Prepare and test your skills
Prepare and test your skills
Defining core generative AI concepts means explaining how artificial intelligence nests into smaller fields and how different foundation models create new content. Generative AI (opens in a new tab) is a subset of machine learning, which is a subset of artificial intelligence. Traditional AI classifies information or predicts outcomes from historical data, while generative AI learns patterns from training data to generate new text, images, audio, video, or code.
The hierarchy of artificial intelligence places machine learning inside artificial intelligence and generative AI inside machine learning. Artificial intelligence covers systems that learn from data to classify, predict, or generate. Machine learning provides models that learn patterns from data rather than following only hand-written rules. Natural language processing (NLP) is the area of AI that works with human language, and large language models perform NLP tasks such as text generation, translation, summarization, and question answering.
Traditional AI built with machine learning focuses on prediction from historical sources such as tabular data and images. Common tasks include classification that assigns a label such as spam or not spam, regression that predicts a number such as a house price, time series forecasting that forecasts sales and demand, and clustering that groups records such as customer segments. Generative AI expands beyond labeling to create summaries, uncover hidden correlations, and generate new content that resembles its training data.
Exam tip: Traditional AI classifies or predicts from historical data, while generative AI creates new content that reflects patterns in its training data.
Large language models, multimodal models, and diffusion models are three foundation model capabilities that differ by input types and output types. A large language model (LLM) is a text-driven foundation model trained on vast amounts of data that uses embeddings to understand context and generate coherent text. A multimodal model understands and generates content from multiple modalities including text, images, video, and audio. A diffusion model provides image-generation capability by learning visual patterns and creating realistic images from text prompts.
| Capability | Input data types | What it generates |
|---|---|---|
| Large language model | Text | Generated text, translations, summaries, and answers |
| Multimodal foundation model | Text, images, videos, and audio together | Responses that combine modalities, such as text with related images |
| Diffusion-based image model | Text descriptions, plus images for editing | New images, edited images, and short videos or animations |
Embeddings are arrays of floating point numbers called vectors that convert text, images, and video into a form models can process. Embeddings map data into a continuous vector space where closer proximity means similar meaning. Data flows from raw text and images into vectors, large language models then operate on language vectors to discern language patterns, and diffusion models operate on visual vectors to discern visual patterns.
Prompt engineering and model tuning are two ways to improve model output that differ by what changes and how much effort is needed. Prompt engineering changes only the request sent to the model, while model tuning changes the model itself by training it on a dataset. The practical order is fixed: first try prompt engineering, then move to tuning only if prompt design alone cannot handle the specialized task.
Prompt engineering is the iterative process of writing a prompt, reading the response, and adjusting the prompt until the output matches the goal. Common techniques are zero-shot prompting with no examples, one-shot prompting with one example, and few-shot prompting with multiple examples. Model tuning customizes default behavior through methods such as supervised tuning, reinforcement learning from human feedback (RLHF) tuning, or model distillation. A tuned model can then follow simpler prompts, which can reduce request cost and latency, but the tuned version must be deployed to an endpoint before use.
Describing the machine learning approaches means comparing supervised learning from labeled answers, unsupervised learning from unlabeled structure, and reinforcement learning from reward feedback. Machine learning (opens in a new tab) models learn from data in these three ways to predict, group, or act. The guide for learning is different in each case: correct labels, inherent patterns, or cumulative reward.
Supervised learning is a machine learning approach that trains a model on labeled data to map inputs to desired outputs. Each labeled example pairs an input with its correct answer, such as a class label for classification or a numeric value for regression. During training the model compares its prediction with the label and adjusts to reduce the difference, so the quality, quantity, and consistency of labels determine prediction quality.
Training quality is checked with evaluation metrics such as precision and recall. Low precision and recall often mean the model needs more training data or has inconsistent labels, while perfect precision and recall can mean the test data is too easy to generalize from. Data is commonly split into training, validation, and test sets, where the training set teaches the model, the validation set checks progress, and the test set measures final performance on unseen data. Overfitting occurs when near-duplicate examples appear in both training and validation, so the model performs well on training data but poorly on new data.
Unsupervised learning is a machine learning approach that finds structure in unlabeled data without correct answers provided in advance. Data flows from a raw dataset into the learning algorithm, and the algorithm returns groups, themes, or rules that describe how the data is organized. Choose unsupervised learning when labels are missing or costly, or when the first goal is to explore what the data contains.
