Prepare and test your skills
Prepare and test your skills
Information synthesis and knowledge discovery is the use of generative AI to turn large volumes of unstructured text, images, and video into summaries, structured facts, and insights that support business decisions. The pattern combines large-scale search over content with structured extraction from documents. Business teams use the results to create tailored materials, track competitors, improve customer service, and automate manual analysis work.
Embeddings (opens in a new tab) are vector representations of content that let a system find, group, and compare similar items across a large collection. Text embeddings accept input text and an optional output dimensionality, with a default of 768 dimensions. Image embeddings accept either raw image data or a Cloud Storage URI and can include optional contextual text and a dimension choice, where the contextual text vector is available only for single image creation. Video embeddings accept a Cloud Storage URI with optional start and end offset times in seconds, optional contextual text, and a dimension choice.
The options for each modality differ in inputs and output size:
| Modality | Input | Optional parameters | Output dimension |
|---|---|---|---|
| Text | Input text | Output dimensionality | Default 768 |
| Image | Raw data or Cloud Storage URI | Contextual text, dimension | Not specified |
| Video | Cloud Storage URI | Start/end offset seconds, contextual text, dimension | 128, 256, 512, or 1408, default 1408 |
The vector method for video returns embeddings only for the first segment of the video. To process many items, a loop generates an embedding for each text or image item, collects the embeddings into an internal table, and then retrieves the final table for search and analysis. Generated embeddings can be sent to a Cloud Storage bucket or a BigQuery dataset, so traffic flows from the embedding process to storage and then to discovery and summarization.
Controlled generation is the practice of asking a generative model to return output in a defined structure instead of free-form prose. The request sets the response MIME type to application/json and provides a response schema that names the required fields. In the weather example, the model gemini-2.5-flash converts an unstructured narrative into a forecast array whose items contain Day, Forecast, Temperature, Humidity, and WindSpeed, where some fields are required and others are nullable when the source does not provide a value. This structure turns business documents into records that can feed reports, dashboards, and downstream automation.
Safety settings control what the model is allowed to return during generation. The categories shown are dangerous content, harassment, hate speech, and sexually explicit content, and each category can be set to a threshold such as BLOCK_LOW_AND_ABOVE. These settings apply to the generation request and help keep knowledge-discovery systems within organizational boundaries while still extracting useful information.
The discovery workflow for business documents connects ingestion, search, extraction, and safety controls in a fixed order:
A left-to-right process flow showing five steps from ingesting text, images and video to generating embeddings, storing in Cloud Storage or BigQuery, extracting JSON fields, and applying safety settings, leading to four business uses.
Only after embeddings are stored can large-scale comparison happen, and only after controlled generation runs can unstructured narratives become analyzable fields. The workflow supports several business uses where synthesis creates value:
Exam tip: Embeddings turn content into vectors for similarity and search, while controlled generation turns unstructured text into structured JSON fields for analysis.
Multi-modal content creation and personalization is the use of generative AI (opens in a new tab) to produce text, images, video, and source code from simple prompts and tailor the result to a specific audience. One generative capability can work across modes to create original material, synthesize information, and automate routine creative work. Businesses use this approach to streamline creative processes and deliver highly personalized user experiences.
Generative AI supports four content modes that are often combined into one customer message:
Text generation is the ability of generative AI to create and adapt written content on demand for different readers. The business value comes from dynamic content that changes for each audience, such as custom one-pagers and dynamic slide decks. A marketing team provides a prompt with the topic, audience, and key points, and the model drafts the tailored version from existing information. This moves content work from slow manual rewriting to fast and repeatable creation.
Image and video generation is the ability of generative AI to produce visual content from text descriptions and other inputs. These modes streamline creative processes by giving teams a fast starting draft for campaigns, product stories, and training materials. A marketer can request variations in style, setting, or message without a full reshoot or redesign. The team then reviews, selects, and refines the best option, so human creativity focuses on direction and quality rather than starting from a blank page.
Source code generation is the ability of generative AI to draft code from natural language descriptions and examples. The generated code helps developers build and personalize digital experiences more quickly, such as websites, apps, and data-driven features. A developer describes the desired behavior, and the model suggests functions, fixes, or starter templates for that behavior. The developer then tests and adjusts the result, so automation speeds routine coding while people remain responsible for correctness and security.
Personalization is the use of generative AI to tailor content and experiences to individual needs, interests, and context. Personalized marketing connects text, images, and video into a coherent message where the wording, visuals, and offer fit each customer segment. The same foundation also supports proactive marketing, where AI-driven competitive intelligence helps teams detect market shifts and respond rapidly. Together, creation and intelligence move a business from reactive and one-size-fits-all outreach to sharper and more relevant engagement.
Exam tip: Custom one-pagers and dynamic slide decks are examples of dynamic personalized content, while rapid response to market shifts is an example of proactive use of AI-driven intelligence.
