Databases in the AI Era

Vector Search, RAG, and the Era of Good Data

This is the final part of 12 in "Databases for Non-Developers". It covers how databases are changing in the AI era, and why "good data" matters more than ever.

---

An employee asks the company AI, "How do I apply for parental leave?" The AI looks through company policies, HR guides, and FAQs, then answers. What matters here isn't just the AI's ability to produce natural-sounding language. It's the data foundation that finds the right material and delivers the parts that are needed.

And with that, the whole series comes full circle. Part 1's "why manage data well" and Part 6's "Garbage In, Garbage Out" return in the AI era, louder than before.

In the AI Era, GIGO Gets Amplified

In Part 6 we saw the GIGO principle: put bad data in, get bad results out. In the AI era, this principle doesn't disappear — it gets amplified. There are three reasons.

1. Amplified speed When a person copies wrong numbers into a report, it takes time — and along the way, someone has a chance to say, "Wait, these numbers look odd." AI spreads polluted data into thousands of answers in seconds. Errors now outrun human review.

2. Amplified plausibility A wrong number in a spreadsheet looks suspicious. But AI delivers even wrong content as a smooth, confident sentence. The tendency of generative AI to make up plausible answers with no basis is called "hallucination" — but an AI built on bad data is wrong even without hallucinating, because the data itself is wrong. And because the error arrives in fluent prose, it's that much easier to believe.

3. Baked-in learning With an analytics system, fix the data and the results change right away. But an AI model trained on bad data doesn't improve just because the data gets fixed later. Biased, polluted patterns harden inside the model. This is less "let's clean up the data and then adopt AI" and more "an AI adopted without clean data makes the bad data part of its nature."

Three Data Questions AI Must Answer

Generative AI can produce plausible sentences, but it doesn't automatically know your company's latest refund policy or a specific customer's actual order status. AI projects often focus only on model selection and miss these questions:

• Which material will be used as the basis for answers?
• How do we tell current material apart from retired material?
• Who is responsible for the accuracy of documents?
• Does search return only material the person asking is allowed to see?
• Can we show users the sources and original text behind an answer?

These look like new questions, but they're really an extension of data quality (Part 6), governance (Part 7), and security (Part 9). They aren't new challenges of the AI age — they're existing principles that have simply grown more important.

A Vector Is a Numeric Representation for Comparing Meaning

AI systems can turn sentences or images into bundles of numbers called embeddings. Material with similar meaning is placed close together in this numeric space. Vector search finds related material by comparing these distances.

Even if a user asks, "I want my money back," the system can still find the "refund procedure" document. It overcomes the limits of traditional search, which requires exactly matching words.

But being close in meaning is not the same as being true or being something you're allowed to use. Relevance, accuracy, currency, and permission each have to be checked separately.

RAG Makes AI Look Up Material Before Answering

Retrieval-augmented generation (RAG) is an approach where the system first finds material related to the question, then hands that material to a language model to craft an answer.

1. The user asks a question
2. The system searches for related material
3. The found material and the question are sent to the AI together
4. The AI builds an answer based on the material
5. Where possible, the sources are shown as well

RAG helps companies use their internal material and cite their sources, but it doesn't eliminate wrong answers entirely. The search can go wrong, or the model can interpret even the right material differently. And above all — if the material being retrieved is outdated or wrong, RAG faithfully amplifies that error. For important decisions, checking the original text and human review are still necessary.

Will Vector Databases Replace Traditional Databases?

For most services, no. Information where exact values and relationships matter — order amounts, inventory counts, user permissions — stays in relational databases and the like. A hybrid structure that adds vector capabilities on top for searching the meaning of documents is common.

AI and Personal Data

If you feed customer service conversations into an external AI service, removing names may not be enough — a combination of an order number, part of an address, or an unusual situation could still identify the person (the re-identification risk from Part 8). Depending on the service and contract, retention periods for input data, whether it's used for training, and where it's processed all differ.

These principles are needed:

• Send only the information that's truly necessary
• Remove or mask sensitive values
• Check the service's data-use terms and retention period
• Apply per-user document access permissions from the search stage onward
• Minimize sensitive information even when logging questions, retrieved material, and answers
• Review important answers alongside their source documents

The Risks of an Era Where You Ask a Database in Plain Language

AI can turn "show me last month's sales" into a database query. That's a genuine benefit — it widens non-developers' access to data. At the same time, it introduces risks: wrong calculations, queries outside someone's permissions, exposure of sensitive information, and expensive queries being run.

You need read-only permissions, allowed data scopes, verified metrics, previews before execution, limits on cost and usage, and a process for reviewing results. Just because plain language is easy doesn't mean the meaning of the data becomes automatically clear. Even the single phrase "last month's sales" hides Part 6's questions: "Does it include cancelled orders?" "Which system is this the source of truth from?"

In the End, It Comes Back to Human Questions

Even in the AI era, the important questions barely change. If anything, only organizations that can answer these questions will fully reap AI's benefits.

• Where does this data come from?
• Who is responsible for its accuracy?
• Is it still up to date?
• Is this person allowed to see it?
• Can we check the basis for the answer?
• What damage would be done if it were wrong?
• When is human review needed?

Key takeaways Vector search finds material with similar meaning, and RAG provides found material as the basis for AI answers. But relevance doesn't guarantee accuracy or permission. In the AI era, GIGO (Garbage In, Garbage Out) doesn't disappear — it's amplified, through three paths: speed, plausibility, and baked-in learning. Good AI doesn't start with a good model. It starts with good data — that is, with the existing principles of quality, governance, and security.

AI Data Project Checklist

• Are the official materials to use for answers — and the ones to exclude — clearly defined?
• Can outdated documents be found automatically or on a regular schedule?
• Do documents users can't see stay out of search results?
• Does the answer show the original source and its date?
• Does sensitive information stay out of external models, logs, and evaluation data?
• Are high-stakes questions handed off to a human?
• Is quality evaluated regularly using sets of real questions?

Closing the Series — Knowing Databases Means Being Able to Ask Good Questions

Non-developers don't need to install database products or write SQL themselves. But in an era where data drives decisions in business and daily life, you need to be able to ask a few questions.

• What are we collecting, and why?
• Which data is the source of truth?
• What does this number actually mean?
• Who can see it and change it?
• If it's wrong or goes missing, how do we notice and recover?
• How long do we keep it, and how do we delete it?
• If it's breached, who is harmed and how?
• Is it okay to hand this data to AI?

A good database isn't a system piled high with the latest technology. It's a system that remembers the facts accurately, delivers them to the right people at the right time, limits unnecessary exposure and misuse, and can recover when something goes wrong.

A database isn't equipment sitting in the tech team's basement. It's where a product's promises, an organization's memory, and customer trust meet. And in the AI era, the quality of that memory becomes the organization's intelligence.

---

Series contents

1. What Is a Database — How Is It Different from an Excel File?
2. How Databases Keep Data Accurate
3. What Kinds of Databases Are There?
4. Databases, Data Warehouses, and Data Lakes
5. Why Choosing the Right Database Matters
6. Data Models and Data Quality
7. What Is Data Governance?
8. Less Personal Data — Only What You Need
9. The Basics of Data Security
10. Why Do Data Breaches Happen, and What Should You Do When One Does?
11. Cloud and Serverless Databases
12. Databases in the AI Era ← You are here