How RAG works for real estate — and why it kills hallucination

The single most common objection to AI in real estate is "but the AI will make things up." It's a real concern. Language models, asked a question they don't know the answer to, will produce a confident, plausible-sounding answer that is wrong. In a brokerage context, the wrong answer might be an HOA fee, a square footage, a school zone, a commission split. The cost of being confidently wrong is high.

The technical answer to this problem is called Retrieval-Augmented Generation, or RAG. It's the architectural pattern that lets an AI agent answer questions from a brokerage's actual data instead of from the patterns it learned during training. RAG is what separates a system that hallucinates from one that doesn't.

What hallucination actually is

A language model is, in technical terms, a next-token predictor. Given everything it's read so far in a conversation, it predicts the most likely next word. It does not know anything. It does not have a database. It has parameters — billions of numerical weights that encode patterns from its training data.

When you ask such a model "what is the HOA fee at 4221 Anderson Avenue in Coral Gables," the model has not been trained on that specific listing. It will generate text that sounds like an HOA fee — "$1,247 per month" — because text describing HOA fees in Coral Gables condos follows a recognizable pattern. The number is fiction. The model has no way of knowing.

This is hallucination. It's not a bug; it's the basic operating mode of a model with no access to your data.

What RAG changes

RAG inserts a retrieval step before generation.

1. The system receives the user's question.
2. A retrieval component searches the brokerage's actual knowledge base (listings, MLS data, past conversations, knowledge documents) for the most relevant chunks.
3. Those chunks are inserted into the prompt as context.
4. The model is told: "Answer this question using only the following information. If the information isn't here, say you don't know."
5. The model generates an answer grounded in real data.

The model is no longer guessing. It's reading off a card it just pulled from a real index. If the answer isn't on the card, the system is configured to refuse rather than invent.

The components of a real RAG system

A working RAG system for a brokerage has four components.

The corpus. Everything the agent might need to answer from: active listings with full attributes (HOA, taxes, comps, school zones), past lead conversations, knowledge documents (the brokerage's onboarding doc, neighborhood guides, FAQ on financing). The corpus is curated. Bad data in, bad answers out.

The chunker. Listings, documents, and conversations are split into chunks small enough to fit in a model's context window. A 12-page neighborhood guide becomes 30 chunks of ~500 words each.

The embedder. Each chunk is converted to a high-dimensional vector — a numerical representation that captures meaning. Two chunks about HOA fees end up near each other in the vector space even if they use different words. The vectors are stored in a vector database (pgvector, Pinecone, Weaviate, etc.).

The retriever. When a question comes in, it's embedded the same way. The retriever finds the chunks whose vectors are closest. Top-k (typically 5–20) chunks are returned. Modern systems combine vector search with keyword search (BM25/FTS) and a reranking step that re-orders the results by relevance to the specific question.

The combined result — vector + keyword + rerank — is called hybrid search. It's the standard in production RAG systems because each technique catches what the others miss.

Why this works in real estate specifically

Three reasons RAG is a particularly good fit for brokerage workflows.

The answer set is bounded. A lead asks about a property in your inventory, the answer is in your inventory data. Asks about a neighborhood, the answer is in your neighborhood guide. Asks about a financing question, the answer is in your knowledge base. Almost every legitimate lead question can be answered from a bounded, brokerage-controlled corpus.

The cost of being wrong is high. Invented HOA fees, invented commissions, invented school zones are not minor errors — they're material misrepresentations. RAG with a refusal threshold means the system says "I'm not sure, let me check with the agent" instead of making something up.

The corpus is yours to control. Unlike a public web search, the brokerage decides what's in the knowledge base. Outdated listings are removed. Approved disclosures are versioned in. The compliance team can audit exactly what the AI has been told it's allowed to say.

A worked retrieval

A lead asks Sara on WhatsApp: "What are the HOA and school zones at 4221 Anderson?"

The system does this in sequence, in under 800ms:

1. Embeds the question into a 1536-dimensional vector.
2. Searches the brokerage's listing vector store. Finds the chunk for 4221 Anderson (top match, similarity 0.92), plus two adjacent listings on the same block.
3. Searches the school-zone knowledge base. Finds the Coral Gables school-zoning chunk that covers Anderson Avenue.
4. Reranks. The 4221 Anderson chunk and the Coral Gables school-zone chunk are the top two.
5. Constructs a prompt: "Using only the following information, answer the user's question. If you can't answer fully from this, say so. [Chunk 1: 4221 Anderson, HOA $1,247/mo, 2BR/2.5BA, Coral Gables Elementary zone] [Chunk 2: Coral Gables school-zone guide]."
6. The model generates: "The HOA at 4221 Anderson is $1,247/month. It's zoned for Coral Gables Elementary, Ponce de Leon Middle, and Coral Gables Senior High."

The model didn't know any of those facts before the retrieval ran. They were placed in front of it. If the listing had been removed from inventory, the retrieval would have returned nothing, and the system would have responded: "Let me check on that property — I don't have current information."

Where RAG can still go wrong

RAG isn't magic. The failure modes are different from raw hallucination but worth knowing.

Stale data. If the corpus isn't kept in sync with the source of truth, RAG will confidently retrieve outdated HOA fees, sold listings, removed price drops. Sync is operational discipline, not a model property.

Bad chunking. If a critical sentence gets split across two chunks, the retriever might miss it. Good chunking (semantic-aware splits, overlap, metadata-rich) avoids this.

Retrieval miss. Sometimes the right chunk doesn't show up in the top-k. Hybrid search (vector + keyword + rerank) addresses most of this, but it's not perfect. The refusal threshold catches the residual.

Prompt drift. The instruction "answer only from the retrieved context" sometimes gets overridden if the model has strong prior beliefs. Stronger system prompts and refusal-prone tuning reduce this.

A serious RAG system measures all four of these in production and tracks them as quality KPIs.

What this means for a brokerage owner

The two things to ask any AI vendor:

1. "Show me what the agent will retrieve when I ask this question." A real RAG system can produce an audit view that shows: question, retrieved chunks, sources, final answer. If the vendor can't show this, the retrieval probably isn't really happening — the model is guessing.

2. "What does the agent do when it doesn't know?" The right answer is: refuses and asks the lead to hold while a human checks. The wrong answer is: confidently produces a number. If the agent never refuses anything in a demo, the refusal threshold isn't tuned — the agent is hallucinating, just convincingly.

A well-built RAG system, in real estate, makes the AI as accurate as your data. The data is the constraint. The model is the easy part.

---

Closi's RAG implementation is the knowledge base layer behind Sara, with hybrid search, semantic chunking, and a configurable refusal threshold. The training module is where you load and tune the corpus for your brokerage.