Design Applications practice questions

Classification can be single-label (one category per input) or multi-label (any number of categories per input).

The deciding facts are that one review can describe several problems and that the team wants every applicable category. Single-label classification would force a choice and drop issues. Sentiment addresses a different question, and a summary sentence is unstructured text. Multi-label classification returns each matching category from the fixed set; Databricks ai_classify supports this with its multilabel option.

1. Identify the output: categories from a fixed list of six.
2. Check whether one input can belong to several categories: here it can.
3. Choose multi-label classification so no applicable category is dropped.

Exam tip: If an input can fall into several known categories and you need all of them, specify multi-label classification.

Extraction turns unstructured text into named fields; it is the task to choose when the output is a set of specific values.

Finance needs three specific values per email as separate columns. Field extraction returns exactly those values, which on Databricks maps to ai_extract with a schema or a prompted model with a structured output. Classification only decides a category, a summary is unstructured prose, and sentiment is unrelated to the values.

1. List the outputs: three named values per email.
2. Named values copied from the text means extraction.
3. Reject tasks that return a category, prose or a score.

Exam tip: Specific values out of free text into columns: choose extraction.

Summarization condenses meaning and may paraphrase; extraction returns exact values as fields. The task must match what the consumer does with the output.

The auditors reconcile values against another system, so they need each value present and exactly as written. That is an extraction requirement, best expressed as one output field per value. A longer summary or a deterministic summary keeps the paraphrasing problem, and classifying first does not change the free-text output.

1. Note how the output is consumed: values are reconciled against a system.
2. Reconciliation needs exact values, so paraphrase is a defect.
3. Change the task from summarization to extraction with named fields.

Exam tip: When values must be exact and reconcilable, extract fields; do not summarize.

Summarization condenses content while preserving key points; translation changes language; extraction returns fields.

The requirement is a shorter version of the same document, in the same language, keeping results and risks. That is summarization. Translation is ruled out by the same-language requirement, a full table of figures does not condense, and a single risk label loses almost all of the content.

1. Identify the transformation: long text to short text.
2. Check the language requirement: it stays the same, so no translation.
3. Choose summarization.

Exam tip: Shorter text, same language, main points kept: summarization.

Translation changes language without changing meaning; requirements often need it on both the input and the output side of a pipeline.

The agents' language (English) differs from the customers' languages, and the agents write the replies. So incoming messages must be translated to English and outgoing replies translated back. Translating only the inbound side leaves customers with English replies, summarization loses detail and does not change language, and generating replies removes the agent from the loop, which the scenario does not ask for.

1. Map who reads and who writes, and in which language.
2. Inputs must become English; outputs must return to the customer's language.
3. Agents write replies, so no reply generation is needed.

Exam tip: Map each hop of the conversation to a language; translation may be needed both into and out of the pipeline.

Embedding models map text to vectors whose distance reflects meaning, which suits similarity and de-duplication tasks at scale.

Near-duplicate detection is a similarity problem. Embeddings capture meaning across different wordings, cost one model call per item and are compared with cheap vector math or a vector index. Asking a chat model about every pair is accurate but scales with the square of the catalog. Matching summaries is unreliable and still costly, and a shared category does not mean a duplicate.

1. Recognise the task: semantic similarity across wording differences.
2. Consider scale: per-pair model calls grow quadratically.
3. Choose embeddings plus vector similarity.

Exam tip: Similarity, de-duplication and semantic matching at scale: embeddings, not chat prompts.