How Vector Embeddings Power Search in Legal AI Tools

In shortA vector embedding is a list of numbers a machine-learning model assigns to a piece of text, so that passages with similar meaning get similar numbers. Legal AI tools use embeddings to find the passage that answers a question even when the wording differs from the source document, which is why they outperform keyword search on contracts and discovery files.

Ask a keyword-only document search tool "when does the indemnification clause expire" and it hands back every page that contains the word indemnification, dozens of hits, none of them necessarily the answer. That is the real cost of keyword search on legal documents: time spent scanning passages that mention a term without answering the question asked. Vector embeddings, the technology behind grounded AI tools like Vorticel, solve a narrower and more useful problem. A vector embedding is a list of numbers that a machine-learning model assigns to a piece of text based on its meaning, so that passages saying similar things end up with similar numbers even when the wording is different. This article walks through how that works, in plain English, and what actually happens to a firm's documents along the way.

What Is a Vector Embedding?

A vector embedding is a list of numbers, typically hundreds or thousands of them, produced by a model trained on huge amounts of text to place similar meanings near each other in that number space. Think of it as assigning every sentence a set of coordinates: sentences about contract termination cluster together in that space, sentences about billing disputes cluster somewhere else, and the model never has to be told those clusters exist — it learns them from patterns in the text it was trained on.

The word choice does not have to match for two passages to land near each other. "Termination for cause" and "grounds for dismissal" can produce embeddings that sit close together because they mean similar things, even though they share almost no words. That is the entire advantage over keyword matching: embeddings compare meaning, not spelling.

The comparison itself is arithmetic, not judgment. Once a question and a passage each have a numeric embedding, software measures the distance between the two lists of numbers — the smaller the distance, the closer the meanings. A search over a thousand-page document set becomes a search over a thousand short distance calculations, which is fast enough to run in the time it takes to type a question.

How Does a Document Become Searchable in a Tool Like Vorticel?

Turning an uploaded file into something a question can search happens in a fixed sequence, and every step is necessary for retrieval to work:

  1. The document is uploaded over an encrypted connection and tagged to the firm's account.
  2. It is encrypted at rest with per-firm libsodium encryption before it touches disk.
  3. Text is extracted from the file and split into smaller passages.
  4. Those passages, not the original file, are sent to Voyage AI to be converted into embeddings.
  5. The resulting embeddings are stored and indexed so a future question can be compared against them.

The original file itself never leaves encrypted storage at any point in that sequence. What leaves, once, is extracted text sent specifically to be embedded — see Vorticel's security and confidentiality page for the full breakdown of who touches what.

Why Does Semantic Search Find Passages That Keyword Search Misses?

A keyword search engine matches literal strings: search "fire an employee" and it will not surface a clause titled "termination for cause" unless that exact phrase happens to appear nearby too. An embedding-based search compares the meaning of the question to the meaning of every passage and ranks by closeness in that number space, so wording differences stop being the deciding factor.

Search typeWhat it matchesWhere it struggles
Keyword searchExact words or close variants present in the textSynonyms, paraphrasing, and legal terms of art that describe the same idea differently
Embedding (semantic) searchMeaning of the question against the meaning of each passageRequires a trained embedding model and a vector index, more infrastructure than a text search box

Neither approach is magic. Embedding search still depends on the embedding model being good at legal language specifically, and on chunking documents into passages that keep related sentences together rather than splitting a clause in half.

What Happens After the Matching Passages Are Found?

Finding relevant passages is only half the job; someone still has to turn a handful of retrieved passages into a usable answer. In a grounded AI tool, the passages that scored closest to the question's embedding are handed to a language model with instructions to answer only from what it was given. Vorticel uses Anthropic's Claude for this step: it drafts an answer strictly from the retrieved passages, cites the source document and page for each claim, and says so honestly when none of the retrieved passages actually answer the question, rather than filling the gap with general knowledge. For a fuller walkthrough of that distinction, see what grounded AI means for legal document research.

