FinTax
Retrieval-first research on Finnish tax law: find the right provision before generating anything.
- statutes ingested from Finlex with provenance
- 7
- statutes ingested from Finlex with provenance
- research questions, written down before experiments
- 6
- research questions, written down before experiments
Overview
General-purpose LLMs answer Finnish tax questions fluently but unreliably. They mix up tax years, invent section numbers and blend in other countries' rules. If the answer comes from the wrong provision, generation can't fix it. So FinTax treats retrieval as the research problem: build a benchmark of real sole-proprietor questions with gold provisions, then measure BM25, dense and hybrid retrievers on it before any RAG is built.
Research questions
- How well do BM25, dense and hybrid retrievers find the correct provision for real questions written in Finnish?
- Which chunking unit works best for legal text: section (§), subsection (momentti), fixed windows or hierarchical?
- How much does handling Finnish morphology (lemmatisation, compound splitting) help lexical retrieval?
- Are multilingual embedding models good enough for Finnish legal language?
- How should temporal validity be represented so that retrieval returns the version of the law that applies to the tax year?
Done so far
- Statute ingestion from the Finlex open-data API, recording the source, retrieval date, statute number and version of every document.
- Dated in-force snapshots. Finlex's consolidated texts mix current and not-yet-in-force wording, and the build step separates them so every provision has a validity window.
- A strict-typed, tested codebase. Raw data is immutable, and all cleaning happens in code.
Next
Verohallinto guidance ingestion, the FinTaxBench question set with a labelling guide, chunking variants, and the first BM25, dense and hybrid baselines reported with Recall@k, MRR and nDCG. Generation (FinTaxRAG) only starts once retrieval has a written evaluation.
There are no retrieval results yet. This page will be updated when the first baselines are in.
Related projects
RAG Sales Analysis
Ask business questions about a sales dataset in plain English, fully offline.
PDF Q&A Chatbot
Chat with your documents, with answers that cite their sources.

Failure Modes in Multi-Agent LLM Systems
A chaos-engineering study of how multi-agent coding pipelines break.