Task 3.2: Design and implement multimodal understanding workflows

• AAs an input_image with detail set to low

  Correct

  Low detail uses a lower-resolution version of the image, using the fewest tokens with the fastest response.

• BAs an input_image with detail set to high

  Incorrect

  High detail captures finer features but uses more tokens and takes longer, which this job doesn't need.

• CAs an image URL plus the file name and EXIF data in the prompt

  Incorrect

  The model doesn't process file names or metadata, and extra prompt text only adds tokens.

• DAs an input_image with detail set to auto

  Incorrect

  Auto leaves the choice to the model, so it may spend more tokens than a one-line caption needs.

The detail level trades accuracy on fine features against tokens and speed. Use low for simple images and short captions, high when small text or fine detail matters, auto when you don't know in advance.

• AAn Azure AI Search index of shelf images queried for each product name

  Incorrect

  Search retrieves matching images or text; it doesn't count units or find gaps in a photo.

• BThe same chat model with detail set to high and a longer prompt

  Incorrect

  The docs warn that vision models give approximate counts and struggle with precise spatial localization, whatever the detail setting.

• CA Content Understanding image analyzer with count and location fields

  Correct

  Content Understanding extracts schema-defined fields from images and lists shelf analysis and inventory, detecting and counting products, among its uses.

• DAn image generation model that redraws each shelf with labels added

  Incorrect

  Generation models create new images; a redrawn shelf isn't a measurement of the real one.

General vision chat is good at describing scenes but approximate at counting and locating objects. For repeatable, structured extraction from images, define the fields in a Content Understanding analyzer.

• AA vision chat model sent one frame per minute with a request for topics

  Incorrect

  Sampled frames miss speech and cuts, and stitching the answers back into timed chapters is custom work.

• BA Content Understanding video analyzer with custom segmentation by topic

  Correct

  Video analyzers produce transcripts, key frames and descriptions, and segmentation splits a video into logical chunks based on categories you define, in RAG-ready output.

• CSpeech to text on the audio, then split the transcript every five minutes

  Incorrect

  Fixed time slices ignore where topics change and drop the visual content.

• DVideo translation of each recorded session

  Incorrect

  Video translation dubs speech into other languages; it doesn't create chapters.

Content Understanding treats video in two stages: content extraction (transcripts, shots, key frames) and field extraction with segmentation. Custom segments give chapter boundaries and descriptions ready for an index.

• ASplit each video into one file per language before analysis

  Incorrect

  The speakers switch languages within a clip, so there's no clean per-language file to cut; multilingual transcription applies a locale per phrase instead.

• BSet the locale to en-US for every one of the videos

  Incorrect

  A single locale misreads the Spanish segments.

• CLeave the locale unset or set to auto for multilingual output

  Correct

  Multilingual transcription is enabled when no language or locale is specified or the language is set to auto.

• DRun Azure Translator text translation on each video file first

  Incorrect

  Translator works on text, not video files.

Configure locales when content uses one known language, and use multilingual transcription for mixed-language media; unsupported locales fall back to the closest one, which can be wrong.

• Aprebuilt-imageSearch

  Correct

  It generates a one-paragraph description of each image for search and retrieval.

• Bprebuilt-layout

  Incorrect

  prebuilt-layout is a document analyzer, not an image describer.

• Cprebuilt-read

  Incorrect

  prebuilt-read targets text in documents, not descriptions of photos.

• Dprebuilt-invoice

  Incorrect

  That analyzer extracts invoice fields from documents.

For image retrieval, Content Understanding offers prebuilt-imageSearch, whose descriptions can be embedded and searched.

• ASend the HEIC files as PDF file inputs to the model instead

  Incorrect

  A renamed image isn't a PDF.

• BSplit each photo into smaller HEIC tiles before sending

  Incorrect

  Smaller files of an unsupported type still fail.

• CRaise the image detail level to high on each request

  Incorrect

  Detail doesn't change which file types are accepted.

• DConvert each photo to JPEG or PNG before sending

  Correct

  Supported image inputs are PNG, JPEG, WEBP and non-animated GIF.

Check image input requirements before sending: supported file types, the per-request image count and size limits, and readable files.

• AKey frame extraction in the video analyzer output

  Incorrect

  Key frames are images, not speaker labels.

• BSegment descriptions in the video analyzer output

  Incorrect

  Descriptions say what happens in a segment, not who is speaking.

• CA text blocklist applied to the transcripts

  Incorrect

  Blocklists filter terms; they don't identify speakers.

• DDiarization in the video analyzer's transcripts

  Correct

  Video content extraction includes diarization, attributing parts of the transcript to speakers.

Content Understanding's first stage for video extracts transcripts with diarization and key frames; the second stage adds fields and segments.

• AThe segment's category label from the content categories

  Incorrect

  A category label isn't an image.

• BThe key frames listed for each segment in the Markdown

  Correct

  The video analyzer's Markdown includes ordered key-frame thumbnails for each segment.

• CThe WEBVTT transcript of each matching segment

  Incorrect

  The transcript holds speech text, not images.

• DThe analyzer's confidence scores for each segment

  Incorrect

  Confidence scores are numbers, not thumbnails.

Each segment in the video output lists its own key frames, so an app can show them beside that segment's description.