Two main methods differ by the structure they return. Clustering groups similar examples into clusters where members are more similar to each other than to other clusters, such as customer segmentation or grouping support tickets by topic. Association finds rules about items that tend to occur together, such as if a shopper buys bread the shopper often buys butter. Clustering answers which examples belong together, while association answers which items appear together in the same transaction.
Reinforcement learning is a framework in which an agent learns by interacting with an environment to maximize cumulative reward. The agent is the learner that chooses actions, the environment is the system acted upon, and the reward signal tells how good each outcome was. Behavior flows from the agent to the environment, and reward feedback flows from the environment back to the agent to shape future behavior.
The agent aims for total cumulative reward over many steps, so it must balance short-term gains against long-term outcomes. Reinforcement Learning from Human Feedback (RLHF) adds human judgment to the loop so human preferences guide the reward toward helpful, accurate, and safe outputs. Choose this approach when model behavior must align with human values and safety, not just raw task performance.
Identifying the stages of the machine learning lifecycle means following data from ingestion through preparation, training, deployment, and ongoing management. Machine learning on Google Cloud (opens in a new tab) moves in order from scattered raw data to clean tables to trained models to monitored endpoints. Each stage depends on the previous stage, so weak ingestion or preparation limits everything downstream.
Data ingestion and preparation turn scattered raw data into clean training sets. Data ingestion brings outside data into Google Cloud by batch or streaming, while data preparation cleans, transforms, and validates it. Cloud Storage stages raw unstructured files and pipeline templates, BigQuery holds cleaned structured tables for analysis and training, and Dataflow transforms data between storage and analytics using the Apache Beam model.
A typical preparation flow runs in three ordered stages. First, source files land in Cloud Storage buckets for staging using transfer services or the gcloud CLI. Then data flows from the staging bucket through the Dataflow pipeline, where each step cleans, enriches, and reshapes records in order. Then the pipeline writes curated output from Dataflow to destination BigQuery tables for querying and training. Batch ingestion copies bounded sets on a schedule, while streaming ingestion captures continuous events through Pub/Sub or Kafka for near-instant access.
Exam tip: Use Cloud Storage to stage raw and unstructured files, and use BigQuery to hold the cleaned, structured tables that models train on — Dataflow moves and transforms data from the first to the second.
Model training and deployment build or tune a model and then connect it to resources to serve predictions. Vertex AI is the central machine learning platform for training and deploying models and AI applications. Training follows two paths: pre-built models for common tasks that accelerate time to value, and custom training that tailors a model to specialized needs.
Tuning for specialized tasks follows an ordered process. First, build a training dataset that represents the task. Then select a tuning method such as supervised tuning, RLHF tuning, or model distillation. Then train the model with the chosen method and dataset. Deploying a model associates the model with an endpoint and machine resources so it can serve online, low-latency predictions. Google foundation models already have project-specific endpoints and are immediately available, but any tuned version must be deployed before use.
| Deployment environment | Resource management | Control |
|---|---|---|
| Fully managed | Vertex AI instantiates and manages machine type and accelerators | Simpler operation, less fine-grained control |
| Self-managed, for example on Google Kubernetes Engine (GKE) | Team manages its own serving resources | More fine-grained control |
Model management keeps trained models organized, reusable, monitored, and compliant after training. Vertex AI Model Registry stores, manages, and deploys trained models so previous work can be reused in pipelines or online prediction. Agent Platform Feature Store provides a central repository for organizing, storing, and serving machine learning features for online and offline use. Versioning of pipelines, containers in Artifact Registry, and datasets ensures reproducibility and auditability.
Monitoring tracks pipeline execution and deployed models over time. Cloud Logging and Cloud Monitoring watch pipeline execution, while Vertex AI monitoring watches deployed models for performance changes and data drift in prediction requests. MLOps pipelines automate transformation, training, deployment, and versioning, so dependencies between steps are defined and data flows automatically. Governance through Knowledge Catalog, BigQuery sharing, and access controls keeps data discoverable and compliant while enabling reuse.
Choosing a foundation model for a business use case means filtering by technical fit, then by operational cost and reliability, then by security and customization needs. A foundation model is a candidate only when its modality matches the data and its context window, performance, cost, and compliance fit the workload. A model with excellent quality is not valid if it cannot accept the input type or hold the input size.
