We indexed every Act, regulation and judgment in Mauritius. Here's what we learned.

Building Themis meant assembling the complete published corpus of Mauritian law. The engineering story is less about scale and more about the shape of the data: bilingual, scattered, and full of traps for a system that needs to cite.

When we started building Themis, the first question was not which model to use. It was whether we could build a corpus complete enough that a lawyer could trust an answer without redoing the search.

The numbers we ended up with are published on the product page: 4,100+ Acts, 57,000+ judgments, 17,800+ regulations and 150+ bills and reprints. Those figures describe the size. What they do not describe is the shape of the work.

The corpus lives in more places than you expect

Legislation is not all in one format. Acts, regulations, statutory instruments and reprints come from different sources, at different intervals, with different conventions for amendments. Some are bilingual. Some are amended by a later Act that itself has been repealed. Some references use short titles that changed over time.

A system that simply ingests PDFs and hopes for the best will miss the amendments. We built a pipeline that tracks each instrument, its amendments, its reprints and its current status, so a citation points to the law as it stands today.

The bilingual problem is real

Mauritian law operates across English and French. The Code Civil Mauricien is French-origin; much of the case law that applies it is in English. The same concept can be expressed differently in the two traditions. A search that looks only for English keywords will miss French-origin authorities, and vice versa.

Our approach was to index both languages and to treat cross-language references as first-class links in the corpus. A query in English can surface a French-origin provision if the concepts align, and every answer carries the original citation.

Judgments are harder than statutes

Statutes are structured. Judgments are not. They come as PDFs with varying headers, page numbers, marginalia and formatting. Extracting the ratio, the cited authorities and the procedural history requires more than text extraction.

We invested in parsing that identifies the parties, the court, the date, the cited authorities and the headnotes where they exist. The result is that a Themis answer can cite a judgment precisely, not just point to a document that might contain the relevant paragraph.

Citation is the feature, not the afterthought

The most important design decision was that every generated answer must include citations that a lawyer can verify. That changes how the model is used: it is a retrieval-and-synthesis system, not a free-text generator. If a source does not exist, the system is trained to say so rather than invent one.

This is why the corpus has to be complete. A missing source is not a minor gap; it is a reason for the user to stop trusting the answer.

Keeping current is the quietest hard problem

New judgments appear regularly. Acts are amended. Regulations are added. The corpus is never finished. We run an update process that ingests new material, checks it against what we already have, and flags conflicts for human review.

What it taught us about client work

The same discipline applies to client automations. A system that files documents without knowing the firm’s naming convention is unreliable. A system that answers questions without citing sources is dangerous. Building Themis first forced us to solve these problems at a scale that makes smaller client systems straightforward by comparison.