This is the mechanism that separates a grounded search tool from a general chatbot: a general chatbot answers from whatever it learned during training, while a retrieval step like this restricts the answer to a specific firm's own uploaded files. The Vorticel documentation covers the retrieval and citation behavior in more technical detail, including how citations map back to the source page.

Is Sending Document Text to an Embedding Model a Security Risk?

This is the objection worth naming directly rather than glossing over: a firm handling privileged material should be uneasy about any step where text leaves its control. The honest answer is that the extracted text sent to Voyage AI to be embedded is the one point where document content leaves Vorticel's own infrastructure, and it happens only for that narrow purpose. The original file stays encrypted at rest the entire time; only Anthropic, and only for a single question, ever sees the small set of passages retrieved to answer it; and neither vendor uses that text to train its models. Firms comparing Vorticel to a general chatbot or to keeping everything in a plain document management system should weigh that specific, minimal exposure against the alternative of a human spending hours manually reviewing files that a keyword search under-serves. See grounded AI vs. general chatbots for legal research for that fuller comparison, and the FAQ page for shorter answers to the questions that come up most.

How Should a Firm Evaluate This Before Adopting It?

Ask any vendor using embeddings three questions before uploading real client material: which embedding provider receives your text, whether the original file or only extracted text is sent, and whether that provider trains on what it receives. A vendor that cannot answer all three plainly has not thought through its own data flow closely enough to be trusted with privileged documents.

For Vorticel, the answers are Voyage AI, extracted text only, and no training on customer data, on a 14-day free trial that does not require a card, with plans starting at $49 a month on the pricing page. IT and compliance reviewers evaluating any embedding-based tool should expect a vendor to answer these same three questions in writing, not just in a sales call.

The fastest way to see whether embedding-based search actually finds what keyword search misses in your own documents is to run it against a real file rather than take a vendor's word for it. Start a free trial, upload one contract your team has struggled to search by keyword, and ask it the question in plain language — you will see the retrieved passage and its citation, not just a list of pages that happen to contain a word.

Frequently asked questions

What is a vector embedding in plain English?

A vector embedding is a list of numbers that a machine-learning model assigns to a piece of text based on its meaning. Passages that mean similar things get numbers that sit close together, while passages about unrelated topics get numbers that sit far apart. Software can then compare a question's numbers to a document's numbers to find the closest match, without either one containing the exact same words.

How is embedding search different from keyword search?

Keyword search matches literal words and misses a passage that answers the question using different wording, like a clause labeled termination for cause when you searched for firing an employee. Embedding search compares meaning instead of words, so it surfaces the right passage even when the phrasing does not overlap. Keyword search is faster to build; embedding search finds more of what you actually meant.

Does Vorticel use my documents to train its AI models?

No. Vorticel does not use your documents, questions, or answers to train any model. Document text is sent to Voyage AI only to generate the embeddings that make search possible, and only the passages retrieved for a specific question are sent to Anthropic to draft that answer. Neither vendor retains your text to train on.

What happens to the original document file during this process?

The original file is encrypted with per-firm libsodium encryption before it is written to disk and stays that way until your firm requests it back. It never leaves encrypted storage. Only the extracted text, split into passages, is sent out to be embedded; the file itself is not what leaves Vorticel's infrastructure at any point.

Can embeddings guarantee a correct answer every time?

No, and any tool that claims otherwise is overselling the mechanism. Embeddings find the passages most likely to be relevant to a question; a language model then drafts an answer from only those passages and cites the document and page it came from. When no retrieved passage actually answers the question, a well-built system says so instead of guessing.

Which AI models does Vorticel use for embeddings and answers?

Vorticel uses Voyage AI, a MongoDB company, to generate the embeddings that power search, and Anthropic's Claude to draft answers from the passages retrieved for a given question. Voyage AI receives extracted document text to embed; Anthropic receives only the specific passages retrieved to answer one question, never the full document.

Try grounded, cited answers on your own documents.

Create your firm's private workspace, upload a few files, and ask the first question in minutes. 14-day free trial, no card required.