Skip to content

1 Architecture

Bosaj edited this page Nov 5, 2026 · 4 revisions

System architecture

Ministry of Justice portal (official PDFs)
        │  tools/crawl_adala.py, download_core.py
        ▼
tools/build_corpus.py ── src/pdf_articles.py (rebuild Arabic from glyph positions)
        │
        ▼
data/corpus.sqlite: laws, articles, FTS5 indexes, document catalogue
        │
        ▼
src/corpus.py: named articles, legal index, Darija word map, BM25
        │
        ▼
src/legal_rag.py: citations with source links, exact-match check, short note
        │
        ▼
server.py (local) or api/index.py (Vercel) ── web/ (React interface)

Principles

  • Nothing is generated. Every quote is a row cut from an official PDF and is compared with the corpus before it is shown.
  • Declining beats guessing. A question that matches too little gets no articles and a pointer to the document search.
  • Questions are private. Only language, field of law and result counts are stored.
  • The server needs no packages. It runs on the Python standard library and SQLite FTS5. pymupdf is only needed to rebuild the corpus.

Other parts

  • src/bo_diff_engine.py: Bulletin Officiel index and text comparison. Its comparisons are hand-written illustrations and are labelled as such in the interface.
  • src/contract_auditor.py: pattern checks for a few clauses the law forbids. Tests confirm the articles it cites exist.

Under load

Requests run on several threads at once, so each thread has its own read-only SQLite connection (Corpus.db). Usage counters can never fail a request, and they are off on serverless hosts, where each instance has its own disk. Law text routes may be cached at the edge for a day. Answers are never cached.

Measured on production on 2026-11-05: 120 of 120 questions answered at 20 at a time, p95 about 1.9 seconds. Before this work, 15 of 120 returned 500.

Clone this wiki locally