Skip to content

LexTreeRAG

What is this?

LexTreeRAG is a vectorless legal RAG system that answers questions about EU law by searching the live EUR-Lex CELLAR database — without needing a vector database or pre-embedded corpus.

How it works

User question
  │
  ├─ 1. Classify question ──────── Claude Haiku
  ├─ 2. Extract keywords + year ── Claude Haiku
  ├─ 3. Search EUR-Lex SPARQL ──── publications.europa.eu
  ├─ 4. Fetch documents ────────── HTML → Markdown  /  PDF → text
  ├─ 5. Build reasoning trees ──── parse Articles → summarise → cache JSON
  ├─ 6. Navigate trees ─────────── Claude Haiku picks relevant articles
  ├─ 7. Python reranking ───────── obligation score + list score
  ├─ 8. Generate answer ────────── Claude Opus 4.6 (streaming)
  └─ 9. Confidence scoring ──────── Claude Haiku → retry if < 50%

Architecture

Pattern B — Unified Streamlit. There is no separate frontend and backend.

Layer File Role
UI + Orchestration app.py Streamlit interface, calls pipeline modules directly
CLI main.py Terminal entry point, same pipeline
Pipeline pipeline/*.py Keyword extraction, EUR-Lex search, tree building, navigation, ranking, answer generation
Cache data/cache/ JSON reasoning trees per document
PDF Archive data/pdfs/ Raw PDFs downloaded from CELLAR

External dependencies

Service Purpose
Anthropic API — Claude Haiku 4.5 Keyword extraction, question classification, article summarisation, tree navigation, confidence scoring
Anthropic API — Claude Opus 4.6 Final answer generation with adaptive thinking
EUR-Lex CELLAR SPARQL Live document discovery (publications.europa.eu/webapi/rdf/sparql)
EUR-Lex CELLAR content HTML and PDF document fetch
Google Gemini (optional) Hybrid second-opinion answer with Google Search grounding

Pages