Embeddings Explained
The Foundation of How AI Understands Real Estate
Version 1.3
Joe Bjorklund & Jobe Rose
Chicago, IL
Technical Accuracy Statement
This paper explains the concept of embeddings, one of the foundational technologies behind modern artificial intelligence. It is based on established principles from machine learning and natural language processing rather than the implementation details of any single commercial AI system.
Where implementation details of proprietary AI platforms are unknown, this paper says so explicitly rather than speculating. Its purpose is to explain concepts that are broadly accepted within the AI research community using real estate examples that make those concepts easier to understand.
Abstract
Artificial intelligence is changing the way people search for homes, neighborhoods, and real estate information. Buyers are increasingly asking complete questions rather than entering a handful of keywords into a search engine. Understanding how AI answers those questions begins with understanding one of its most important building blocks: embeddings.
Embeddings allow computers to represent concepts mathematically while preserving many of the relationships between them. They make it possible for an AI system to recognize that Lincoln Park is more closely related to Lakeview than to Miami Beach, or that historic greystones, walkability, and mature trees often describe similar neighborhoods even when those exact words are not repeated.
This paper introduces embeddings in plain language and explains why they have become one of the foundations of modern artificial intelligence.
1. The Fundamental Problem
Before defining an embedding, it helps to understand the problem embeddings were created to solve.
People and computers process information in fundamentally different ways. Humans naturally recognize meaning, relationships, and context through years of experience. Computers excel at mathematics, but they do not possess intuitive understanding. To a computer, the phrase Lincoln Park is simply a sequence of characters. It does not automatically evoke tree-lined streets, historic greystones, neighborhood parks, restaurants, schools, or a particular style of city living.
Those associations exist because people have learned them over time.
For an artificial intelligence system to reason about neighborhoods, architecture, or real estate, those concepts must first be represented in a mathematical form. Representing meaning mathematically without losing the relationships between ideas is one of the central challenges of modern artificial intelligence.
Embeddings were developed to help solve that problem.
2. Why Traditional Databases Aren’t Enough
Traditional databases are exceptionally good at storing and retrieving facts. They can quickly answer questions such as which ZIP code contains an address, when a building was constructed, or how many units it contains because those answers already exist as stored records.
Questions involving meaning are different.
Suppose someone asks, “Which Chicago neighborhood feels most like Lincoln Park?” There is no single factual answer stored in a database. Similarity depends on many characteristics, including architecture, parks, walkability, schools, density, restaurants, price, and neighborhood character.
Programming every possible relationship by hand would be impractical. Instead, modern AI systems learn statistical relationships from enormous amounts of information. Embeddings provide one of the mathematical structures that make those learned relationships useful.
3. What Is an Embedding?
An embedding is a numerical representation of a concept that enables a computer to compare that concept with other concepts mathematically.
That definition is simple, but its implications are significant.
Rather than assigning a single identification number to a word or place, modern AI represents concepts as collections of numerical values. Those values are not intended for people to interpret directly. Their purpose is to preserve relationships.
Concepts that appear in similar contexts during training tend to develop similar mathematical representations. As a result, related neighborhoods, architectural styles, landmarks, and other ideas become positioned closer together within the model’s representation of information.
This allows an AI system to compare concepts based on learned relationships rather than exact keyword matches.
4. Why This Matters
Imagine someone asks an AI assistant:
“I’m looking for a quiet Chicago neighborhood with historic homes, mature trees, excellent walkability, and convenient access to Lake Michigan.”
The question never mentions Lincoln Park, the Gold Coast, or Lakeview.
Yet an AI system may recognize those neighborhoods as relevant because it has learned relationships between concepts such as historic homes, walkability, mature trees, and specific areas of Chicago.
Embeddings do not contain ready-made answers to every question. Instead, they provide one of the mathematical foundations that make those relationships possible.
5. Common Misconceptions
Embeddings are often misunderstood. They are not databases, maps, dictionaries, or memories stored inside an AI system. They do not contain complete knowledge about a subject, nor do they function as miniature encyclopedias.
Instead, embeddings provide a mathematical representation of concepts that allows similarity to be measured efficiently. They are one component of a much larger system that enables modern language models to process information.
Key Takeaways
Embeddings help solve one of the most important problems in artificial intelligence: representing meaning mathematically.
Rather than storing isolated facts, embeddings preserve learned relationships between concepts. This allows modern AI systems to compare ideas, recognize similarity, and work with natural language in ways that extend far beyond traditional keyword matching.
Understanding embeddings does not explain every aspect of artificial intelligence, but it provides an essential foundation for understanding how modern language models represent information.
Questions for the Curious
- How do AI models create embeddings during training?
- If embeddings represent meaning, how does an AI system generate complete sentences?
- Can two different AI models create different embeddings for the same concept?
- Why do embeddings improve semantic search compared with keyword search?
These questions will be explored in future papers within the AI Reference Library.
About This Paper
AI Reference Papers published by ChicagoRealEstate.ai explain established concepts in artificial intelligence using real-world examples from the real estate industry. Each paper is versioned and updated as the library evolves.
See these concepts applied to real questions in our monthly AI Benchmark Run series, comparing ChatGPT, Claude, and Gemini on the same Chicago real estate questions.