An embedding model is an AI system that converts text into a fixed-length vector of numbers so that text with similar meaning maps to nearby vectors. In commercial real estate, an embedding model reads a lease clause, rent roll note, or offering memorandum paragraph and outputs the vector that makes meaning-based search possible.
How Does an Embedding Model Work?
An embedding model works by passing text through a trained neural network that outputs a fixed-length list of numbers, one vector per input. The model learns during training to place text with similar meaning at nearby points in a high-dimensional space. The same model must embed both documents and queries so their vectors are comparable.
Vector length is set by the model. OpenAI's text-embedding-3-small outputs 1,536 dimensions and text-embedding-3-large outputs 3,072 dimensions by default, per OpenAI's embeddings documentation. The larger model scored 64.6% on the MTEB benchmark that spans retrieval, classification, and clustering, per OpenAI's launch announcement. Both support Matryoshka truncation, which shortens a vector to as few as 256 dimensions for cheaper storage with a modest accuracy loss.
Component | Function |
Input text | The lease clause, note, or query to encode |
Neural network | Trained weights that map text to a vector |
Output dimensions | Fixed length of the vector, such as 1,536 |
Similarity metric | Cosine similarity, used downstream to compare vectors |
Why an Embedding Model Matters
An embedding model matters because it is the single component that decides retrieval quality in any document AI system. If the model encodes "early termination rights" and "cancellation option" as distant vectors, semantic search fails and the downstream answer degrades no matter how capable the language model is. The embedding model sets the ceiling on retrieval.
Model choice trades cost against precision, and the gap is measurable. OpenAI prices text-embedding-3-small at $0.02 per million tokens and text-embedding-3-large at $0.13 per million tokens, a 6.5x difference, per OpenAI pricing. For a portfolio embedded once, the smaller model is far cheaper; for a corpus where retrieval misses are expensive, the larger model earns its cost. The choice is a deliberate engineering decision, not a default.
Example
An embedding model is chosen by comparing cost and dimensions on a real corpus. A portfolio holds 5,000 leases at roughly 45,000 words each, or about 225 million words total. At 0.75 words per token, that is roughly 300 million tokens to embed once.
Model | Dimensions | Price per 1M tokens | Cost to embed 300M tokens |
text-embedding-3-small | 1,536 | $0.02 | 300 x $0.02 = $6.00 |
text-embedding-3-large | 3,072 | $0.13 | 300 x $0.13 = $39.00 |
Embedding the full portfolio once costs $6.00 with the small model and $39.00 with the large model, a difference of $33.00. Because embedding happens once and queries reuse the stored vectors, the one-time cost is small relative to the retrieval accuracy at stake, which is why teams often test both models on a sample before committing.
Variations and Edge Cases
Embedding models vary by what they encode, how they are trained, and where they run. The variant chosen changes cost, speed, and how well the model handles CRE-specific language.
Variant | Behavior |
General-purpose model | Trained on broad text; strong default for most documents |
Domain-tuned model | Fine-tuned on legal or financial text; better on jargon |
Multilingual model | Encodes multiple languages into one shared space |
Self-hosted model | Runs on private infrastructure; no data leaves the system |
Reduced-dimension output | Truncated vector, such as 256 dims; cheaper, slightly less precise |
Embedding Model vs Vector Embedding
An embedding model is often confused with a vector embedding, but one produces the other. An embedding model is the AI system, the trained network that does the encoding. A vector embedding is the output, the specific list of numbers the model returns for one piece of text. The model is the machine; the embedding is the product.
Every vector embedding comes from an embedding model, and swapping the model changes every embedding. A vector embedded by text-embedding-3-small cannot be compared against one from text-embedding-3-large, because the two models place text in different spaces. This is why a document corpus and its search queries must be embedded by the same model version.
Frequently Asked Questions
What is an embedding model in document AI?An embedding model is an AI system that converts text into a fixed-length vector of numbers so that text with similar meaning maps to nearby vectors. In commercial real estate it reads a lease clause or rent roll note and outputs the vector that makes meaning-based search possible.
What is the difference between an embedding model and a vector embedding?An embedding model is the AI system that does the encoding, while a vector embedding is the list of numbers it outputs for one piece of text. Every embedding comes from a model, and changing the model changes every embedding it produces.
How many dimensions does an embedding model output?Output length is set by the model. OpenAI's text-embedding-3-small outputs 1,536 dimensions and text-embedding-3-large outputs 3,072 dimensions by default, per OpenAI documentation, with Matryoshka truncation down to 256 for cheaper storage.
Related Terms
Vector Embedding
Semantic Search
Retrieval-Augmented Generation
Large Language Model
Vector Database