A teaching thread

RAG for real estate companies

How a property firm asks its own leases, title files, and estoppels, and gets the clause back, not a guess.

  1. 01

    The problem

    Real estate firms don't lack documents. They can't find the right sentence in time.

    Leases, title, HOA rules, inspections, zoning memos. The answer is usually already in the file. RAG is how you ask the archive and get it back with a source.

  2. 02

    The idea

    RAG means Retrieval-Augmented Generation.

    Before the model writes, it searches your documents, pulls the passages that match, and answers from those, not from memory.

    The model is the writer. Your files are the source of truth.

  3. 03

    The limit

    Why not paste the whole lease into ChatGPT?

    A mall lease runs 80–200 pages. A portfolio is thousands. Models still lose the middle. A general model has never read your rent roll or last Tuesday's estoppel.

    Retrieval picks the pages. Then the model reads only those.

  4. 04

    The corpus

    What a property company should actually index.

    • Full leases and amendments
    • Title exceptions
    • HOA and condo declarations
    • Inspection and PCA reports
    • Zoning letters and surveys
    • Rent rolls and estoppels

    If it is not in the index, the model should say so.

  5. 05

    The pipeline

    Four steps. Then an answer.

    1. 1Ingest. PDFs, email, and scans become text.
    2. 2Chunk by section. Article 7 stays with Article 7.
    3. 3Embed. “security deposit” can match “tenant's deposit.”
    4. 4Retrieve. Pull the top passages and hand them over with the question.

  6. 06

    The guardrail

    Generation is the easy demo and the easy place to get hurt.

    Tell the model: answer only from these passages. Quote the clause. Name the document and page. If two clauses conflict, say so. If the file is silent, refuse.

    A fluent answer with no citation is a liability.

  7. 07

    The work

    Four workflows that earn their keep.

    1. 1Lease Q&A. “Who pays to replace the roof?”
    2. 2Diligence. Data-room questions while the clock is running.
    3. 3Broker copilot. Answers that cite the listing file.
    4. 4Resident bot. House rules only, and an escalation when it should not guess.

  8. 08

    Permissions

    The hard problem is not search. It is who may see what.

    A resident must not pull another unit's lease. A junior broker must not see seller financials on a listing they do not cover.

    One index with hard filters, or separate indexes. Without that, RAG leaks.

  9. 09

    Failure modes

    Property files break careless systems.

    • NumbersRent, CAM, dates. Extract them. Never round.
    • Stale paperRank the 2024 amendment over the 2019 lease.
    • Bad OCRA scan that never became text was never searchable.
    • BoilerplateSimilar clauses hide the one line that differs.

  10. 10

    Proof

    Do not trust a demo. Build a test set.

    Take 50 questions asset managers already ask. Record the right document, page, and answer.

    Score three things: right page, faithful answer, and a refusal when the file is silent.

    Ship when retrieval is right. Polish the prose after.

  11. 11

    Where to start

    Start with one building, not the portfolio.

    Index that asset's lease file. Chunk by article. Ask the questions people already send the property manager. Compare each citation to the clause a senior analyst would open.

    If it cannot beat Ctrl+F plus that analyst, it is not ready.

  12. 12

    The point

    The firms that get this right will not have the biggest model.

    They will keep leases and estoppels clean, current, and permissioned, so a 9am question returns the clause and the page.

    RAG is a filing system that can talk. Care for the files as much as the model.