Project Mike

Introduction

If you’ve watched Suits, you’ll know Mike Ross. And if you work with the law, you probably envy him. In law, context is indispensable, and unfortunately most of us don’t have a photographic memory.

From my law degree, I know how hard it is to distill the relevant legal questions, find the correct legal basis, and interpret and apply it correctly. Swiss law spreads the relevant material across statutes, court decisions and commentaries, and missing one piece of context easily leads to the wrong conclusion.

A good retrieval system can act as that missing memory. In this project, I built and fine-tuned one for decisions of the Swiss Federal Supreme Court. The four posts below cover the legal background, a benchmark built from commentary citations, a reranker trained on LLM relevance labels, and the distillation of that reranker into the first-stage retriever. If you’re new to Swiss law, Part 1 gives you a crash-course on the most important details.

Blog

Part 1: A Primer on Swiss Law and Retrieval Systems

A crash course on Swiss law and the challenges it poses for information retrieval.

Part 2: Building a Heuristic Benchmark for Retrieval

Setting up a heuristic evaluation framework and benchmarking embedding models.

Part 3: Teaching a Cross-Encoder Swiss Law

Using LLM relevance labels to fine-tune and calibrate a legal reranker.

Part 4: Fine-Tuning First-Stage Retrieval

Using knowledge distillation to finetune first-stage retrieval.

Project Mike#

Introduction#

Blog#

Project Mike

Introduction

Blog