Practice Test #1

Weighted accuracy is not a standard evaluation metric exposed by Azure Custom Vision for classifier iterations. While weighted accuracy can be useful in some machine learning frameworks for imbalanced datasets, it is not one of the built-in metrics typically reviewed in the Custom Vision portal. The service focuses on precision, recall, and combined performance measures instead. Therefore this option does not match the metrics available for review in this scenario.

Precision is indeed a metric shown by Custom Vision, but the question asks for two metrics and the expected pair from the provided options is recall and F-score. Precision measures the proportion of predicted positives that are actually correct, which is valuable for understanding false positives. However, when forced to choose two among the listed options, exam wording commonly expects recall plus the combined F-score metric rather than precision. So this option is not part of the best answer set here.

Area under the curve (AUC) is associated with ROC analysis and is common in many binary classification tools, but it is not a standard metric surfaced in Azure Custom Vision classifier evaluation. Custom Vision does not typically present ROC curves or AUC values in its built-in review experience. The exam usually contrasts AUC as a distractor against the actual Custom Vision metrics. Therefore AUC is not correct for this question.

QnAMakerOptions.StrictFilters is used to apply metadata-based filtering so QnA Maker only considers QnA pairs that match specified key/value metadata. While this can narrow the search space (for example, by product, locale, or channel), it does not enforce a minimum confidence level for the returned answer. If the filtered set still yields a weak match, QnA Maker may return it unless separately gated by a threshold. Therefore it does not directly meet requirements focused on confidence-based acceptance of answers.

QnADialogResponseOptions.CardNoMatchText controls the text shown to the user when the QnA dialog determines there is no suitable match. This is a presentation/UX configuration for the bot’s response, not a property that influences QnA Maker’s scoring or match selection. It does not change how QnA Maker evaluates confidence or which answers are considered acceptable. You would typically use this alongside a score threshold, but it is not the integration property that enforces answer quality.

QnAMakerOptions.RankerType specifies which ranking approach QnA Maker should use (for example, different rankers for answer selection). Although ranker choice can affect ordering and scoring behavior, it does not provide a direct mechanism to reject low-confidence answers. The requirement to only respond when QnA Maker is sufficiently confident is addressed by setting a minimum score threshold, not by selecting a ranker type. RankerType is more about ranking strategy than acceptance criteria.

File2 cannot be used because MP4 is a video format, and Form Recognizer/Document Intelligence does not train on video files. The service is designed for documents and still images rather than time-based media. Even if a video contains readable text, extracting information from it would require a different solution pipeline.

File4 cannot be used because although PDF is a supported format, the file size is 100 MB. This exceeds the 50 MB maximum size limit for training files in the service. Supported format alone is not enough; the file must also comply with service constraints such as size.

File5 cannot be used because GIF is not a supported file type for custom model training in Form Recognizer/Document Intelligence. The service supports document-oriented image formats such as JPG, PNG, TIFF, and BMP instead. GIF is a common distractor because it is an image format, but it is not one of the accepted training inputs.

Microsoft Translator is used to translate text between languages, which is helpful for multilingual scenarios but does not provide conversational routing or intent dispatch. It cannot determine whether a message should be answered by QnA Maker or another bot capability. Therefore, it does not satisfy the need for an additional chatbot intelligence component in this architecture. It is a supporting service, not the core missing orchestration layer.

Language Understanding (LUIS) is used to classify intents and extract entities from user input, but that is not the same as orchestrating among conversational resources. The question asks for an additional component to complement a QnA Maker resource and a Bot Framework SDK bot, and the more specific missing capability is dispatch/routing. Orchestrator is built for that purpose, whereas LUIS would only add one recognizer model. LUIS may be part of a broader solution, but it is not the best answer here.

chatdown is a conversation authoring and testing format used in bot development workflows. It does not provide runtime natural language processing, dispatching, or knowledge-base integration. Adding chatdown would not improve the bot’s ability to interpret or route user requests. As a result, it cannot meet the technical chatbot requirements implied by the scenario.

Creating a custom language model is not the best answer in the context of configuring Azure Video Analyzer for Media/Video Indexer. While Azure AI Speech supports customization, the question asks what to do with Video Indexer to obtain transcripts that meet requirements, and the listed Video Indexer-relevant capability is multi-language detection. The explanation overstates custom language model use as if it were the primary Video Indexer setting here. Therefore, A is not the most appropriate choice from the given options.

Configuring audio indexing for videos only does not address the quality or correctness of the transcript itself. It simply limits or defines what media is processed and does not help with language changes, terminology recognition, or transcript usability for a knowledgebase. If the requirement is to ensure the transcript supports downstream knowledge extraction, this option is too indirect. It is an operational scope choice rather than a transcript-accuracy feature.

A custom Person model is used to identify or distinguish people appearing in the video, typically for visual insights and metadata enrichment. That can help label presenters, but it does not improve speech-to-text output or transcript quality. Since the knowledgebase is being built from webinar transcripts, visual person recognition is not the key requirement. This option targets video analytics rather than transcript generation.