Retrieval-Augmented Generation (RAG)
How AI Combines Learned Knowledge with Current Information
Version 1.7
Joe Bjorklund & Jobe Rose
Baird and Warner
Technical Accuracy Statement
This paper explains Retrieval-Augmented Generation (RAG), an architectural approach used in many modern AI systems to improve the quality and freshness of generated responses. While individual companies implement retrieval differently, the concepts presented here are broadly accepted within the fields of information retrieval and machine learning.
The discussion focuses on the underlying ideas rather than the proprietary implementations of any single AI platform.
Abstract
One of the first questions people ask about artificial intelligence is surprisingly practical.
“If AI was trained months ago, how does it know what happened yesterday?”
It is a reasonable question because it exposes one of the limitations of a standalone large language model. A language model can only generate responses from the knowledge available during training and the information supplied in the current conversation. It does not automatically know today’s home listings, this morning’s interest rates, or a newly published zoning ordinance.
Modern AI systems often overcome this limitation through a technique known as Retrieval-Augmented Generation, commonly abbreviated as RAG.
Rather than expecting the language model to remember everything, a retrieval system first searches for relevant information. The retrieved material is then supplied to the language model, allowing it to generate an answer that combines learned language abilities with current, external information.
This paper explains why retrieval has become one of the most important architectural ideas in modern AI and why it has particular relevance for industries such as real estate, where information changes continuously.
1. The Knowledge Problem
Imagine you ask an AI assistant:
“How many homes are currently listed for sale in Lincoln Park above $2 million?”
A standalone language model cannot reliably answer that question because the number changes constantly.
Even if the model had seen similar information during training, it would have no reliable way to know what is true today.
This illustrates an important distinction.
A language model excels at understanding language, recognizing patterns, and generating coherent responses. It is not designed to function as a continuously updated database.
Current information requires a different approach.
2. Looking Up Before Answering
Consider how a real estate broker answers a detailed client question.
Suppose a buyer asks:
“How many luxury condominiums are available in Streeterville this morning?”
A professional broker rarely answers entirely from memory.
Instead, the broker opens the MLS, reviews the current listings, and then explains the results.
The broker combines experience with fresh information.
Retrieval-Augmented Generation follows a remarkably similar pattern.
Before generating a response, the AI system retrieves relevant information from one or more external sources. That information may come from a database, a document collection, a website, an MLS feed, or another trusted repository. Only after reviewing that material does the language model begin generating its answer.
The language model is not replacing the database.
It is using the database.
3. Why Retrieval Matters
Without retrieval, every answer is limited to what the model learned during training and the information provided in the current conversation.
With retrieval, the model can incorporate information that did not exist when it was trained.
That difference fundamentally changes what AI systems can accomplish.
Rather than asking a language model to memorize an entire world that changes every day, retrieval allows it to consult reliable sources before responding.
In many applications, this produces answers that are both more current and easier to verify.
4. RAG Is Not Search
One of the most common misconceptions is that Retrieval-Augmented Generation is simply another name for search.
It is not.
Traditional search systems usually return a list of documents, leaving the user to read and interpret the results.
A RAG system uses retrieval as the first step, but its goal is different. Instead of presenting documents alone, it provides those documents to the language model, which then synthesizes the information into a coherent response.
Search helps people find information.
Retrieval-Augmented Generation helps AI systems use information.
Although they are related, they are not the same.
5. Why This Matters for Real Estate
Real estate is one of the clearest examples of why retrieval matters.
Listings appear and disappear every day.
Prices change.
Properties go under contract.
Open houses are scheduled and canceled.
Mortgage rates fluctuate.
Neighborhood statistics evolve over time.
No language model can permanently memorize information that changes continuously.
Instead, successful AI-powered real estate systems increasingly combine two complementary capabilities.
The language model contributes its ability to understand language, explain concepts, summarize information, and generate natural responses.
The retrieval system contributes accurate, current, and verifiable information.
Together, they produce answers that are significantly more useful than either component could produce alone.
6. A Practical Example
Imagine a buyer asks:
“Show me the newest luxury homes in Lincoln Park with rooftop decks and attached garages.”
A language model alone cannot know which homes were listed this morning. A retrieval system searches the available listing data for properties matching those criteria.
The results are then supplied to the language model, which generates a natural-language response describing the available homes and highlighting relevant details.
Neither component accomplishes the task alone. The retrieval system finds the information. The language model explains it.
Common Misconceptions
Retrieval-Augmented Generation does not replace a language model.
It does not eliminate hallucinations entirely.
It is not simply another search engine.
It is not a database.
Instead, it is an architectural approach that combines external information retrieval with the language-generation capabilities of a large language model.
Key Takeaways
Large language models excel at understanding and generating language but are not designed to function as continuously updated databases.
Retrieval-Augmented Generation addresses this limitation by allowing an AI system to retrieve relevant external information before generating a response.
This combination of learned language capabilities and current information has become one of the defining architectural patterns in modern AI and is particularly valuable in industries such as real estate, where information changes constantly.
Questions for the Curious
- How does a retrieval system decide which documents are relevant?
- What role do embeddings play in retrieval?
- What is a vector database?
- How does AI rank competing sources of information?
- Why can two RAG systems produce different answers from the same documents?
These questions lead naturally to future papers in the AI Reference Library.
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.