Retrieval-based AI for legal teams, without the jargon

Most worry about AI in a law firm comes down to one fear: the machine making things up. A confident answer that points to a case that does not exist, or a rule that was repealed two years ago, is worse than no answer at all. The good news is that there is a simple, practical way to build an assistant that does not do this. It is called retrieval, and once you understand the idea, the rest stops feeling like magic.

The simple idea behind it

Picture two ways to ask a question.

In the first, you walk up to a very well-read colleague and ask them to answer from memory. They have read an enormous amount, but they read it a while ago, they cannot show you exactly where each fact came from, and on a bad day they will fill a gap with something that sounds right. That is what a plain language model does when you ask it a question cold.

In the second, you ask the same colleague to first go to the shelf, pull the exact statutes and files that relate to your question, read those pages, and then answer using only what is in front of them, telling you which page each point came from. That second method is retrieval. You find the relevant documents first, then ask the model to answer from those documents and nothing else.

The technical name for this is retrieval-augmented generation, sometimes shortened to RAG. You can forget the term right after reading it. The plain version is what matters: look it up first, then answer from what you found.

Why this beats answering from memory

Two things change the moment you put the documents in front of the model.

First, the made-up answers drop sharply. When the assistant is told to work only from a set of supplied pages, it has far less room to invent. If the answer is not in the documents, a well-built system says so instead of guessing. That honesty is exactly what you want from anything touching legal work.

Second, the answers stay current. A model trained last year does not know about a rule that changed last month. But if the retrieval step pulls from an up-to-date collection of the law, the assistant reads the new version and answers accordingly. You update the shelf, not the model. That keeps the work honest about what the law says today, not what it said whenever the model was last trained.

Citations are the whole point

An answer you cannot check is just a rumour with good grammar. The reason citations matter so much here is that they turn the assistant from something you have to trust into something you can verify.

When every claim points back to a specific provision, judgment, or page in your own file, you can do what lawyers already do best: read the source and decide for yourself. The assistant becomes a fast way to find the right passage, not a replacement for your judgment. If a citation does not support the point, you see it immediately and move on. That is a far safer footing than a tidy paragraph with nothing behind it.

A practical way to think about it:

• An answer with no source is a starting point at best, and a liability at worst.
• An answer with a source is a shortcut to the page you needed to read anyway.

Keeping client facts private while you do it

There is a catch worth naming. To answer questions about a live matter, the assistant needs to read the matter. That is sensitive material, and it should never be the price of a useful tool.

The way to handle this is to separate two jobs. Reading the public law is one job. Reading your client's private file is another. They do not need to share the same identities. Before any matter is processed, you can mask the client facts, swapping out names and identifying details so the work happens on a stripped-down version that still makes legal sense but no longer points back to a real person or company.

This is where the Adimen tools fit together. Adimen Corpus (coming soon) applies retrieval over the real law, so questions about statutes and judgments are answered from current sources with citations. Adimen Praxis (coming soon) applies the same approach to your own client files, and connects to Corpus through Velum, which is available today. Velum is the layer that masks client identities before any matter is processed, so the private facts stay on your side while the legal reasoning still gets done. The retrieval method and the privacy method are designed to work as one.

A working day with it

Here is what the difference looks like in practice.

A question lands on a junior's desk about whether a particular notice period applies to a contract. Instead of an afternoon of searching, they ask the assistant. It retrieves the relevant provisions, answers in plain terms, and shows the exact sections it relied on. The junior opens those sections, confirms they say what the assistant claims, and drafts the note. The senior reviewing it can follow the same citations in seconds rather than retracing the research from scratch.

The client facts in that matter never travelled around unprotected, because they were masked before processing. The legal sources were current, because the underlying collection is kept up to date rather than frozen in a model. And nobody had to take a single sentence on faith, because every point came with a place to check it.

That is the quiet promise of retrieval. Not a machine that knows everything, but an assistant that shows its work, stays current, and respects the line around your clients. It does the looking-up fast so you can do the judging well.

If you want to see where this fits in real practice, read the use cases and learn how Velum keeps client identities masked before any matter is processed.