FOMDJ/ai_10012200004

CS4241 RAG exam — Ghana Knowledge Assistant # Ghana Knowledge Assistant CS4241 — Introduction to Artificial Intelligence | End-of-Semester Exam Student: Farima Konaré | Index: 10012200004 Lecturer: Godwin N. Danso | Academic City University --- ## What it does Answers questions about Ghana's presidential election results (2000–2020) and the 2025 national budget. Built as a custom RAG pipeline — no LangChain, no LlamaIndex. Chunking, retrieval, and prompt construction are all hand-coded. **Live demo:** [Add Streamlit Cloud URL after deployment] --- ## How it works ``` User question -> Query expansion (vague geographic queries get domain context added) -> Hybrid retrieval: BM25 keyword + FAISS vector search, weighted 40/60 -> Confidence check: if best score Context window: top chunks truncated to ≤2000 tokens -> Prompt built with last 5 conversation turns injected as memory -> Groq / Llama 3.3 70B generates a grounded answer -> Every stage logged and visible in the UI ``` --- ## Setup ### 1. Install dependencies ```bash python -m venv venv source venv/bin/activate # Windows: venv\Scripts\activate pip install -r requirements.txt ``` ### 2. Download the datasets ```bash python scripts/download_data.py ``` Pulls the Ghana election CSV and 2025 budget PDF into `data/raw/`. ### 3. Set your Groq API key Get a free key at console.groq.com, then create `.streamlit/secrets.toml`: ```toml GROQ_API_KEY = "your_key_here" ``` ### 4. Run ```bash streamlit run app.py ``` First run builds the FAISS index over ~1,565 chunks — takes about a minute. Loads from disk after that. --- ## Project structure ``` ├── app.py # Streamlit chat UI ├── requirements.txt ├── CAPABILITIES.md # Feature overview in plain language ├── data/ │ ├── raw/ # Downloaded datasets (gitignored) │ └── processed/ # FAISS index + chunk cache (gitignored) ├── src/ │ ├── ingestion/ # CSV and PDF loading, chunking │ ├── retrieval/ # Embedder, FAISS sto …