Curated Law and the Knowledge Graph in Agentic Legal Research

What does it mean to curate law for AI agents?

Curating law for AI agents means enriching primary law into a data structure built for how agents read, rather than for how humans browse. Each opinion is tagged with its determinative facts, its rules of law, and its key quotations, each element located with pinpoint specificity. The connections among authorities are recorded as data.

The digests and encyclopedias of the print era curated the law for human eyes, so their structures reflect the limits of human browsing. A print taxonomy is consulted one topic at a time. A headnote serves a reader scanning a page. An agent needs the same editorial knowledge in a different form, as tagged, located elements it can retrieve, compare, and cite.

When one opinion quotes another, the quotation is captured as a tagged object carrying an embedding of its meaning, the numerical representation Chapter 4 explained. Comparing embeddings across the collection then surfaces every statement that cross-references the quoted case, together with semantically close statements quoting other authorities, a web of doctrinal connection that neither a fixed taxonomy nor any query would expose. The citator's treatment data serves double duty in the same structure. A notation that a case has been distinguished is a warning about authority and equally a map, because the distinguishing opinion leads directly to closely related law.

Curation also addresses the hardest step in retrieval-augmented generation (RAG), finding the exact passages that bear on the inquiry. A RAG answer is composed only from the chunks retrieval returns. Locating those chunks in raw text is the difficulty, as Chapter 1 explained. When the determinative facts, rules of law, and quotations are captured in advance, retrieval targets the elements that matter instead of guessing where the useful text sits. Every proposition the agent later writes can then be traced to a pinpoint source passage, the structure Chapter 8's validation audit depends on.

How does an AI agent change the recall and precision trade-off in legal research?

In legal research, an AI agent changes the recall and precision trade-off by reading every authority a broad search returns. A human researcher runs narrow, high-precision searches because no one can read a thousand results. The agent runs the broad search that maximizes recall and reads everything it returns. Precision comes from the reading, not from narrowing the query.

Narrow queries are a major cause of missed authorities, as Chapter 3 documented. A query precise enough to keep the reading manageable excludes the case phrased in unfamiliar vocabulary, the authority relevant for its facts rather than its terms, and the minority position that never ranks high. The agent keeps the broad net and applies its judgment after retrieval, evaluating each authority on its own content, its facts, its rule of law, and its treatment, rather than on its rank in a results list.

The reading budget that confined human research to the high-precision, low-recall corner of the trade-off does not bind a system that reads the full candidate set, hundreds or even thousands of authorities against an associate's realistic budget of a few dozen. Nothing a broad search returns goes unread.

Depth in legal research has always been rationed by human reading time. The agent removes the ration. That single change, more than any other, is what separates the third era of legal research from a faster second era.

What is a legal knowledge graph?

A legal knowledge graph is a data structure that records the law's authorities and the connections among them, so that an AI agent can research by moving from one authority to the next rather than by running searches. The recorded connections include citations, quotations, citator treatment, shared topics in a legal taxonomy, and common rules of law.

The oldest layer is the citation graph, the record of which cases cite which. Legal publishers have kept that record for over a century. Thomas Smith's The Web of Law mapped nearly four million American precedents as a network, finding precedential authority concentrated in a small number of heavily cited cases. A legal knowledge graph extends that record with the substance of each link. When one opinion quotes another, the quotation itself is captured with its location. When a citator marks a case as followed or distinguished, that treatment becomes a path the agent can follow. Shared topics and common rules of law connect authorities that never cite each other at all.

The structure fits the law because a case draws its force from the decisions it follows and loses it to the decisions that overrule or distinguish it. Lawyers have always researched by tracing those ties by hand, from a case to the authorities it cites and on to the later cases that treat it. A legal knowledge graph records them as data, so an AI agent can move along the same paths at a scale beyond any human researcher's reach.

What is retrieval by graph traversal in legal research?

In legal research, graph traversal is retrieval that follows the recorded connections among authorities instead of running a query. The agent moves from a case to everything that cites, quotes, follows, or distinguishes it, the way a careful scholar works outward from a leading decision. No search, keyword or semantic, is involved. The connections themselves are the retrieval path.

Traversal reaches law that no query would surface, because the connection, not the vocabulary or even the semantic similarity, is what makes the authority relevant. A later case that distinguishes a controlling decision may share almost none of its language, so neither a keyword nor a semantic match would pair the two. The recorded treatment joins them in a single step.

From a controlling opinion, the agent moves to the cases that quote its rule and checks the treatment of each. A distinguishing opinion then leads it to the closely related law that case applied instead. The traversal ends in a line of authority the original query never named.

Traversal and search are companions rather than competitors. A broad search seeds the graph with starting authorities. Traversal extends outward from each one, so the two retrieval modes cover each other's misses. Chapter 4 explained the search half of that pairing.

How do keyword search, semantic search, and graph traversal compare in legal research?

In legal research, keyword search matches the words of the query, semantic search matches its meaning, and graph traversal runs no query at all, following the recorded connections among authorities instead. The three methods differ on what they match, what they find, and where they fail.

Each method's blind spot sits inside another's coverage. Keyword search misses authority phrased in vocabulary the query never guessed, a gap semantic search closes by matching meaning. Semantic search in turn overlooks the authority related by connection rather than by similar content, which traversal reaches directly. And traversal cannot leave the graph of its starting cases, so the broad searches supply the entry points.

An agentic system runs all three, which is why no single retrieval method describes it. Keyword and semantic search are available to any lawyer at a search box, as Chapter 4 explained. Graph traversal, and an agent that reads everything the other two return, belong to the agentic era.

Why not just run an AI agent over raw case law?

An AI agent over raw case law can read broadly, but it cannot follow connections that were never recorded, verify treatment against a citator that does not exist, or cite to tagged passages that were never tagged. The curation is what the agent's depth runs on.

The difference between an agent over enriched, connected law and an agent over raw text is the difference between an associate working in a well-organized library and one working from a pile of slip opinions. Whether a case remains good law appears nowhere in the case itself. That answer lives in the later opinions that treat it, which is why the citator exists, as Chapter 6 explains. A rule of law sits unmarked inside pages of procedural history, with no boundary showing where a supporting passage begins or ends.

This curation also answers the cost question Chapter 2 raised. The editorial structures that once took armies of editors can now be generated and kept current by AI, extending as the law develops rather than waiting on a publisher's editorial cycle or a taxonomy fixed in the 1800s.