A parent walks into a family law matter carrying seven years of emails and documents scattered everywhere.
Not a file. Not a bundle a solicitor has already worked through.
Years of messages written across the worst stretch of their life, some of it argument, some of it logistics about a Tuesday pickup, and somewhere inside it the handful of exchanges that decide how a court sees them.
They cannot afford someone to read all of it. They could not hold it in their head if they tried.
And the thing that matters most is usually not the angriest email; it is a flat sentence written long before, which only means something once you know what came after.
That is the problem RYTZ exists to solve, and it is worth being precise about why the obvious tools do not.
Search by keyword and you get the messages containing the words you guessed.
Search by meaning and you finally find the ones that say the same thing differently, a real advance, and still not enough.
Neither can tell you that this message is the reason that one exists. Neither builds a chronology. Both hand you fragments and leave the work where it started.
Getting past that took us somewhere unusual, and this is a note about where.
RYTZ holds two knowledge graphs rather than one.
The first is the law: the legislation corpus, its sections and the relationships between them.
The second belongs to a single client, their evidence, their correspondence, the chronology of their matter.
A retrieval bridge joins them. The split itself is not the unusual part, and I want to be accurate about that: a shared domain graph beside per-tenant private graphs is a documented pattern with vendor tooling behind it, and knowledge graphs over a legislation corpus are well-established in both research and product.
What is unusual is what the second graph is built from and what it is asked to produce.
This is written for anyone who has wondered what actually sits under a claim like "AI for legal work" and for anyone about to point a language model at a pile of documents and hope.
Three generations of finding things
The reason two graphs exist at all is that we kept following retrieval further than we meant to.
The obvious starting point is keyword search: the model most people carry in their heads, because it is what a search engine taught them.
It matches the words you typed. In law that fails immediately, because the same obligation is written differently across decades of drafting, and none of those wordings need share your vocabulary.
Semantic retrieval fixes the vocabulary problem. Cut the corpus into chunks, give each chunk a numerical representation of its meaning, and closeness in that space becomes closeness in meaning.
Two passages that share no words at all can sit beside each other because they say the same thing.
That was where RYTZ started, and for a while it was enough.
What it cannot do is tell you how two things are related. Semantic similarity says these passages resemble one another.
It does not say this section amends that one, or that this obligation only bites once that threshold is met.
- 01
Keyword
Matches the words you typed. Fails the moment the same idea is written in different language, which in law is always.
- 02
Semantic
Matches meaning, not vocabulary. Two passages with no words in common can finally find each other.
- 03
Graph
Matches relationships. Not just what resembles what, but what amends, qualifies, triggers or supersedes what.
A knowledge graph is the third generation. It does not model the points; it models the lines between them. Flat similarity becomes something with structure: the difference between a map of where things are and a map of how they connect.
The corpus graph knows the law
The first graph is the one you would expect. Every section, every instrument, every relationship between them: what amends what, what a provision depends on, which sections are read together in practice. It is the same for every user of the platform, because the law is the same for every user of the platform.
The case graph knows one person
The second graph is where the value actually is, and it is built fresh for each client.
The legal AI platforms most people have heard of, among them Harvey, CoCounsel and Legora, sell to law firms and in-house teams, and they are built to work across a firm's document sets.
They are not built to assemble one person's chronology; the comparisons written about them say as much, describing portfolio analysis rather than dispute sequencing.
RYTZ starts from the opposite end: one matter, one person, and the order of what happened as the thing being produced.
That is the parent from the opening, and their seven years of email goes through the same treatment. It is extracted, chunked and dated. But what the graph is looking for is different.
What was actually said. What was actually meant.
How something written in March connects to something written four years later.
The graph assembles the links a person could not hold in their head across that span, and could not afford to pay someone to reconstruct by hand.
What the two graphs are for
The corpus graph tells you what the law says. The case graph tells you what happened. Neither one answers a client's question on its own. The answer lives in the bridge between them.
Documents are not discarded
There is a version of this architecture where the graph replaces the source, and that is not what we built.
Every document a user uploads still moves through the pipeline and is still stored for retrieval.
The graph is an index over the evidence, never a substitute for it.
This is also the difference between a system built for one job and a system built for anything.
The frontier models are very good at reasoning over what you hand them. They are not structured around one task.
RYTZ is: gather the evidence, build the chronology, read the correspondence, find the gaps, find the openings.
That pipeline is the product. The model/s are a component inside it.
The tooling exists. The distribution does not.
None of this is new to the legal industry. Not the corpus graph, not the chronology, not the retrieval that knows what relates to what.
It is what a firm with a technology budget has been buying for years.
The intelligence was never the scarce thing. Access to it was.
Australian lawyers are, on one measure, the most AI-exposed in the English-speaking world.
Clio's 2025 Legal Trends Report put 98% of Australian legal professionals using AI in some capacity, ahead of the United States, Canada and the United Kingdom.
That number is worth reading carefully, because it is broad rather than deep.
A 2026 LEAP survey of 700 legal professionals across six countries found only 16% of Australian respondents using legal-specific AI daily or as part of core workflows, against 49% globally and 37% using integrated AI regularly, against 57% globally.
Separate research into the Victorian profession put AI use in practice at 36.7%.
So the profession touches AI more than almost anyone and has embedded it less than most. Breadth without depth.
Now hold that beside the person from the opening.
They are not in the 98% or the 16%. They are not in any of these surveys, because the surveys are of the profession, and they cannot afford the profession.
The tooling that would let them understand their own evidence has existed for years behind a price they were never going to pay.
That is the part worth changing. Not the invention of the intelligence; that argument is over.
The question is whether it stays inside the firms that can buy it, or reaches the people whose lives actually turn on the answer.
Everything in this architecture is chosen for the second.
Build for the model you will have
The whole thing is deliberately lean, and I would call it fragile in the way any young system is fragile.
AI models are never going to be as dumb as they are today.
That is the assumption everything rests on. Capability improves; the foundation is what determines whether you can use the improvement when it arrives.
So the work is in the structure: how evidence is chunked, how conversations are read, how it all loops back together. And it is in the part that is easy to treat as secondary and never is, which is presenting it so that someone in the middle of the worst year of their life can understand what they are looking at and act on it with some clarity.
That is the goal. Everything else is plumbing.