Legal research

A corpus of primary legal sources that a model searches and quotes from directly, instead of recalling what it once read about the law. Three jurisdictions are loaded: Switzerland, France and Italy, 1,105,378 documents between them.

Access sits behind a per-account permission. On this installation no ordinary account holds it yet, so the feature is reachable for the operator only for now.

What is loaded

Loaded1,105,378 documents

Switzerland30,444

Confederation5,321
Cantonsall 2616,300
Communes196 communes8,795

Source: Official enactments from the systematic collections.
Licence: These rows carry no licence line, so none is printed with them.

France918,113

National918,113
- LEGI719,189
- KALI198,924
Below national levelnot loaded

Source: DILA open data.
Licence: Licence Ouverte 2.0 (Etalab). Crediting the source is a condition of that licence, so the credit is printed with every citation.

Italy156,809

National155,681
Regional1,128

All of it from the Autonomous Province of Trento.
Communalnot loaded

Source: Normattiva, published by the IPZS.
Licence: CC BY 4.0, and CC BY 4.0 for Trento as well. Crediting the source is a condition of that licence, so the credit is printed with every citation.

Samples, not corporasample

European Union5
United States4
Canada3

These documents sit in the store, and the picker marks them as a sample rather than a corpus. An answer drawn from them says so, because three documents are not a legal system.

Left out on purposenot loaded

Germany

Statutory text is unprotected, but the database right over the collection around it is unresolved.
Austria

Reuse is granted only after written notice to the authority, and that notice has not been filed.

Where it stops

The corpus is not complete, and the answer says where it stops. French and Italian communal law is not in it. A question that needs it gets told so, instead of being handed national law in its place.

Some of what is loaded cannot be read either. These Swiss files are scans with no text layer to search:

Swiss scans with no text layer1,190
Of those, communal1,182
Communes affected110 of 196

The answer names that count for the commune you asked about, rather than letting the gap pass as silence.

The line

A question outside the store is refused before the search runs. The relevance floor underneath that refusal is measured, not guessed: it comes from two sets of questions, one the corpus does answer and one it does not.

26 questions the corpus does answer

None of them scored below 0.45 coverage.

The line · 0.45

Between 0.42 and 0.45, no question in either set landed here.

15 questions it does not answer

None of them scored above 0.42 coverage.

The answerable set bottoms out at 0.45 and the refusable set tops out at 0.42, so 0.45 is where the line sits. Below it you get a refusal that names the terms the corpus did not recognise, not a confident paragraph built from whatever happened to match.

What comes back

Answers cite articles and judgments so you can open the source and read it yourself. That is the intended way to use this: the model finds and quotes the provision, you decide what it means for your situation.

Four things ride along with every answer:

This is research, not legal advice. The server attaches that line rather than asking the model to include it, so it cannot go missing.
The sources the answer quoted.
The credit the French and Italian licences require.
A section headed with what this corpus does not cover.

Before you act on it

No lawyer has looked at your matter, and the corpus does not know your facts. Take the citations to someone qualified.

Something missing or wrong here? Say so. The documentation is part of the product, not an afterthought. All topics · Legal & privacy · Home