Prepare and test your skills
Prepare and test your skills
Data accessibility is whether AI services can reach needed data in a usable format, at an acceptable cost, and with enough availability. Infrastructure requirements are the compute, storage, and networking resources that support each stage of the machine learning (ML) lifecycle. In Google Cloud, the two are linked because the performance, cost, and scalability of AI applications depend directly on the underlying infrastructure.
Data accessibility in Google Cloud AI means placing data where AI services can use it without extra copying or reformatting. BigQuery ML (opens in a new tab) brings ML to the data, so SQL practitioners can train, evaluate, and run inference using existing SQL tools and skills. This approach speeds model development and makes ML usable by more teams. BigQuery AI adds predictive ML, inference against large language models (LLMs) such as Gemini, embeddings and vector search, built-in agents, data pipelines, and agent tools. Users can work with BigQuery ML through the Google Cloud console, the bq command-line tool, the REST API, or Colab Enterprise notebooks.
Availability also affects accessibility. Some BigQuery AI features may not be available when using reservations created with certain BigQuery editions. Edition and reservation choices therefore determine which AI capabilities can reach data already stored in the platform.
Infrastructure for AI and ML must match the needs of each ML lifecycle stage for compute, storage, and networking. Training needs different resources than inference, and storage must deliver data without creating bottlenecks. Google Cloud provides design guidance for this planning:
The right choice depends on the workload stage and its specific compute, storage, and networking needs. Poor storage planning raises cost and slows training and inference, while well-matched infrastructure keeps data available to AI processes.
Data formats determine whether data can flow smoothly into Google Cloud AI services. BigQuery AI functions accept natural language inputs and work with structured and unstructured data, including text, documents, and images. They can generate structured output, such as extracting names, addresses, or object descriptions from text, documents, or images. They can also generate embeddings for semantic search and clustering, and table-valued functions can add columns of embedded text, image, audio, video, or document data to a table.
For integration, BigQuery AI functions access Gemini and partner LLM models from Gemini Enterprise Agent Platform, Cloud AI APIs, or built-in BigQuery models. Task-specific functions use Cloud AI APIs for natural language processing, machine translation, document processing, audio transcription, and computer vision. When the data format matches what the service can read directly, fewer conversion steps are needed and data stays more accessible.
Cost and availability shape which data-access choices are practical. BigQuery AI functions fall into general-purpose, managed, and task-specific categories that trade control for cost and simplicity:
| Function category | Control and cost behavior | Common use |
|---|---|---|
| General-purpose AI functions | Full control over model, prompt, and parameters | Custom analysis and generation |
| Managed AI functions | Streamlined syntax; optimized for cost and quality; BigQuery chooses a model | Routine filtering, rating, classification, and aggregation |
| Task-specific functions | Use Cloud AI APIs | Natural language processing, translation, document processing, audio transcription, computer vision |
BigQuery also provides the AI.COUNT_TOKENS utility function to estimate tokens in an input prompt before running a query. Pricing, token usage, cost tracking, monitoring, and error management help control AI costs. Location limits and BigQuery edition limits also apply, so the same data may be reachable through some AI features but not others.
Exam tip: BigQuery ML brings ML to the data instead of requiring data movement, which keeps data accessible when it already lives in BigQuery.
Data quality for model integrity is measured by completeness, consistency, and relevance, while format, availability, and cost decide how usable a dataset is. A generative AI model produces accurate and less biased outputs only when the data used to evaluate and operate it is whole, uniform, and on-topic. In Google Cloud, these dimensions are supported by the Gen AI evaluation service, Sensitive Data Protection, and data residency controls.
Completeness means a dataset contains every record and field a task requires, with no missing prompts, variables, or responses. In the Gen AI evaluation service (opens in a new tab), a dataset loaded for evaluation must contain either prompts or records to use in a prompt template and, optionally, model responses. An uploaded CSV or JSONL file is limited to 200 rows. When prompts are generated instead of uploaded, the service fills the variables defined in the prompt template, and the number of samples to generate can be set. Missing fields or empty variables leave gaps that metrics cannot score, which weakens evaluation and can hide bias.