Modality, context window, and performance are the first technical filters for model selection. Model modality is the content type a model accepts and produces, where a text-only model fits summarization and drafting and a multimodal model fits tasks that combine image and text. The context window is the maximum information measured in tokens that a model can consider in one request. Performance metrics include latency from request to response, throughput of requests handled, quality of output, and reliability of availability.
Matching follows an ordered check. First, identify the required input and output types and keep only models whose modality matches. Then estimate the largest expected token count and keep only models whose context window can hold it. Then set targets for latency, throughput, quality, and availability and compare remaining models. Then choose the least expensive or simplest model that still meets all three. Larger windows and larger models generally increase cost and response time.
Exam tip: Modality and context window are pass-or-fail filters, while performance metrics are compared among the models that pass.
Reliability and cost-efficiency decide whether a valid model can stay available under changing demand without overspending. Reliable deployment handles workload fluctuations, supports high availability, and recovers from failure. Google Cloud Observability collects monitoring, logging, and profiling data to reveal trends, while multi-region storage such as Cloud Storage dual- or multi-region buckets replicates data automatically. Autoscaling adjusts resources to demand so capacity exists during peaks and waste falls during idle periods.
Pricing models determine the billing rate for those resources. Committed use discounts (CUDs) give discounts for committing to usage over one year or three years. Resource-based CUDs commit to a quantity of Compute Engine resources in a specific region, while spend-based CUDs commit to spend on a product at the billing-account level. Cost Management tools such as budgets, alerts, tagging, and labeling track spending by project or department, and architectures should be reviewed regularly as pricing and features change.
Security and customization decide how a model must protect data and when it must be adapted. The shared responsibility model divides duties so Google Cloud protects the underlying hardware, storage, and network, while the customer protects its data, identities, access, and configuration. Compliance rules such as HIPAA, GDPR, PCI DSS, FedRAMP, CJIS, and CCPA can block a use case even when the financial case looks strong.
Data residency controls where model data can be stored and processed, enforced with resource location constraints that limit deployments to pre-approved regions. The enterprise foundations blueprint (opens in a new tab) provides deployable Terraform assets for governance, security controls, visibility, and shared services using policy, architecture, and detective controls. Customization starts from the recommended baseline and adds fine-tuning only when a general model cannot meet vocabulary, accuracy, or control needs. Foundational protections for identity, network, and data come first, critical regulations come next, and advanced practices such as zero trust come last.
Identifying business use cases for generative AI means grouping work into synthesis and discovery, multimodal creation and personalization, and operational automation. BigQuery (opens in a new tab) and Vertex AI workflows support these patterns by connecting embeddings, controlled generation, and virtual agents to business data. One capability can create original material, summarize complex sources, and trigger follow-on actions.
Information synthesis turns large volumes of unstructured text, images, and video into summaries, structured facts, and insights. Embeddings are vector representations that let a system find, group, and compare similar items, with text defaulting to 768 dimensions and video supporting 128, 256, 512, or 1408 dimensions. Controlled generation asks a model to return structured output by setting the response MIME type to application/json and providing a response schema, such as converting a weather narrative into Day, Forecast, Temperature, Humidity, and WindSpeed fields.
The discovery workflow runs in fixed order. First, ingest unstructured content. Then generate embeddings for each item. Then collect embeddings and store them in Cloud Storage or BigQuery. Then use controlled generation with a schema to extract structured fields. Then apply safety settings such as BLOCK_LOW_AND_ABOVE for dangerous content, harassment, hate speech, and sexually explicit content. Only after embeddings are stored can large-scale comparison happen, and only after controlled generation runs can narratives become analyzable records.
Multimodal creation produces text, images, video, and source code from simple prompts and tailors results to an audience. Text generation creates articles, summaries, one-pagers, slide decks, and marketing copy. Image and video generation create illustrations, product visuals, clips, and personalized messages from descriptions. Source code generation drafts snippets, functions, and starter templates from natural language descriptions.
Personalization tailors wording, visuals, and offers to each customer segment, while proactive marketing uses AI-driven competitive intelligence to detect shifts and respond rapidly. A marketing team provides topic, audience, and key points, and the model drafts the tailored version from existing information. Developers and marketers then review, test, and refine the draft, so automation speeds routine work while people remain responsible for quality, correctness, and security.
