Fundamentals of gen AI · Topic 4 of 6
Data for Gen AI
How data is used in gen AI: data quality and accessibility, structured versus unstructured data, labeled versus unlabeled data, and the business implications.
- 4 min
- Quality: completeness, consistency, relevance. Accessibility: availability, cost, format.
- Structured has a fixed schema; unstructured (documents, images, audio, video) has none and is most of enterprise data.
- Labeled data carries answers and costs more to produce; unlabeled data is raw.
What it is
Gen AI runs on data: to train models, to tune them, and to ground their answers in facts. The exam expects you to explain the characteristics of data quality and accessibility, tell structured from unstructured data, and labeled from unlabeled data, with real examples and the business implications.
How it works
Data quality and accessibility
The exam guide lists six characteristics:
| Characteristic | Meaning | Business implication |
|---|---|---|
| Completeness | No important values or records are missing | Gaps lead to wrong or biased outputs |
| Consistency | The same data agrees across systems and over time | Conflicting facts produce conflicting answers |
| Relevance | The data actually relates to the task | Irrelevant data adds cost and noise |
| Availability | The data can be reached when needed by the right people and systems | Data locked in silos cannot feed AI |
| Cost | Collecting, storing and preparing the data is affordable | Expensive data can make a use case unviable |
| Format | The data is in a form the model or pipeline can use | Scanned paper or proprietary formats need conversion first |
The first three describe quality; the last three describe accessibility. Both are needed: perfect data nobody can reach is as useless as accessible data that is wrong.
"Garbage in, garbage out" applies strongly to gen AI. A model grounded in outdated policies will confidently give outdated answers.
Structured and unstructured data
| Type | What it looks like | Real-world examples |
|---|---|---|
| Structured | Organized in rows and columns with a fixed schema | Sales tables, customer records, inventory, financial transactions |
| Unstructured | No predefined format | Emails, contracts, PDFs, images, audio recordings, videos, chat logs |
Most enterprise data is unstructured. Traditional analytics struggled with it; gen AI and multimodal models are good at reading it, which unlocks value from documents, recordings and images that were previously ignored.
Labeled and unlabeled data
| Type | Meaning | Example | Used for |
|---|---|---|---|
| Labeled | Each example has the correct answer attached | Photos tagged "damaged" or "undamaged"; tickets tagged by category | Supervised learning and tuning |
| Unlabeled | Raw data with no answers attached | A folder of photos; a database of support tickets | Unsupervised learning, pre-training foundation models, grounding |
Labeling costs time and money, often needing experts. Foundation models reduce the need: they learn from huge unlabeled datasets, and a small labeled set is often enough to tune them.
How gen AI uses data
| Use | Data needed |
|---|---|
| Pre-training a foundation model | Vast unlabeled data (done by model providers) |
| Tuning a model to a task or style | A smaller set of good labeled examples |
| Grounding answers in facts | Current, accurate enterprise documents and databases |
| Prompting | The context you include in the prompt |
A worked scenario
A telecom company wants an assistant that answers questions about customer contracts. The contracts are unstructured PDFs; billing data is structured in BigQuery. Some contracts are scanned images, a format problem solved with document processing. Two systems disagree on plan names, a consistency problem fixed before grounding. The team labels a few hundred past questions with correct answers (labeled data) to evaluate the assistant.
How the exam asks about this
- A data problem is described and you name the characteristic (completeness, consistency, relevance, availability, cost, format).
- Examples are listed and you sort them into structured or unstructured.
- "Data tagged with correct answers" is labeled; raw data is unlabeled.
What to remember for the exam
- Quality: completeness, consistency, relevance. Accessibility: availability, cost, format.
- Structured has a fixed schema; unstructured (documents, images, audio, video) has none and is most of enterprise data.
- Labeled data carries answers and costs more to produce; unlabeled data is raw.
- Gen AI uses data for pre-training, tuning, grounding and prompting.
High-quality data that the AI system cannot access, because of silos or format, does not help. Plan for accessibility, not just quality.
Common mistakes
Calling a PDF structured because it has headings. Documents are unstructured.
Assuming more data always helps. Irrelevant or low-quality data adds cost and errors.
Underestimating labeling cost. Expert labels can be the most expensive part of a project.
Check yourself
Which three characteristics describe data accessibility in the exam guide?
Availability, cost and format.
Is a collection of recorded customer calls structured or unstructured data?
Unstructured data.
What is labeled data?
Data where each example has the correct answer or category attached.
Topic 4 of 6 in Fundamentals of gen AI
Test yourself on Fundamentals of gen AI
The full Google Cloud Generative AI Leader question bank has 512 questions, with timed practice exams and an explanation for every answer.
This study guide is independent and is not affiliated with, endorsed by, or sponsored by Google Cloud. Vendor names, certification names and exam codes are trademarks of their respective owners and are used for identification only.