Semantic search in CRE is how you find a clause by what it means instead of what it says. A keyword search for "co-tenancy" misses the provision drafted as "occupancy requirement" or "anchor operating condition," and across a thousand leases those synonyms are the rule, not the exception. Semantic search closes that gap by converting text into vector embeddings, numeric representations of meaning, so that two clauses expressing the same right sit close together in vector space regardless of wording. The result is a portfolio you can ask a question of, rather than a filing cabinet you have to search one drawer at a time. This is the retrieval layer that determines whether a lease portfolio is an asset or an archive.
Key Takeaways
Semantic search finds clauses by meaning, not exact words, so a co-tenancy provision surfaces whether it is called co-tenancy, occupancy requirement, or anchor condition.
The mechanism is vector embeddings: text becomes a high-dimensional vector, and clauses with similar meaning sit close together, measured by cosine similarity between negative one and positive one.
Keyword search is precise but brittle; it fails on synonyms and paraphrase, which is exactly how legal drafting varies across a portfolio.
OpenAI's text-embedding-3 models encode text into vectors of up to 3,072 dimensions, each dimension capturing an abstract feature of meaning, per OpenAI documentation.
Hybrid search, combining keyword and semantic retrieval, outperforms either alone, and is the practical default for portfolio clause search.
What is semantic search in commercial real estate?
Semantic search in commercial real estate is retrieval by meaning rather than by literal string match. It represents each clause as a vector embedding, a list of numbers that encodes what the text means, and finds matches by measuring how close two vectors are. A query for renewal rights returns the extension option even when the word "renewal" never appears.
The distinction from keyword search is the whole point. Keyword search asks "which documents contain this exact string." Semantic search asks "which documents mean roughly this," which is a fundamentally different question and the one a lease analyst has. Legal drafting is adversarial to string matching: the same economic right is written a dozen ways across a dozen counterparties and their counsel. A system that only finds "co-tenancy" will silently miss the "operating covenant" that does the same work, and a silent miss in due diligence is the expensive kind.
Stated plainly: keyword search finds the word, semantic search finds the meaning, and in a lease portfolio the meaning is what you are looking for.
How do vector embeddings make clause search work?
Vector embeddings make clause search work by turning text into geometry. An embedding model maps each clause to a point in high-dimensional space so that clauses with similar meaning land near each other. Similarity is then a distance calculation, typically cosine similarity, which ranges from negative one for opposite meanings to positive one for identical ones.
The mechanics are worth understanding because they explain both the power and the limits. A vector embedding is produced by a model trained so that meaning maps to position. OpenAI's text-embedding-3-large, per OpenAI's documentation, encodes text into a vector of up to 3,072 dimensions, where each dimension captures some abstract feature the model learned. Two clauses about the same right produce vectors pointing in nearly the same direction, so the cosine of the angle between them approaches one. Two unrelated clauses point in different directions, and the cosine drops toward zero.
Here is the mechanism as a worked sequence:
Step | What happens | Concrete example |
Embed the corpus | Every clause becomes a vector, stored in a vector database | 1,000 leases yield tens of thousands of clause vectors |
Embed the query | Your question becomes a vector using the same model | "Where do we have percentage rent?" becomes a vector |
Measure similarity | Cosine similarity ranks every clause against the query | Top matches score near 0.85; unrelated clauses near 0.1 |
Return ranked clauses | Highest-similarity clauses surface first | "Percentage rent," "overage rent," and "breakpoint" all appear |
The vectors live in a vector database, which is built to search millions of high-dimensional vectors quickly. That is what makes "across a thousand leases" a single query instead of a project. The database does not store the meaning of "percentage rent"; it stores the geometry, and geometry is fast to search.
Is semantic search always better than keyword search?
No. Semantic search is better at recall across paraphrase and synonyms, but keyword search is better at exact precision, and each fails where the other succeeds. Keyword search will find a specific defined term or a party name exactly; semantic search will find a concept however it is worded. Serious systems use both.
The evidence here is honest rather than promotional. In a 2025 study comparing lexical and semantic vector methods on medical document classification, the lexical method BM25 achieved higher predictive accuracy than several semantic embedding models on most data splits. Semantic embeddings are not a universal upgrade. They excel at meaning-matching and degrade when you need an exact token, a precise number, or a specific defined term that string matching nails cleanly.
This is why hybrid search is the practical default. As one useful framing: keyword search gives you precision on the terms you can name, and semantic search gives you recall on the terms you cannot. A portfolio clause search wants both, because the analyst both knows exactly what a "gross-up provision" is called and needs to catch every clause that functions like one under a different name.
Dimension | Keyword search | Semantic search |
Matches on | Exact strings and defined terms | Meaning and paraphrase |
Strength | Precise, predictable, cheap | Finds synonyms and rephrasings |
Weakness | Misses synonyms entirely | Can rank a near-miss above an exact match |
Best for | Party names, defined terms, figures | Concept and clause discovery across a portfolio |
What does semantic clause search change for a lease portfolio?
Semantic clause search changes a portfolio from something you file into something you query. Instead of remembering which leases have a problematic co-tenancy trigger, you ask, and the system surfaces every functional match across every counterparty's drafting. Exposure that was buried in wording variation becomes a list.
The operator consequence is portfolio-level visibility. When an anchor tenant threatens to go dark, the question "which of our leases let other tenants reduce rent or terminate if this store closes" is answerable in minutes across the whole book, not weeks of manual review. When a new co-tenancy risk emerges, you can scan for its shape everywhere it might hide. This is the difference between managing exposure you can see and being surprised by exposure you could not find. Firms that make their leases queryable convert a static archive into a live risk instrument; firms that do not keep re-reading the same documents every time a new question arises, and the cost of that re-reading scales with every lease they own.
Frequently Asked Questions
What is the difference between semantic search and keyword search in CRE?
Keyword search matches exact strings, so it finds "co-tenancy" but misses "operating covenant" that means the same thing. Semantic search matches meaning using vector embeddings, so it finds the concept however it is phrased. Keyword search is more precise on defined terms; semantic search has far better recall across the wording variation typical of lease drafting.
Do I need a vector database for semantic clause search?
For a small set of documents, no. For portfolio scale, yes. A vector database stores clause embeddings and searches millions of high-dimensional vectors quickly, which is what makes querying a thousand leases a single fast operation rather than a batch job. It is the infrastructure that makes semantic search practical at scale.
Can semantic search miss a clause that keyword search would find?
Yes. Semantic search can rank a topically similar but wrong clause above an exact match, and it can underperform on precise defined terms or specific figures. This is why hybrid search, combining keyword precision with semantic recall, is the reliable default for clause discovery across a lease portfolio.
Conclusion
A lease portfolio is only an asset if you can ask it questions. Keyword search cannot carry that load alone, because legal drafting expresses the same right in too many words, and every synonym is a place a keyword query goes blind. Semantic search, built on vector embeddings and cosine similarity, closes the gap by matching meaning instead of strings, and a vector database makes it fast enough to run across a thousand leases at once. It is not a replacement for keyword precision; the honest answer is that the strong system uses both. For the leasing analyst and the asset manager, the reframe is simple: the value in a portfolio is not in owning the documents, it is in being able to find the one clause that matters, however it happens to be written.