Operational automation uses generative AI virtual agents to reduce manual work in analysis and support. Generative AI virtual agents synthesize historical support cases, documentation, inventory records, emails, and old chats into specific, reasoned responses aligned with business expectations. A generative AI chatbot interprets full context, nuances, and multimodal input such as text plus a photo of a damaged product, while a traditional chatbot focuses on keywords and fixed answers.
| 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 full context across text, voice, and image | Synthesizes multiple sources into a reasoned response |
Data flows from the chat interaction to the support ticketing system and then to the assigned team. The chatbot creates each ticket with customer details, technical categorization and priority based on urgency, sentiment, and complexity, full transcript, and summary of the core issue. Prioritization ensures urgent or complex issues move ahead of routine requests.
Describing how data types are used in generative AI means linking data quality, variety, and ownership to model accuracy, competitive advantage, and governance risk. Generative AI applications (opens in a new tab) depend on high-quality, varied data, and proprietary datasets can enable features competitors cannot easily copy. If inputs are inaccurate or incomplete, the model can perform poorly or produce false outputs.
Data quality, variety, and proprietary data are strategic assets for AI performance. Data quality is the accuracy, completeness, and relevance of data used to train, tune, and evaluate. Data variety is the mix of types a model can process, including text, images, video, audio, and structured tabular data. A proprietary dataset is data a business uniquely owns or collects that improves effectiveness or supports differentiated capabilities.
Business implications include competitive advantage from unique data, higher accuracy from better data, efficiency from reliable automation, and improved customer experience from personalization and recommendations. AI investments should start from measurable goals such as efficiency, cost reduction, experience, or advantage and work backward to the solution. Return on investment is tracked through direct financial gains, operational efficiency such as faster time-to-market, and customer metrics such as satisfaction and retention.
Structured, semi-structured, and unstructured data differ by format and primary AI role. Structured data uses fixed rows, columns, and values such as purchase history and sales records and trains traditional AI to classify and predict. Semi-structured data mixes labeled fields with free text such as invoices, receipts, contracts, and emails and supports extraction, compliance checks, and fraud detection. Unstructured data has no fixed schema and includes documents, reviews, call transcripts, threat reports, images, video, and audio that train and ground generative AI.
| Data type | What it looks like | Primary role |
|---|---|---|
| Structured data | Fixed rows, columns, and values | Train traditional AI for classification, regression, forecasting, and clustering |
| Semi-structured data | Identifiable fields plus free-form content | Extract fields, check compliance, and detect anomalies |
| Unstructured data | No fixed schema, multiple modalities | Train and ground generative AI for summary and creation |
Traditional AI fits prediction on structured data, while generative AI fits summarization, creation, and reasoning across modalities. When training data is missing or too small to custom train, the documented path is generative AI with prompt engineering, not custom training. If high control over algorithms, loss functions, explainability, layers, and hyperparameters is required, custom training of traditional AI fits; if faster go-to-market is the priority, generative AI with tuning fits.
Training, fine-tuning, and grounding keep generative AI outputs relevant and trustworthy. Training builds broad foundation capability from vast multimodal data. Fine-tuning adapts the base model to specialized tasks such as classification or entity extraction using curated examples. Grounding connects output at response time to verifiable sources such as internal documentation, which reduces hallucinations where the model gives a confident answer that training data cannot support.
Grounding often uses retrieval-augmented generation (RAG), which relies on embeddings and vector search to add facts to the prompt. First, the user query is converted into an embedding vector. Then the query embedding is compared to document embeddings where closer proximity means similar meaning. Then the system retrieves the most relevant information and adds it to the prompt, so the response reflects facts and current information.
Explaining data quality and accessibility means showing why complete, consistent, relevant, and reachable data produces useful models and confident decisions. Accessible data means complete, relevant data is available in a usable format when the model needs it. When data is hard to reach or stored in inconsistent formats, predictions weaken and leaders decide with less confidence.
Completeness, consistency, and relevance decide whether training data is fit to teach a model. Completeness asks if all needed products, events, and details are present for training and evaluation. Consistency asks if related records line up across systems. Relevance asks if data matches the real language, products, and actions the model must predict.
In AI Commerce Search, completeness is checked on the Data quality page for catalog and user event data, and on the Catalog and Event tabs for import counts and stock status. Consistency is threatened by unjoined events where a user event refers to a product not yet uploaded; such events are still logged and handled but cannot train the model for future predictions, so the unjoined rate should stay very low. Relevance for conversational agents comes from real end-user expressions such as human support logs, support emails, forums, FAQs, and social questions, while long-form non-conversational text and non-user responses should be avoided.
