Design Applications
Free Databricks Certified GenAI Engineer Associate practice — 6 questions on Design Applications, with explanations. No sign-up. Full 12-question mixed test →
A financial services company is building a RAG application that must retrieve relevant documents from a 500GB corpus of regulatory filings, then use an LLM to generate compliance summaries. The application requires sub-second retrieval latency and must handle 10,000 concurrent users during market hours. Token costs must be minimized while maintaining accuracy. Which architectural approach BEST meets these requirements?
This solution addresses all constraints: Vector Search with pre-computed embeddings provides sub-second retrieval at scale; Redis caching reduces redundant LLM calls for common queries (critical for cost with 10K concurrent users); smaller open-source model significantly reduces token costs while few-shot prompting maintains accuracy for specialized compliance tasks. Pre-computed embeddings eliminate query-time embedding latency.
An e-commerce platform needs to generate personalized product descriptions using customer browsing history and product attributes. The system processes 2 million products across 50 categories. Descriptions must be factually accurate and include specific product specifications that cannot be hallucinated. The LLM frequently generates descriptions mentioning features the product doesn't have. Which design pattern MOST effectively prevents hallucination while maintaining personalization?
This approach enforces ground-truth constraints at generation time. Retrieving specifications from Unity Catalog ensures factual accuracy; constrained prompting with JSON schema validation (using tools like Pydantic or function calling) forces the LLM to select only from provided values rather than generate free-form specifications. This prevents hallucination structurally, not just probabilistically. The LLM can still personalize tone and style while being constrained to factual specifications.
A healthcare application uses a conversational agent built on Databricks to answer patient questions about lab results. The system must handle multi-turn conversations where context from previous messages affects interpretation (e.g., 'What about my cholesterol?' after discussing blood work). Conversations average 8 turns and the application serves 50,000 daily active users. Token costs are a primary concern. Which state management strategy BEST balances cost and context accuracy?
Semantic compression addresses the core challenge: at 8 turns average with 50K users, raw history creates massive token consumption. This approach extracts and stores only salient facts (e.g., 'patient asked about cholesterol on turn 3, value was 215'), then retrieves relevant compressed context. This dramatically reduces tokens while maintaining semantic accuracy. Vector similarity ensures the right context is retrieved when needed (e.g., cholesterol question retrieves prior cholesterol discussion).
A legal tech company is deploying a RAG system that must cite specific page numbers and paragraphs from source documents in generated responses. The system uses Databricks Vector Search to retrieve relevant chunks from 100,000 legal documents. During testing, the LLM frequently provides correct information but cites incorrect page numbers or attributes information to the wrong document. Chunk size is 512 tokens with 50-token overlap. Which modification MOST effectively improves citation accuracy?
Citation errors occur because retrieved chunks lack explicit source metadata that the LLM can reference. Including document_id, page_number, paragraph_number as metadata with each chunk (stored in Vector Search metadata fields) ensures this information flows to the LLM. Structured output formats (JSON schema, function calling) force the LLM to explicitly populate citation fields from the metadata rather than hallucinating page numbers. This is an indexing and prompt engineering solution to a retrieval-generation alignment problem.
A media company is building an application that generates video thumbnails and titles using multimodal LLMs. Videos are stored in Unity Catalog Volumes (average 500MB each, 10,000 videos). The application must process each video once for initial indexing, then handle 1,000 user queries per hour requesting thumbnail/title variations. Token costs for processing full videos are prohibitive. Which architectural pattern BEST optimizes cost while maintaining quality?
This approach separates expensive processing (embedding all frames, done once during indexing) from cheap query-time operations (retrieve top 5 frames, send only those to LLM). At 1,000 queries/hour, sending 5 frames is far cheaper than processing 500MB videos. Vector search over frame embeddings intelligently selects the most visually distinct/representative frames, so the LLM sees video diversity without processing all content. This is a classic pre-computation + query-time refinement pattern.
A customer support application uses a compound AI system with three components: (1) intent classifier (routes to specialized models), (2) retrieval system (fetches knowledge base articles), and (3) response generator (creates answers). During A/B testing, the system shows 15% lower accuracy than a single GPT-4 call, despite individual components each testing at >90% accuracy. Investigation shows most errors occur when the classifier misidentifies intent, causing retrieval of irrelevant articles. Which design change MOST effectively improves end-to-end accuracy?
This addresses the core failure mode: classification errors cascade through the pipeline. Confidence-based fallback catches uncertain classifications before they cause retrieval errors. Re-ranking retrieved articles against the original query (not classified intent) provides error correction - even if intent is misclassified, relevant articles can still surface. This adds redundancy and error tolerance without abandoning the efficiency benefits of specialized routing. The confidence threshold provides a tunable quality/cost tradeoff.
Ready for the real thing?
The full course: two full-length practice tests, video lessons for every exam domain, hands-on labs and detailed explanations.
$199.99 $66.99 with code FREETEST33 — valid through August 23.
Get my $66.99 deal →