Case law weighed by how it is cited, not just by keyword

The problem

Keyword search over case law has a peculiar failure mode: it returns everything and ranks nothing. A search surfacing four hundred judgments containing a phrase has not helped the researcher, because the question was never “which judgments mention this”, it was “which of these actually decides the point.”

Precedent has structure. A judgment matters in proportion to how it is cited, by whom, and for what. That structure is present in the corpus and invisible to text search.

What we built

A research platform built on the citation graph rather than the text index.

The corpus (cases, statutes, judgments and precedents) is indexed for retrieval, but ranking draws on how cases cite one another, so a judgment the field has built on outranks one that merely shares vocabulary. A visualisation engine renders that structure, letting a researcher see which authorities a line of reasoning actually rests on.

Two further pieces: a document summarisation and case-linking model, and community relevancy signals: practitioners marking what proved useful for a given query, which is domain judgment no ranking function recovers on its own.

Alongside the research tool, a marketplace connecting clients with legal professionals.

What was hard

Building the corpus. A citation graph is only as good as its links, and legal citation in practice is inconsistent: the same authority appears in different forms across judgments, and resolving those to a single node is most of the work before any ranking can happen.

Outcome

The product went on to raise funding and is still live, selling subscriptions.