Task 1.2: Building AI solutions using Google Cloud AI APIs or foundational models.

• AThe Form Parser, which returns the key-value pairs found on each page

  Incorrect

  Form Parser returns generic key-value pairs and tables, not the city's 25 named fields as a defined schema.

• BThe pretrained Invoice Parser, run on the permit forms

  Incorrect

  A pretrained parser extracts its own document type's fields; permit forms are not invoices.

• CA Custom Extractor using its generative AI training method

  Correct

  The custom extractor is for document types with no pretrained processor, and its generative mode needs from zero to a few dozen documents.

• DA Custom Classifier trained on the 30 sample applications

  Incorrect

  A classifier assigns a document to a class; it does not extract fields.

Document AI's custom extractor builds entity extraction for new document types that have no pretrained processor. Its generative AI (foundation model, optionally fine-tuned) method needs about 0 to 50 documents for production quality, while custom models and template-based models need more consistent layouts or more training documents.

• ACollect the day's questions into one batch inference job and return the answers to the lawyers the next morning

  Incorrect

  Batch inference is cheaper but asynchronous, which breaks the few-seconds answer requirement.

• BBuy Provisioned Throughput for the model to keep the requests from ever being throttled

  Incorrect

  Provisioned Throughput guarantees capacity; it does not reduce the tokens billed for the repeated bundle.

• CCache the bundle with explicit context caching and send each question with a reference to it

  Correct

  Caching the repeated content cuts the cost and latency of requests that reuse it, and the cache can be refreshed when the bundle changes.

• DFine-tune the model on the contract bundle so it no longer has to be sent with each question

  Incorrect

  Tuning teaches style and behavior, not reliable recall of a contract that changes every week.

Context caching reduces the cost and latency of requests that contain repeated content. Implicit caching is automatic; explicit caching lets you declare the content to cache and refer to it from later prompts, which fits a large bundle reused thousands of times a day.

• ADOCUMENT_TEXT_DETECTION

  Correct

  This feature is optimized for dense text and documents, including handwriting, and returns page, block, paragraph and word structure.

• BTEXT_DETECTION on each photo, which reads text in images

  Incorrect

  TEXT_DETECTION reads text such as signs in images; the response is not optimized for dense documents.

• CLABEL_DETECTION, which returns the general labels found in an image

  Incorrect

  Labels describe what an image shows, not the words written on it.

• DOBJECT_LOCALIZATION, which finds and boxes the objects in each photo

  Incorrect

  Object localization returns objects and bounding boxes, not text.

The Vision API offers TEXT_DETECTION for text in general images and DOCUMENT_TEXT_DETECTION, whose response is optimized for dense text and documents (including handwriting) and includes page, block, paragraph, word and break information.

• ASend requests to the global endpoint

  Incorrect

  The global endpoint has the most headroom but may process requests in any location, which breaks the EU requirement.

• BSend requests to the eu multi-region endpoint

  Correct

  The jurisdictional eu endpoint keeps ML processing within EU member states while pooling capacity across data centers instead of one region.

• CRetry throttled requests with truncated exponential backoff

  Correct

  On shared PayGo capacity a 429 signals temporary contention, and exponential backoff is the recommended way to absorb it.

• DMove the workload to Flex PayGo

  Incorrect

  Flex PayGo is for latency-tolerant work and accepts more throttling, the opposite of what a live assistant needs.

• ERaise the temperature parameter on each request

  Incorrect

  Temperature changes how random the output is; it has no effect on throughput or throttling.

Global endpoints give the most availability but may process requests anywhere, so they provide no data residency guarantee. Jurisdictional multi-region endpoints (us and eu) keep ML processing inside the boundary and pool capacity across several data centers. On Standard PayGo a 429 reflects temporary contention for shared resources, so the client should retry with truncated exponential backoff.

• AFlex PayGo, which is priced lower per token

  Incorrect

  Flex PayGo is for latency-tolerant work and accepts more throttling, the opposite of a live assistant's needs.

• BPriority PayGo, which costs more per token than Standard PayGo

  Correct

  Priority PayGo gives more consistent performance than Standard PayGo without the upfront commitment of Provisioned Throughput.

• CA one-month Provisioned Throughput commitment, sized for the peak, instead of Standard PayGo

  Incorrect

  Sizing a commitment for unpredictable peaks leaves capacity unused most of the time, which the team wants to avoid.

• DBatch inference for all requests

  Incorrect

  Batch inference is asynchronous and cannot answer users in real time.

Priority PayGo is a consumption option that provides more consistent performance than Standard PayGo without Provisioned Throughput's upfront commitment, at a higher per-token rate. It suits business-critical workloads with fluctuating or unpredictable traffic.

• AA glossary that maps each product name to its translation

  Incorrect

  A glossary fixes individual terms; it does not carry over tone and phrasing from whole sentences.

• BA Gemini model fine-tuned on the sentence pairs and served from an endpoint

  Incorrect

  Tuning and serving a model is the training and management work the company wants to avoid.

• CDocument translation for the articles, which keeps their layout

  Incorrect

  Document translation preserves formatting; it does not adapt the wording to the company's style.

• DAdaptive translation with the sentence pairs as an example dataset

  Correct

  Adaptive translation uses example translations to tailor responses, selecting the most similar reference sentences for each request.

With adaptive translation, Cloud Translation uses example translations to customize its output. With a dataset of sentence pairs (TSV or TMX), it automatically picks the five reference sentences most similar to each source sentence; no model has to be trained.

• ARETRIEVAL_DOCUMENT for the articles and RETRIEVAL_QUERY for the queries

  Correct

  Retrieval is an asymmetric use case: documents and queries are embedded with their own task types.

• BSEMANTIC_SIMILARITY for both the articles and the queries

  Incorrect

  SEMANTIC_SIMILARITY is for text similarity and is not intended for retrieval use cases.

• CCLUSTERING for the articles and CLASSIFICATION for the queries

  Incorrect

  These task types are for grouping and labeling texts, not for matching queries to documents.

• DRETRIEVAL_QUERY for the articles and RETRIEVAL_DOCUMENT for the users' queries

  Incorrect

  This reverses the roles; documents use RETRIEVAL_DOCUMENT and queries use RETRIEVAL_QUERY.

Embedding task types optimize embeddings for a use case. Search retrieval is asymmetric: embed the documents with RETRIEVAL_DOCUMENT and the queries with RETRIEVAL_QUERY. SEMANTIC_SIMILARITY, CLASSIFICATION and CLUSTERING are symmetric, single-input task types.

• AAutomatic language detection for each call

  Incorrect

  Language detection picks the spoken language; it does not separate the speakers.

• BWord-level confidence scores returned with each word

  Incorrect

  Confidence scores show how sure the recognizer is about each word, not who said it.

• CSpeaker diarization in the recognition request

  Correct

  Diarization detects when speakers change and labels the voices by number in the transcript.

• DProfanity filtering on the transcripts

  Incorrect

  Profanity filtering masks words; it does not attribute them to speakers.

Cloud Speech-to-Text can recognize multiple speakers in the same audio. With speaker diarization enabled in the request, it detects speaker changes and labels each word with a speaker number.