Consistency means data follows the same format, structure, and location rules through the whole pipeline. The Gen AI evaluation service accepts datasets in CSV or JSONL format, so mixing formats breaks the load step. Location consistency is enforced by data residency rules. An AML AI Instance resource is created with a region, and every artifact created under that parent stays in the same region. AML AI also requires all input and output data to be in the same region and project, which prevents accidental movement between Google Cloud regions. Organizational constraints for resource locations control where Google Cloud resources can be created and are applied at resource creation time.
Relevance means the data measures what the model is actually judged on, so scores reflect the behavior that matters. In the Gen AI evaluation service, at least one metric is added to score the quality of a candidate's responses, and the chosen metric determines what the evaluation rewards. For some adaptive rubrics, the rubric generated from each prompt can be steered by opening Advanced and adding custom instructions, such as Evaluate the dataset on cultural sensitivity. Adding a comparison candidate scores two candidates side by side, which makes differences in relevance visible. Relevant data reduces bias because the evaluation signal matches the intended task rather than an unrelated pattern.
Format and availability determine whether data can reach the model at all. The Gen AI evaluation service reads CSV or JSONL files and prompt templates with variables, so data in another format must be converted first. Access depends on identity and access management (IAM). Creating a project requires the Project Creator role, which contains the permission to create projects, and the evaluation workflow requires the Storage Admin role on the project. Evaluation results are written to a Cloud Storage bucket named during setup. Data that is complete, consistent, and relevant still cannot improve a model if the service cannot read its format or the user lacks access.
The Gen AI evaluation service turns data quality checks into a repeatable workflow that runs from the Google Cloud console and produces a score per prompt alongside the model response. The ordered workflow is:
A left-to-right five-step process flow showing load dataset, add candidate, add metric, name and choose storage, then evaluate to get per-prompt scores in Cloud Storage.
Partner models such as Anthropic and Llama can also be evaluated. A partner model must be enabled in Model Garden before selection during setup, and pricing for third-party models is based on model inference charges in Model Garden.
Data quality also depends on protecting sensitive data so it stays accurate and is not exposed. Sensitive Data Protection (opens in a new tab) classifies confidential data such as credit card numbers, addresses, phone numbers, and personally identifiable information (PII). Data can be tagged and tokenized before storage in Google Cloud. Knowledge Catalog stores, manages, and provides access to metadata about that data. VPC Service Controls places sensitive data in its own service perimeter and controls movement across the perimeter, which reduces data exfiltration from Google-managed services. IAM access controls and multifactor authentication (MFA) protect users who need access to sensitive data.
Model integrity depends on how the platform handles customer data during processing. Document AI (opens in a new tab) does not use customer content such as documents and predictions for any purpose except to provide the service, and never uses customer data to train models. Google does not claim ownership of content sent to Document AI, and documents and custom models are treated as private customer data. Retention depends on processing type:
| Processing type | Where data is processed | Retention |
|---|---|---|
| Batch | Stored encrypted with an ephemeral key so no human has access | Typically deleted immediately after processing, with a failsafe time to live of one day; abnormal ends may persist up to one day |
| Synchronous | Processed in memory and encrypted in flight | Not persisted to disk; request metadata such as time received and request size is logged temporarily |
Customers control Google support access through Access Approval, which starts with a support ticket and lets the customer authorize or deny access, and Access Transparency, which shows what support does while access is granted. Document AI also supports data residency, deny policy, VPC Service Controls, and customer-managed encryption keys (CMEK).
Exam tip: Completeness means having all required records and fields, while consistency means those records follow the same format and region rules — a dataset can be complete yet inconsistent if rows use different formats or live in different regions.
