Assembling and Deploying Applications: Databricks Generative AI Engineer Associate Practice Questions

Correct answer: C

WHY C is correct: Hybrid keyword-similarity search combines vector-based embedding search with traditional keyword-based BM25 search. The Databricks documentation explicitly calls out that this is 'particularly useful in RAG applications where source data has unique keywords such as SKUs or identifiers that are not well suited to pure similarity search.' Results from both methods are fused using Reciprocal Rank Fusion (RRF). WHY NOT A: Increasing embedding dimension improves the model's ability to represent semantic concepts in dense vector space, but SKUs are arbitrary alphanumeric sequences that don't carry distributional semantic signal — higher dimensions don't help with exact-match retrieval of opaque identifiers. WHY NOT B: Direct Vector Access Index changes how embeddings are updated (manual vs. auto-sync) but does not add keyword search capability. It does not improve exact-match retrieval of SKU queries. WHY NOT D: Metadata filtering pre-filters documents before similarity scoring — it can work for exact SKU lookups if the SKU is stored as a metadata column. However, the question implies mixed queries that include SKU keywords alongside semantic content. Hybrid search handles both simultaneously without requiring a pre-known exact SKU filter value.

Correct answer: B

WHY B is correct: This is the canonical LCEL minimal chain. The pipe operator (|) chains Runnable objects left-to-right: the PromptTemplate formats the input dict into a prompt message, ChatOpenAI() calls the LLM and returns an AIMessage, and StrOutputParser() extracts the text string. The chain is invoked with chain.invoke({'question': '...'}). WHY NOT A: The + operator is not a valid chain composition operator in LCEL. Runnable objects are composed with | (pipe), not +. WHY NOT C: LLMChain is the legacy pre-LCEL API. While it works, it is not LCEL and does not support .stream(), .batch(), or async operations natively. The question specifies LCEL and the pipe-operator approach. WHY NOT D: StrOutputParser() is not a wrapper constructor that accepts other objects as positional arguments. It is an instantiated Runnable that receives LLM output via | — it cannot be constructed with the prompt and LLM as parameters.

Correct answer: B

WHY B is correct: The Databricks Prompt Registry (integrated via MLflow's prompt engineering and AI Gateway) provides native versioning for prompts, tracks performance metrics per version, supports alias-based promotion (e.g., 'production' alias pointing to a specific version), and enables rollback by reassigning aliases. This is the purpose-built solution for the described prompt lifecycle management workflow. WHY NOT A: Delta time travel tracks table data history but is not optimized for prompt version management with associated evaluation metrics, aliases, or promotion workflows. It lacks the metadata structure needed to compare prompt versions by performance. WHY NOT C: Using separate MLflow experiments per prompt version creates a proliferation of experiments not linked to a shared prompt lineage. MLflow experiments are designed for model training runs, not for managing a versioned prompt entity with aliases. WHY NOT D: Git-based prompt management provides version history but requires manual processes for metric tracking, rollback, and promotion. It lacks the integrated evaluation metric association and alias-based deployment control provided by the Prompt Registry.

Correct answer: B

WHY B is correct: Databricks Apps allows teams to build and host interactive web applications (Gradio, Streamlit, React, etc.) directly within the Databricks workspace. A Gradio or Streamlit chat interface wrapping the Model Serving endpoint provides a browser-accessible UI with conversation history display, user-friendly input, and Databricks workspace authentication — exactly what non-engineer business analysts need. WHY NOT A: Providing raw REST API credentials and instructing analysts to use curl/Postman requires technical expertise that business analysts typically lack. It provides no persistent conversation history UI and is entirely inappropriate for non-engineer users. WHY NOT C: A Delta Live Tables batch pipeline is a data transformation workflow, not an interactive conversational interface. Analysts cannot engage in real-time conversation with an agent through a DLT pipeline. WHY NOT D: The 'Test endpoint' button in the MLflow / Unity Catalog model UI is a developer tool for testing input/output formats during development and debugging. It is not designed for ongoing production user interaction, does not maintain conversation history, and requires Unity Catalog access privileges that business analysts may not have.