Accessible data in BigQuery improves AI results by letting models work where data already lives instead of waiting for copies. AutoML (opens in a new tab) and BigQuery ML bring machine learning to the data so SQL practitioners can build, evaluate, and run inference inside BigQuery. Data flows from BigQuery tables directly to training, evaluation, and inference steps inside BigQuery, so the workflow depends on one governed platform.
Teams reach the same data through the Google Cloud console, bq command-line tool, REST API, and Colab Enterprise notebooks where SQL and Python can be used interchangeably. Standardized formats let inputs and outputs flow without custom parsing, such as GoogleSQL queries, defined output schemas, and numerical embeddings. General-purpose functions such as AI.GENERATE accept prompts with full control over model and parameters, managed functions such as AI.CLASSIFY, AI.SCORE, and AI.AGG let BigQuery choose settings to optimize cost and quality, and task-specific functions use Cloud AI APIs for translation, document processing, transcription, and vision.
Exam tip: Choose general-purpose functions when you must control the model, prompt, and parameters; choose managed functions when BigQuery can choose the model to optimize cost and quality for routine tasks.
Economic and operational trade-offs compare data collection, inspection, storage, and governance costs against model usefulness and reliability. Sensitive Data Protection inspection of a BigQuery table incurs storage inspection costs, and saving findings to BigQuery adds separate BigQuery charges. Tokenization of personally identifiable information with Dataflow and Sensitive Data Protection happens before data becomes persistent in BigQuery, which reduces exposure.
Operational choices trade control against automation. Manual scaling gives precise control but lacks dynamic responsiveness, while autoscaling adjusts automatically to demand. Model distillation lets a smaller student model learn from a larger teacher model to run on limited devices and potentially reduce overfitting. Vertex AI Model Evaluation assesses performance and bias before deployment, while Vertex AI Model Monitoring watches live prediction traffic for skew and drift after deployment.
Identifying differences between structured and unstructured data means matching data format to AI task, search behavior, and storage system. Structured data follows a fixed schema, while unstructured data has no fixed table-like schema. Data stores (opens in a new tab) in Agent Search, formerly Vertex AI Search, hold only one type of data, so the type determines how information is organized, searched, and used.
Structured and unstructured data support different AI strengths because of their formats. Structured data is organized into tables, spreadsheets, comma-separated values files, BigQuery tables, or DataFrames with precise fields such as transaction_id, user_email, and transaction_date. Unstructured data includes text, images, audio, video, and code such as documents, reviews, call conversations, invoices, contracts, and product images. Structured metadata describes both types with definitions, owners, sensitivity, and quality rules, which helps ground generative AI and reduce hallucinations.
| Data type | Format and examples | Typical AI use |
|---|---|---|
| Structured data | Tables and records with fixed fields | Traditional AI classification and regression, such as spam filtering or price prediction |
| Unstructured data | Free-form text, images, audio, video, code | Generative AI creation, summarization, translation, chatbots, and code generation |
Structured data is often used by traditional AI for classification and prediction, while unstructured content is the main training material for generative AI. Knowledge Catalog builds a dynamic context graph from both types and powers BigQuery governance for discovery, quality checks, and secure sharing.
Real-world examples show how each data type appears in search applications. Structured data examples include financial transactions, inventory tables, ecommerce product catalogs, doctor directories, and hotel records with fields such as id, title, location, available_date, rating, and room_types. Unstructured examples include reports, scanned PDFs, infographics, images with metadata, videos, news articles, music files, and podcasts in formats such as TXT, PDF, HTML, DOCX, PPTX, and Markdown.
Special types refine this picture. Structured media data is structured data with a media-specific schema where each record includes title, URI to content location, categories, duration, and available date. Website data is indexed from public URLs where excluded patterns take priority over included patterns, and it can combine unstructured pages with structured meta tags and schema.org data. Healthcare FHIR data is structured R4 resources from Cloud Healthcare API, and Agent Search for healthcare is deprecated after May 15, 2027.
Storage architectures hold each format and define how it is queried and scaled. Data format is how information is organized, while storage architecture is the system that holds and serves it. Structured data with predefined schemas fits systems with schemas and query languages, while unstructured files fit object storage that holds whole objects.
| Storage architecture | Data model | Best for |
|---|---|---|
| Relational database, for example Cloud SQL | Tables, rows, columns, primary keys | Transaction processing with SQL, joins, and ACID transactions |
| Firestore in Datastore mode | Kinds, entities, properties, keys | Highly available schemaless data at scale without joins |
| BigQuery | Tables with schemas | Analytical querying over structured data |
| Cloud Storage | Objects in buckets | Large blobs and unstructured documents |
For search ingestion, structured data must be JSON Lines or NDJSON under size limits, and unstructured files can be up to 200 MB with up to 100,000 files per import. Layout parsing helps non-searchable PDFs, OCR helps searchable PDFs with many tables, and retrieval-augmented generation requires chunking enabled at data store creation. When refreshing structured data, the new schema must be the same or backward compatible, where adding only new fields is backward compatible.
