Kaggle Legal RAG

This repo contains the law-side retrieval experiments, data artifacts, prompts, and submission files for the Swiss legal citation retrieval task.

Folders

code/dev/

Active development code. This is where the main retrieval pipeline lives.

code/dev/utils/

Shared utilities, including API wrappers for calling hosted/open-source models so experiments can run faster than local model inference.

code/dev/law-only/lawside-retrieval/

Main law-side retrieval pipeline. The central script is law_document_side_retrieval.py.

code/dev/law-only/lawside-retrieval/trials/

Successful optimization layers used by the law-side pipeline, including court-law expansion, TF-IDF reranking, cross-encoder reranking, and LLM citation filtering.

code/archives/

Older one-off scripts kept for reference, mostly for creating intermediate datasets, audits, and helper artifacts.

code/prompts/

Prompt builders used by LLM-based reranking and filtering experiments.

data/

Input parquet files and prepared datasets, including law documents, court documents, train/val/test rows, embeddings, and subquery embeddings.

context/

Generated experiment artifacts: submissions, score caches, summaries, intermediate CSV/parquet outputs, and notes used to track runs.

Key Docs

pipeline.md

Step-by-step description of the final law-side retrieval and LLM filtering pipeline.

experiments.md

Chronological experiment log with commands, scores, and decisions.