What grounded AI means for legal document research

In shortGrounded AI is an approach where the model answers only from a specific set of documents you provide, and every factual claim is tied to the passage it came from. For legal work this means an answer you can check against the page, and a refusal instead of a guess when the documents do not contain the answer.

Grounded AI means the model is only allowed to answer from documents you supply, and every factual claim in the answer is tied to the passage it came from. For legal document research that changes two things at once: you can check any answer against the cited page in seconds, and when your documents do not contain the answer the tool says so instead of inventing one.

Grounded AI is a design pattern, sometimes called retrieval-augmented generation, in which a search step runs before the model writes anything. The search finds the passages in your own files most relevant to the question; the model is then instructed to draft an answer using only those passages and to cite each one. Nothing from the model's general training is supposed to enter the answer.

The problem this solves is familiar to anyone who has tried a general chatbot on real legal work. The answers read well. Some of them are wrong in ways that are hard to spot, and none of them tell you where a sentence came from. An associate can spend longer verifying a chatbot's answer than it would have taken to find the clause by hand, which is the opposite of what the tool promised.

How does a grounded system actually find the right passage?

It converts text into numbers that capture meaning, then compares those numbers. When a document is uploaded, it is split into passages of a few hundred words each, and each passage is turned into an embedding, a long list of numbers produced by a model built for that purpose. Your question is turned into an embedding the same way, and the system retrieves the passages whose numbers sit closest to the question's.

This is why a grounded tool can find a clause that says written notice of sixty days when you asked about the notice period, even though the words do not match. It is also why the quality of the embedding model matters for legal text, which uses vocabulary and structure that general-purpose models handle less well.

What happens after retrieval?

The retrieved passages, and only those passages, are sent to a language model together with the question and an instruction to answer strictly from the supplied text. The model drafts the answer and attaches a citation for each claim, naming the document and the page the passage came from.

Here is the sequence in a Vorticel-style system, step by step:

  1. The firm uploads a document; it is encrypted per firm before it is stored, and its text is extracted.
  2. The text is split into passages, and each passage is embedded so it can be searched by meaning.
  3. A user asks a question in plain English; the question is embedded and the closest passages from that firm's own library are retrieved.
  4. Those passages go to the language model with instructions to answer only from them and cite each one.
  5. If the passages do not support an answer, the tool returns a refusal rather than calling on outside knowledge.

Notice what does not happen. The whole document library is not sent to the model. The model is not asked what it knows about the topic. The original file never leaves encrypted storage; only extracted text does, and only the retrieved passages reach the model.

Why does the citation change the workflow, not just the answer?

Because verification becomes a click instead of a search. A grounded answer that says the agreement requires sixty days of written notice, citing a named PDF at page seven, can be confirmed by opening page seven. That takes seconds. An uncited answer to the same question forces you to re-find the clause yourself, at which point the tool saved nothing.

The citation also makes errors visible. Models can summarize a passage imperfectly; with a citation you catch the mismatch immediately, and with a source you can quote the actual language in your memo instead of the model's paraphrase.

Grounded tool or general chatbot: which fits which job?

QuestionGrounded document toolGeneral chatbot
Where does the answer come from?Passages retrieved from your own uploaded filesThe model's training data plus whatever you paste
Can you check it?Yes, each claim cites a document and pageUsually not; sources are absent or invented
What happens when the answer is not available?It refuses and says soIt produces a plausible answer anyway
What leaves your firm?Extracted text for embedding; retrieved passages per questionWhatever you paste, under the chatbot's consumer terms
Best suited toQuestions about your own contracts, discovery, correspondenceDrafting, brainstorming, general explanation

Both have a place. The mistake is using a general chatbot for questions whose answers live in privileged documents, because that is where the combination of guessing and unclear data handling does the most damage. A useful rule for a small firm: if the answer should come from a file the firm holds, use the grounded tool; if the task is drafting or explaining a concept, a general assistant is fine as long as no client text is pasted into it.

What should a firm look for when evaluating a grounded tool?

Three things, in order. First, does every answer cite a specific document and page, and can you open the passage from the citation? Second, when you ask something the documents cannot answer, does the tool refuse? Test this on day one with a question about a case you know is not in the files. Third, is the retrieval scoped strictly to your firm, with no shared index across customers? The how-it-works page describes how Vorticel handles each of these, and the same questions apply to any vendor.

This is also the point where doing it by hand stops scaling. A paralegal can find one clause in one agreement quickly. Finding every notice provision across forty agreements, with page references, is an afternoon. A grounded tool built for this does the retrieval and the citing, and leaves the judgment to you.

The honest limitation

A grounded tool is only as good as the documents you give it. If a retainer is not uploaded, the tool cannot cite it, and it will tell you so. Some firms hear that as a weakness. It is the feature: the alternative is a tool that fills the gap from memory and hands you a citation to nowhere. The other real limitation is scanned, image-only PDFs, which need a text layer before extraction can work; check whether your files have one.

Try it on one real question

Pick an agreement your firm already knows well, start a free trial, upload it, and ask the question you would normally answer by scrolling. Then ask a second question you know the document does not address. The first answer should arrive with a page citation; the second should be a refusal. If both happen, the tool is grounded. The trial runs 14 days and needs no card.

Frequently asked questions

What is the difference between grounded AI and a general chatbot?

A general chatbot answers from everything it learned during training, which is why it can produce a plausible case name that does not exist. A grounded system first retrieves passages from documents you supplied and then instructs the model to answer only from those passages, citing each one. If nothing relevant is retrieved, it refuses.

What is an embedding, in plain English?

An embedding is a list of numbers that represents the meaning of a passage of text. Passages about the same idea end up with similar numbers, so a system can find the passages closest in meaning to your question without matching exact words. This is what lets a search for notice period find a clause that says written notice of sixty days.

Does a citation to document and page guarantee the answer is correct?

No. A citation guarantees you can check the answer quickly, which is the point. The model can still summarize a passage imperfectly, so the citation is what turns a claim into something verifiable in seconds rather than something you must take on faith or re-research from scratch.

Why is refusing to answer a feature rather than a failure?

In legal work a confident wrong answer costs more than no answer. A grounded tool that says it could not find the answer in your documents tells you something useful: either the document is not uploaded, or the answer genuinely is not in your files. A tool that guesses hides both possibilities.

Can grounded AI search the internet for case law?

A grounded document tool is deliberately limited to the files your firm has uploaded. It is not a case-law database and does not browse the web. That limit is what makes its answers checkable and keeps client material from mixing with outside content. Use it for your own documents, and a research database for authority.

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.