Multi-Hop Question Answering with KG-Infused RAG Using Wikidata5M

Türkiye Domain Production-ready Streamlit/Python project for Türkiye-domain analysis, verified multi-hop question generation, KG-guided spreading activation, query expansion, answer generation, experiments, case studies, GNN artifact display, and XAI evidence using Neo4j as the only KG backend.

---

📌 Project Principles

• The project never ships dummy records.
• UI panels read from Neo4j or real artifacts generated by the scripts.
• Missing data is displayed as an explicit empty state.

---

⚙️ Setup

python -m venv .venv
.venv\Scripts\activate
pip install -r requirements.txt
copy .env.example .env

Edit .env if your Neo4j or Llama 3 settings differ.

---

🧠 Neo4j Defaults

• Bolt URI: bolt://localhost:7687
• User: neo4j
• Password: passwordv0.4
• Database: neo4j

The database is expected to already contain Wikidata5M-style Entity nodes and relationships.

---

🚀 Generate Real Outputs

python scripts/discover_turkiye_entity.py
python scripts/analyze_domains.py
python scripts/build_question_dataset.py --domain auto --min-two-hop 30 --min-three-hop 15 --min-comparison 5
python scripts/run_experiments.py
python scripts/generate_case_studies.py
python scripts/compute_gnn_stats.py
python scripts/compute_xai_artifacts.py
python scripts/verify_paths.py

All generated outputs are written under outputs/ and are based on Neo4j query results or real pipeline computations.

---

▶️ Launch

streamlit run app.py

Use the sidebar to:

• Verify the Neo4j connection
• Choose domain
• Tune activation rounds
• Select case studies
• Refresh computed artifacts

---

🧪 Tests

pytest

Tests require a reachable Neo4j instance with the imported graph.