Operational automation and workflow optimization is the use of generative AI to reduce manual work in data analysis and in customer-facing and internal operations. A generative AI virtual agent (opens in a new tab) connects to business data sources, understands conversational requests, summarizes complex information, and triggers follow-on actions in other systems. One automated workflow can therefore understand the request, find an answer, and create records for what happens next.
Generative AI virtual agents are automated assistants that synthesize information from many business data sources to answer questions and resolve issues. Their sources include historical support cases, websites, product documentation, inventory records, emails, and old chat conversations with resolutions. The virtual agent combines these sources into specific, reasoned, and actionable responses that are based on the provided data and aligned with business expectations. By summarizing across sources instead of searching one document at a time, the virtual agent optimizes data analysis for both customers and support teams.
Understanding complex customer requests is the ability to interpret full context, language nuances, and different input types in a support conversation. A generative AI chatbot understands the full context of a customer's issue, while a traditional AI chatbot primarily focuses on specific keywords. The generative AI chatbot also accepts the customer's preferred input method, including text, voice, and image, and uses all inputs together to improve resolution. For example, a customer can share a photo of a damaged product during the chat, and the chatbot combines the photo with the written description to improve diagnostic and troubleshooting accuracy.
The two agent types differ in how they interpret input and how they answer:
| Agent type | How it understands input | How it responds |
|---|---|---|
| Traditional AI chatbot | Focuses on specific keywords | Matches keywords to a fixed answer |
| Generative AI virtual agent | Interprets complex language, nuances, and full context across text, voice, and image | Synthesizes multiple data sources into a specific, reasoned response |
Exam tip: Choose a traditional chatbot pattern when matching keywords is enough, and choose a generative AI virtual agent when the task requires full context, nuance, or multimodal input such as a photo plus a description.
The case management and insight-generation workflow is the automated process where a chatbot creates, enriches, prioritizes, and routes support tickets. The support ticketing system receives data directly from the generative AI chatbot after every interaction, so agents do not retype customer information. The chatbot analyzes urgency, sentiment, and complexity to prioritize the ticket, which ensures that urgent or complex issues move ahead of routine requests. For new and complex issues, the chatbot assigns the ticket to the correct team with supporting context such as a summary of the issue and conversation.
Data flows from the chat interaction to the support ticketing system and then to the assigned team. The chatbot creates and populates each ticket with the information the support team needs to act:
application/json and a response schema to convert unstructured narratives into structured fields such as Day, Forecast, Temperature, Humidity, and WindSpeed.Embeddings are vector representations of content that let a system find, group, and compare similar items across a large collection. Text embeddings accept input text with an optional output dimensionality defaulting to 768 dimensions, image embeddings accept raw image data or a Cloud Storage URI with optional contextual text and dimension choice, and video embeddings accept a Cloud Storage URI with optional start and end offset times in seconds, optional contextual text, and dimension choices of 128, 256, 512, or 1408 defaulting to 1408. Generated embeddings can be sent to a Cloud Storage bucket or a BigQuery dataset, so traffic flows from the embedding process to storage and then to discovery and summarization.
Controlled generation is the practice of asking a generative model to return output in a defined structure instead of free-form prose. The request sets the response MIME type to application/json and provides a response schema that names the required fields, such as using the model gemini-2.5-flash to convert an unstructured narrative into a forecast array with Day, Forecast, Temperature, Humidity, and WindSpeed where some fields are required and others are nullable when the source does not provide a value. This structure turns business documents into records that can feed reports, dashboards, and downstream automation.
A traditional AI chatbot focuses on specific keywords and matches keywords to a fixed answer, while a generative AI virtual agent interprets complex language, nuances, and full context across text, voice, and image and synthesizes multiple business data sources into a specific, reasoned, and actionable response. The virtual agent draws on sources such as historical support cases, websites, product documentation, inventory records, emails, and old chat conversations with resolutions, and it can combine inputs such as a photo of a damaged product with a written description to improve diagnostic and troubleshooting accuracy.
Data flows from the chat interaction to the support ticketing system and then to the assigned team, with the support ticketing system receiving data directly from the generative AI chatbot after every interaction. The chatbot analyzes urgency, sentiment, and complexity to prioritize the ticket and assigns new and complex issues to the correct team with supporting context such as a summary of the issue and conversation. Each ticket is created and populated with customer details, technical issue categorization and priority, the full transcript of the conversation, and a summarization of the core issue for fast review.
A global retail company wants to streamline its digital marketing operations and deliver highly engaging customer touchpoints. The creative team currently spends extensive manual effort drafting localized promotional text, designing matching product imagery, producing short video ad scripts, and tailoring promotional assets to individual customer preferences.
Which generative AI capability should the organization implement to automate these workflows and deliver personalized experiences?
Last updated