AI Vocabulary
Embedding: how does a text become a point in space?
An embedding turns a text into a series of numbers representing its meaning. Two texts of similar meaning produce similar series, even with no words in common. This mechanism is what lets a system retrieve a relevant passage without your having used the right keywords.
The idea in one image
Imagine each text receiving coordinates, like a point on a map. Texts about the same thing end up as neighbours; unrelated texts end up far apart.
The map has far more than two dimensions — several hundred, usually — but the principle holds: searching becomes measuring a distance.
The word has no settled equivalent outside technical literature, where one occasionally finds vector representation. In practice, the English term is used everywhere.
What it changes compared with a classic search
A keyword search looks for occurrences: type termination and it finds documents containing that word, missing those that speak of early cancellation or of the end of the agreement.
An embedding search looks for meaning. It retrieves all three formulations, because they occupy neighbouring positions on the map.
That is what makes it possible to interrogate a document corpus in ordinary language, without knowing the exact vocabulary of whoever drafted it.
A legal example
You are looking, across a corpus of contracts, for those limiting the supplier's liability. The drafting varies: a cap expressed in months of fees, in a percentage of price, by excluding certain heads of loss, or by reference to a figure in a schedule.
A keyword search obliges you to imagine all those formulations and look for each in turn. An embedding search brings them together on its own, because they occupy a common area of the map.
That is the case where the mechanism shows its full effect: one legal object expressed in four different ways.
What it does not solve
Proximity of meaning is not legal relevance. Two non-compete clauses sit very close on the map, including when one is valid and the other void; the distinction you care about does not show there.
Negation is represented poorly. The tenant shall and the tenant shall not occupy neighbouring positions while saying opposite things. That is a known limitation with direct contractual consequences.
And chunking weighs heavily. To be turned into points, documents are fragmented; a clause whose scope depends on a definition 40 pages away will be represented without it.
One further technical point has practical consequences: embeddings are computed by a particular model, distinct from the one that drafts answers. Changing it requires recomputing across the entire corpus, a heavy operation rarely anticipated at the time of the initial choice.
Why it matters to a lawyer
Because this is the mechanism that decides what the system sees before answering, and its limitations explain a good share of disappointing results. When an answer misses the point, the cause lies here more often than in the model.
The useful question therefore concerns what happens when retrieval fails: does the system signal that it found nothing relevant, or answer anyway with the least bad thing it has?