martel-blip/ai_10022200110

Retrieval-augmented QA over Ghana presidential election results (1992-2020) and the 2025 Budget Statement. Hybrid BM25 + dense retrieval, custom NumPy vector store, Groq LLM, Streamlit. CS4241 project. # Ghana Elections + 2025 Budget — RAG **CS4241 project · index 10022200110** A retrieval-augmented QA system over two Ghanaian public-interest documents: - Ghana Presidential Election results for 2012, 2016, 2020 - The Government of Ghana **2025 Budget Statement and Economic Policy** Stack: **Groq (Llama 3.1 8B Instant) · sentence-transformers · custom NumPy vector store · hybrid BM25 + dense retrieval · Streamlit**, deployed on Streamlit Community Cloud. --- ## Feature summary - **Ingestion** — 866 raw docs → 397 chunks (98 election year-region groups + 299 sentence-window PDF chunks, 300w / 50w overlap). - **Embeddings** — `all-MiniLM-L6-v2`, 384 dims, L2-normalized, disk-cached. - **Vector store** — from-scratch NumPy, cosine = dot product, `argpartition` top-k. - **Hybrid retrieval** — dense + custom BM25, fused with Reciprocal Rank Fusion (k=60), then dense-cosine reranking on the pooled candidates. - **Prompt & hallucination guard** — strict system prompt, numbered `[C1]` citations, post-hoc grounding audit flags suspicious answers. - **Feedback-loop innovation** — Chunk Reputation Boosting: each thumbs-up/down updates per-chunk reputation and influences future retrieval, bounded to prevent runaway effects. - **Adversarial evaluation** — ground-truth, numeric, out-of-scope, and false-premise cases with reported hallucination and refusal rates. - **Streamlit UI** — Chat, Retrieval inspector, Failure-case demo, Evaluation, Feedback report. --- ## Repository layout ``` rag_project/ ├── app.py # Streamlit UI ├── data_loader.py # (your module) CSV + PDF loader ├── chunker.py # (your module) chunking with filters ├── modules/ │ ├── __init__.py │ ├── embedder.py # MiniLM wrapper, disk cache │ ├── vector_store.py # NumPy-only k-NN │ ├── retriever.py # BM25 + dense + RRF + rerank + failure-case demo │ ├── prompt_builder.py # grounded prompt + citati …