markchweya/Baraka

Baraka is a multilingual Kenyan retail-banking & SACCO support assistant that helps customers ask questions, get instant FAQ-based answers, and submit complaints that are automatically routed to the right department, while giving admins a simple console to manage FAQs, complaints, and chat logs. # Baraka ✨ — Multilingual Banking & SACCO Support Assistant (Streamlit) Baraka is a modern, light-themed **multilingual customer support assistant** for Kenyan retail banking and SACCOs. It combines: - **FAQ retrieval** (custom FAQs + a base banking dataset), - **automatic complaint routing** to departments (ATM, Cards, Loans, Transfers, etc.), - an **Admin Console** for managing FAQs and complaint tickets, - **multilingual chat** with optional language locking (e.g. “reply in Kiswahili”). Baraka is designed to **avoid sensitive credential collection** (no PINs, OTPs, passwords). --- ## Features ### ✅ Multilingual Chat (with language lock) - Auto-detects language (when OpenAI is enabled) - Replies in the user’s language - You can force language in chat: **`reply in Kiswahili`**, **`reply in Somali`**, etc. ### ✅ Complaint / Inquiry Ticketing - Customers submit complaints via the **Complaint** page - Baraka routes to a department automatically using: - keyword rules, then - TF-IDF similarity routing - Saves each complaint into a local SQLite database (`bankbot.db`) ### ✅ FAQ Retrieval (Custom First → Base Dataset → AI fallback) Baraka answers in this order: 1. **Custom FAQs** (admin-created, per department) 2. **Base banking dataset** (Bitext retail banking dataset from HF) 3. **OpenAI fallback** (only if enabled and only using top retrieved context) ### ✅ Admin Console (role-based) Admins can: - Add / edit / delete FAQs - View and manage complaint tickets (status, priority, internal notes) - View chat logs (source + similarity scores) ### ✅ Modern Light UI Includes a glassy, modern, light UI with BaseWeb styling overrides (selectboxes, menus, inputs). --- ## Tech Stack - **Streamlit** UI - **SQLite** persistence - **Pandas** for data handling - **scikit-learn** TF-IDF retrieval + cosine similarity - **Hugging Face dataset (parquet)** as base FAQ dataset - **OpenAI API** for translation + fallback answering --- ## Project Structure - `app.py` — the enti …