Identifying differences between labeled and unlabeled data means comparing data with correct answers attached to raw data without answers. Labeled data pairs each input with a ground-truth tag, while unlabeled data is raw content with no predefined category. Machine learning on Google Cloud (opens in a new tab) uses labeled data for supervised prediction and unlabeled data for pattern discovery and large-scale pre-training.
Labeled data teaches supervised learning by showing both the question and the correct answer. An invoice image paired with its correct total, or a form paired with its correct field values, gives the model ground truth to imitate. The model reviews many historical examples, compares its output to the tags, and updates its settings to reduce errors. Two common tasks are classification that predicts a category and regression that predicts a continuous number.
Labeling fits into document preparation in fixed order. First, create a processor such as an Invoice Parser. Then import a dataset or configure uptraining. Then label example documents with correct field values. Then train or uptrain the processor on those examples so it can handle new documents. Machine learning labels are ground-truth tags on training examples, which differ from Google Cloud resource labels that are key-value pairs on resources for organization and billing.
Exam tip: ML labels are ground-truth tags attached to training examples, while Google Cloud resource labels are key-value pairs attached to resources to organize them and break down billing charges.
Unlabeled data reveals structure without labels through similarity and hidden relationships. Unsupervised learning groups similar items without being told the groups, so clusters emerge from the data. Self-supervised learning creates its own signals from the data structure to learn useful representations without human labels. The result maps the data to guide labeling, search, or downstream training.
Practical tools make unlabeled data searchable at scale. Vector embeddings represent each item as numbers where similar items sit close together, and Vector Search indexes with metadata make those items searchable by similarity. Unlabeled content is often unstructured images, documents, audio, or free text stored in Cloud Storage, governed without moving it through BigQuery object tables, signed URLs, and Knowledge Catalog. Shared Agent Platform datasets, Feature Store, and Model Registry let teams reuse datasets, features, and models instead of collecting new copies.
Generative AI lowers data cost by learning first from massive unlabeled datasets and then specializing with a much smaller labeled dataset. Pre-training builds broad language, style, and multimodal patterns once with high compute, storage, and time. Fine-tuning adapts that base model to a specific task with much less data and compute through transfer of knowledge.
| Training stage | Data used | Resource profile |
|---|---|---|
| Pre-training | Vast, largely unlabeled datasets across modalities | High compute and time, done once by provider |
| Fine-tuning | Smaller, task-specific labeled datasets | Much less data and compute, done by adapting team |
Model Garden provides pre-trained starting points including Gemini, Imagen, Gemma, Llama, and partner models. Specialized behavior can then be added with supervised tuning for classification, sentiment analysis, or entity extraction, or with prompt engineering using a few examples. When custom labeled data is scarce, generative AI with prompt engineering and small supervised tuning is preferred over training a complex model from scratch.
Continue with this study guide: Identifying the differences between labeled and unlabeled data
A left-to-right process flow showing five ML lifecycle stages from data ingestion in Cloud Storage through Dataflow and BigQuery to Vertex AI training, endpoint deployment, and model management.
Traditional artificial intelligence classifies information or predicts outcomes from historical data, while generative AI learns patterns from training data to generate new text, images, audio, video, or code. Artificial intelligence contains machine learning, and machine learning contains generative AI.
A large language model takes text in and produces text, a multimodal foundation model combines text, images, video, and audio, and a diffusion model generates or edits images from text prompts. A large language model uses embeddings to understand context and generate coherent text, while a diffusion model learns visual patterns to create realistic images from text prompts.
Try prompt engineering first and move to tuning only if prompt design alone cannot handle the specialized task. Prompt engineering changes only the request sent to the model, while model tuning changes the model itself by training it on a dataset.
Retrieval-augmented generation grounds responses by converting the user query into an embedding vector, comparing it to document embeddings by proximity, and adding the most relevant information to the prompt. This connects output at response time to verifiable sources, which reduces hallucinations where the model gives a confident answer that training data cannot support.
Last updated