Commit History
-
Inscription libre, import de tableur, épinglage, signalement, jeu de test
Samuel ADONE +1: Claude Opus 5Inscription : un bénévole crée son compte même si son numéro n'a jamais été enregistré. Le compte naît sans groupe et ne peut pas être superadmin ; les superadmins sont notifiés pour l'affecter. SELF_REGISTRATION=0 restaure la liste blanche seule. Import : POST /api/admin/users/import lit un .xlsx ou un .csv et retrouve seul les colonnes prénom / nom / téléphone (en-têtes français accentués, colonne « Nom complet », numéros abîmés par Excel, séparateur point-virgule, CP1252, absence d'en-tête). Aperçu avant écriture, colonnes corrigeables à la main. Rôles : l'admin de groupe ajoute et retire des bénévoles et épingle les messages ; seul le superadmin distribue le rôle d'admin. Nouveau signalement de compte, distinct du blocage, et recherche d'annuaire par numéro espacé. Stockage : `meta` filtré champ par champ pour qu'aucune image ne puisse être écrite en base, rétention calée sur la cible réelle (500 messages/groupe, 30 jours), requêtes agrégées sur les écrans rechargés en permanenc…2bc5c2a -
first commit
Samuel ADONE5e0d7d2 -
gitignore
Samuel ADONE7616d37
-
Fix build: pin Ollama, download from GitHub (install script 404s on dropped .tgz)
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>b49f1b9 -
Fix grounding: qwen2.5:1.5b + embedding re-rank + topic-relevance prompt
The 0.5b model hallucinated on questions whose answer isn't the top BM25 hit (e.g. invented a "car importation tax" figure, parroted excerpt headers). Fixes: - Re-enable embedding re-rank (auto) — bridges lexical gaps (car tax -> excise/ customs); chunks pre-warmed in a background thread so queries stay fast. - qwen2.5:1.5b: grounds reliably where 1b/0.5b hallucinated (e.g. deficit). - Prompt now requires same-subject excerpts and to decline (not invent) when the topic isn't covered. Verified: car tax correctly declines; maternity/revenue/ deficit/tobacco/motor-vehicle measures all correct. - Restore recall (top_k 4, budget 4000, num_ctx 4096); EMBED_MODEL pulled at boot. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>8f702fb -
Faster model + multi-turn context + responsive UI with linked citations
- Model: qwen2.5:0.5b at temperature 0 with a concise prompt — benchmarked fastest (~3x) AND most accurate (5/5 on number-extraction test) of the small models; greedy decoding fixes the prior wrong-figure issues. - Conversation memory: prior turns now steer both retrieval and the model, so terse follow-ups ("financially how does it change?") stay on topic. - UI: full-width, mobile-responsive single column; indexed-docs list moved to a collapsible accordion at the bottom; citations link to the official PDF at the cited page (#page=N) with the matched excerpt as a hover tooltip. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>7b684b6 -
Switch to llama3.2:1b: accurate on numbers, still fast (0.5b gave wrong figures)
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>92cbd80 -
Use smaller, faster qwen2.5:0.5b model for free CPU tier
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>72bcb9f -
Tune for free CPU tier: smaller ctx, trimmed retrieval, no re-rank, capped tokens
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>429fd32 -
Install zstd for Ollama (required by its install script on amd64)
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>c596df9 -
Mauritius Budget 2026-2027 RAG assistant (Ollama + Gradio)
RAG over the National Budget 2026-2027 PDFs: PyMuPDF ingest + chunking, BM25 retrieval with optional Ollama embedding re-rank, streaming answers with document/page citations, and a pure-Python Gradio UI. Self-contained Dockerfile runs Ollama and the app together on port 7860 for HF Spaces. PDFs stored via Git LFS. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>c5f394c