AlikoGrace/ghana-health-policy-raq-qa

Retrieval Augmented QA over Ghanaian Health Policy Documents # Retrieval-Augmented QA over Ghanaian Health Policy Documents ## Research Abstract We ask whether Retrieval-Augmented Generation (RAG) can accurately answer policy-related questions from real-world Ghanaian health documents. We scraped 36 official PDFs from moh.gov.gh, implemented three chunking strategies (word, overlap, sentence), and built a dense retriever using FAISS + MiniLM. FLAN-T5 was used to compare zero-context (baseline) and retrieval-augmented generation. On 10 gold QA pairs extracted from the documents, RAG improved factual accuracy by 33% and eliminated hallucinations. This experiment demonstrates the potential of RAG in low-resource, document-heavy public sector domains like African health governance. --- ## Motivation Ghana’s health policy documents are extensive but difficult to query directly. Important guidelines on the use of antibiotics, malaria treatment, or healthcare waste,are buried in long PDF reports that are not structured for search or question answering. This project applys Retrieval-Augmented Generation (RAG) to build a smarter QA system grounded in official documents. It supports domain-specific queries like: > “What is Ghana’s protocol for treating malaria in the first trimester?” > “What does the national health policy say about medical waste disposal?” --- ## Project Design The system was designed with multiple research-focused components: ### Data Collection - Scraped **36 government-issued PDFs** from moh.gov.gh - Filtered out corrupted or scanned documents using automated word count checks ### Chunking Experiments Three chunking strategies were implemented and logged: 1. **Fixed 300-word windows** 2. **Overlapping windows (300 words, 100-word stride)** 3. **Sentence-based segmentation (via NLTK)** All chunk logs and statistics were saved for later retrieval quality comparison Only **word300** was used in the current evaluation phase ### QA Pipeline - Built a FAISS index using `sentence-transformers (MiniLM)` …