auto Loader using cloudFiles with a checkpoint location Correct answer.
  2. spark.read.json on the entire directory each run
  3. dbutils.fs.cp of the directory to a new path
  4. spark.read.format("binaryFile") on a single file

Explanation

Auto Loader tracks which files have already been processed using a checkpoint and schema log, so each run ingests only new arrivals. Reading the whole directory reprocesses everything, and copying files does not perform ingestion.

Question 39Data Ingestion and Preparation for Context

A pipeline reads a CSV where a numeric column occasionally contains the literal string "unknown". Which approach prepares the data most robustly for embedding metadata filters?

  1. Read the column as integer with PERMISSIVE mode and keep nulls silently
  2. Read the column as integer and let the job fail on bad records
  3. Drop the column entirely from the dataset
  4. Read the column as string, then cast with try_cast and quarantine rows that fail Correct answer.

Explanation

Reading as string and casting with try_cast yields nulls instead of exceptions, letting you route invalid rows to a quarantine table for inspection. Failing the job or silently nulling values hides data-quality problems.

Question 40Data Ingestion and Preparation for Context

You want to deduplicate a context table so that each source document id keeps only its most recent version by ingestion timestamp. Which approach accomplishes this?

  1. Use repartition by document id without any ordering
  2. Use dropDuplicates on all columns
  3. Use a window partitioned by document id ordered by timestamp descending and keep row_number 1 Correct answer.
  4. Use distinct on the embedding column

Explanation

A window function ordered by timestamp descending assigns rank 1 to the latest row per document id, which you then filter to. Dropping duplicates across all columns only removes exact matches, and distinct on embeddings is unrelated to versioning.

Question 41Data Ingestion and Preparation for Context

When writing prepared chunks to a Delta table that will later be embedded, why is it useful to enable change data feed on that table?

  1. It enforces schema evolution on the table
  2. It compresses the table to reduce storage cost
  3. It allows downstream jobs to read only row-level inserts and updates since a version Correct answer.
  4. It automatically generates embeddings for new rows

Explanation

Change data feed exposes row-level changes with commit versions so downstream embedding jobs can process only new or modified chunks. It does not compute embeddings, enforce schema, or compress data.

Question 42Data Ingestion and Preparation for Context

A colleague proposes storing raw HTML pages and their extracted plain text in the same column of the context table. What is the main preparation problem with this design?

  1. It forces the pipeline to use Python instead of SQL
  2. It mixes markup noise with clean text, degrading embedding quality and filtering Correct answer.
  3. It prevents the table from being stored in Delta format
  4. It disables time travel on the table

Explanation

Embedding raw markup alongside extracted text introduces boilerplate tokens that dilute semantic similarity and complicate metadata filters. Storage format, language choice, and time travel are unaffected by column content.

Question 43Data Ingestion and Preparation for Context

Which setting ensures that a streaming ingestion job into a context bronze table can add new columns when the upstream source adds fields?

  1. Set the checkpoint to a new location on every run
  2. Set the maxFilesPerTrigger to one
  3. Set the trigger to availableNow
  4. Set mergeSchema to true on the write and enable automatic schema evolution Correct answer.

Explanation

Enabling schema evolution on the write lets new upstream columns be added to the target table without rewriting the pipeline. Checkpoint location, trigger mode, and file batch size affect execution behavior, not schema handling.

Question 44Data Ingestion and Preparation for Context

You must join a chunk table to a document metadata table where document ids in the metadata table are occasionally duplicated with different titles. What is the safest preparation step before the join?

  1. Cast both id columns to double before joining
  2. Use a cross join to keep all metadata combinations
  3. Deduplicate the metadata table to one row per document id using a deterministic rule Correct answer.
  4. Drop the metadata table and rely on chunk text only

Explanation

Duplicated keys on the metadata side fan out chunk rows and inflate the corpus, so reducing metadata to one row per id with a deterministic rule keeps the join one-to-many as intended. A cross join multiplies rows further, and casting ids to double risks precision loss.

Question 45Data Ingestion and Preparation for Context

Which Delta operation is best suited to reorganize a large chunk table by a frequently filtered metadata column to improve data skipping?

  1. RESTORE to an earlier version
  2. DESCRIBE HISTORY on the table
  3. VACUUM with a 168-hour retention
  4. OPTIMIZE with Z-ORDER on the metadata column Correct answer.

Explanation

OPTIMIZE with Z-ORDER co-locates related values in the same files, letting Delta skip files that cannot match a metadata filter. VACUUM removes old files, RESTORE reverts versions, and DESCRIBE HISTORY only reports past commits.

Question 46Data Ingestion and Preparation for Context

A team wants to parse semi-structured event logs where each line contains a JSON object but the file also has a header comment line. Which ingestion approach handles this most reliably?

  1. Read as text, filter out comment lines, then parse the JSON with from_json Correct answer.
  2. Read as Parquet with schema inference
  3. Read as binaryFile and embed the raw bytes
  4. Read as CSV with a comma delimiter

Explanation

Reading as text lets you discard non-JSON lines before applying from_json to the remaining records. CSV parsing would mis-split JSON content, and binary or Parquet reads do not match the line-oriented JSON format.

Question 47Data Ingestion and Preparation for Context

Which practice best protects a context pipeline from ingesting the same source file twice when the upstream system sometimes rewrites files under the same name?

  1. Use a new target table for each run
  2. Rely on file name alone as the deduplication key
  3. Track file modification time and size along with the path in the ingestion state Correct answer.
  4. Disable checkpoints to allow reprocessing

Explanation

Because a rewritten file can keep the same name, deduplication keys should include modification time and size so genuinely changed files are detected while unchanged ones are skipped. Name-only keys miss rewrites, and disabling checkpoints or creating new tables defeats incremental tracking.

Question 48Data Ingestion and Preparation for Context

Which statement best explains why you would store embeddings in a Delta table as an ARRAY of floats rather than as a single delimited string?

  1. Arrays preserve numeric type and length for downstream vector functions and validation Correct answer.
  2. Strings are always rejected by Delta Lake on write
  3. Strings cannot be stored in Unity Catalog tables
  4. Arrays automatically index vectors for similarity search

Explanation

An array of floats keeps each embedding element typed and lets you validate dimension consistency and feed vector functions directly. Delta accepts strings, and neither representation creates a vector index automatically.

Question 49Data Ingestion and Preparation for Context

A pipeline needs to redact personally identifiable information from documents before embedding. Where should redaction occur in the medallion flow to minimize risk?

  1. In the silver transformation before chunks are written for embedding Correct answer.
  2. Only in the gold serving layer after embeddings exist
  3. After embeddings are computed by editing the vector values
  4. In the bronze raw table by overwriting source files

Explanation

Redacting before chunks are embedded ensures sensitive text never reaches the embedding service or the vector store. Redacting later leaves PII in intermediate artifacts and embeddings, and editing vectors does not remove the underlying text.

Question 50Data Ingestion and Preparation for Context

You observe that a MERGE into the chunk table is slow because the source has many small files. Which preparation step most directly addresses the small-file problem before the merge?

  1. Run OPTIMIZE on the source table to compact files Correct answer.
  2. Enable time travel on the source table
  3. Add a comment to the source table
  4. Increase the VACUUM retention to 30 days

Explanation

Compacting many small files into fewer larger ones reduces per-file overhead during the MERGE scan. Retention settings, comments, and time travel do not change file counts or read efficiency.