Data readiness means organizational data is complete, consistent, relevant, available, and stored in a format that AI systems can use. In Google Cloud, BigQuery (opens in a new tab) is where data readiness becomes business results, because its AI functions accept natural language inputs and use pre-trained Gemini Enterprise Agent Platform models and built-in BigQuery models. When data is ready, teams build generative AI faster, spend less on processing, and get outputs they can trust.
Ready data lowers cost because AI work happens where the data already lives. BigQuery ML brings machine learning to the data instead of requiring data movement, which removes copy costs and speeds model development. Managed AI functions are optimized for cost and quality, and BigQuery can choose the model and parameters so teams do not overpay for unneeded control. For large datasets, optimized mode for AI.IF and AI.CLASSIFY trains a lightweight distilled model on a sample of data and then uses that model for inference on most rows. This avoids calling a large language model for every row and reduces token use and latency across thousands or billions of rows. The AI.COUNT_TOKENS utility estimates tokens in a prompt before running a query, so cost is known in advance.
Reliable generative AI depends on data that is accurate and current, not only present. Retrieval-augmented generation (RAG) (opens in a new tab) improves factual accuracy and response consistency by referencing additional data from a trusted source, including data newer than model training data. Managed AI functions improve consistency by using prompt engineering and selecting the right model and parameters for the task. Embeddings, which are high-dimensional numerical vectors that represent entities such as text or audio, capture data semantics so content can be compared and reasoned about. Embeddings support search, recommendations, and record deduplication, which further raise output reliability.
The format and availability of data decide which AI capabilities a team can use. BigQuery AI functions can analyze structured data and unstructured data such as text, images, audio, video, and PDF files. General-purpose AI functions handle any structured or unstructured data, while managed AI functions analyze text, image, audio, video, or PDF data without customization. Embedding generation supports multimodal search, such as using text to search images, plus semantic search for recommendations, substitution, and deduplication. Accessible formats therefore widen the set of generative tasks a team can run without extra pipelines.
The two main categories of generative AI functions in BigQuery differ in team control and automatic optimization:
| Function type | Control | Choose when |
|---|---|---|
| General-purpose AI functions | Full control and transparency over model, prompt, and parameters | The task needs a specific model or prompt, or output must follow a custom schema |
| Managed AI functions | Streamlined syntax; BigQuery selects model and parameters to optimize cost and quality | The task is routine filtering, rating, classification, or aggregation |
General-purpose functions include AI.GENERATE for flexible inference, AI.GENERATE_TEXT for table-valued inference with partner and open models, and AI.EMBED and AI.GENERATE_EMBEDDING for embeddings. Managed functions include AI.IF for filtering, AI.SCORE for rating, AI.CLASSIFY for categorization, and AI.AGG for aggregation.
Exam tip: Managed AI functions let BigQuery choose the model to optimize cost and quality, while general-purpose AI functions keep full control over model, prompt, and parameters.
Completeness means having all required records and fields, consistency means following the same format and region rules, and relevance means measuring the behavior the model is judged on. A generative AI model produces accurate and less biased outputs only when the data used to evaluate and operate it is whole, uniform, and on-topic.
BigQuery ML brings machine learning (ML) to the data in BigQuery instead of requiring data movement, which removes data-movement cost and speeds model development. This approach lets SQL practitioners train, evaluate, and run inference using existing SQL tools and skills, which makes ML usable by more teams.
Use managed AI functions in BigQuery AI for routine filtering, rating, classification, and aggregation where streamlined syntax and automatic optimization for cost and quality are preferred because BigQuery chooses the model. Use general-purpose AI functions for custom analysis and generation where full control and transparency over model, prompt, and parameters are needed.
A retail organization is preparing large-scale datasets—including unstructured customer feedback documents, audio logs, and catalog images—to train and fine-tune generative AI models on Google Cloud. The architecture team needs to establish a central storage layer that meets the following criteria:
Which Google Cloud service should the organization select as its primary data lake storage foundation?
Last updated