Deadsunx/kora

Retrieval-augmented QA over French-language African business law (OHADA). An ablation study: which RAG components actually earn their cost, measured on a hand-validated gold set. # Kora **Retrieval-augmented question answering over French-language African business law (OHADA).** > Status: **all eight phases complete.** Every number below links to the > write-up that produced it. **→ Read the technical report** — the whole project in one document, organised around how each finding was found. (markdown version) --- ## Why this project Most retrieval systems are built and evaluated in English. The techniques that work there are assumed to transfer, and usually nobody checks. This project checks. The corpus is the **Actes uniformes OHADA** — the body of business law shared by 17 mostly francophone African states. It is a good testbed for three reasons: - **It is French, and legal French.** Dense retrievers trained mainly on English web text tend to blur exactly the tokens that matter here: article numbers, fixed terms of art like *société anonyme* or *sûreté mobilière*, cross references between texts. - **It is highly structured.** Livre → Titre → Chapitre → Article. That structure is a gift for chunking and a fair test of whether structure-aware chunking beats naive splitting. - **Answers must be attributable.** "Article 640 AUSCGIE" is checkable. A legal assistant that cannot cite, or that will not say *I don't know*, is worse than useless — so citation accuracy and abstention are first-class metrics here, not afterthoughts. ## The actual thesis The deliverable is not a chatbot. The deliverable is an **ablation table**: a defensible account of which components earn their cost, in accuracy and in latency, on this corpus. A demo that answers questions is the by-product. This ordering is deliberate. The evaluation harness (Phase 2) was built *before* any retrieval improvement (Phase 3), because an improvement you cannot measure is a preference, not a result. ## What exists **3,056 articles** parsed from source PDFs across 10 acts · **64 human-validated questions**, 56 answerable · **23 recorded runs** · 252 tests · single RTX 4070 La …