Correct answer: B

WHY B is correct: 50M vectors at 768 dimensions is well within the standard endpoint's ~320M capacity. Delta Sync with Databricks-managed embeddings handles automatic embedding computation. Continuous sync mode keeps the index current as the source Delta table changes — ideal for a live product catalog. Standard endpoint query latency is lower than storage-optimized (storage-optimized adds ~250ms overhead). WHY NOT A: Storage-optimized endpoints have ~250ms additional query latency overhead, making sub-100ms queries difficult. They also don't support continuous sync mode per the documentation. They are designed for very large datasets (1B+) where standard endpoints are insufficient — unnecessary for 50M vectors. WHY NOT C: Direct Vector Access Indexes require manual REST API calls to update the index when products are added/removed. For a continuously updated catalog, this adds significant operational complexity and risks index staleness. Delta Sync auto-handles catalog changes. WHY NOT D: As noted, storage-optimized endpoints add ~250ms query latency and don't support continuous sync. Both properties disqualify this option for a real-time, sub-100ms catalog search system.

Correct answer: A

WHY A is correct: LCEL supports nesting: each parenthesized sub-pipeline (summarize_prompt | summarize_llm | StrOutputParser()) produces a Runnable that outputs a text string. That text string is then passed as input to the second stage's prompt formatter (classify_prompt), and the full chain is chained together with |. This correctly implements the sequential summarize-then-classify pattern. WHY NOT B: RunnableParallel runs its branches concurrently on the SAME input — it does not chain the summary output into the classification input. It would attempt to classify the original document rather than the generated summary. WHY NOT C: The + operator is not a valid LCEL operator for chaining Runnables. This mixing of | and + will raise a TypeError. WHY NOT D: RunnableSequence is a valid internal LCEL class, but it expects each step to output the correct type for the next step. Passing the raw LLM object (summarize_llm, classify_llm) at the same level as prompt objects without parsers between them will fail to correctly thread texts between the steps.

Correct answer: C

WHY C is correct: 1.2 billion vectors exceeds the standard endpoint's ~320M capacity — a single storage-optimized endpoint supports ~1B at 768 dimensions. Storage-optimized pricing is optimized for large vector counts. Since the archive is only updated annually, non-continuous (triggered) sync is sufficient and avoids unnecessary continuous sync overhead. The 500ms latency tolerance accommodates the storage-optimized endpoint's ~250ms additional latency. WHY NOT A: Standard endpoints support only ~320M vectors at 768 dimensions. 1.2B vectors cannot fit in a single standard endpoint. WHY NOT B: Splitting across 4 standard endpoints is technically feasible but creates significant operational complexity — the application must fan out queries across 4 endpoints and merge results. It is more expensive than a single storage-optimized endpoint at scale and architecturally fragile. WHY NOT D: Storage-optimized endpoints do not support continuous sync mode per Databricks documentation. Additionally, 'real-time updates' contradicts the annual refresh requirement — continuous sync would be wasteful and architecturally misaligned.

Correct answer: B

WHY B is correct: The Databricks Vector Search Python SDK's similarity_search() method accepts num_results for the result count limit and a filters parameter that takes a dictionary of column-value pairs for metadata filtering. {'status': 'approved'} restricts results to documents whose 'status' column equals 'approved'. WHY NOT A: This call is missing the status filter. It will return the top-5 most similar documents regardless of their status column value, potentially returning non-approved documents. WHY NOT C: where_clause is not a parameter in the Databricks Vector Search SDK's similarity_search() method. Vector Search filtering uses the filters dictionary parameter, not SQL WHERE clause strings. WHY NOT D: top_k and metadata_filter are not the correct parameter names in the Databricks Vector Search SDK. The correct parameters are num_results and filters respectively. Using incorrect parameter names will either cause errors or be silently ignored.