Graph-RAG XAI Explorer

Explainable Graph-RAG for multi-hop question answering on HotpotQA.

A relational GNN re-ranks passages over a knowledge graph of entity "bridges", a local LLM reads the result, and every step, retrieval, ranking, reasoning path, and answer, is inspectable through a live D3 explorer.

Course project, Current Trends in Machine Learning, MSc, 2nd year.

---

Interface

The reasoning graph (left) shows the retrieved passages and the bridge entities connecting them; the panel on the right shows the answer, the multi-hop reasoning paths, and the ERASER necessity/attribution scores for each passage. Click a bridge entity to see the counterfactual answer if it were removed.

---

What this is

Answering a HotpotQA question means combining facts from two different Wikipedia paragraphs, linked by a shared entity (the "bridge"). Most RAG pipelines retrieve those two paragraphs independently and hope the reader stitches them together; this project makes the bridge itself part of the model and exposes it as an explanation.

Pipeline

1. Dense retrieval (BAAI/bge-large-en-v1.5) selects candidate passages for the query.
2. Bridge graph construction: entities are linked via ReFinED and connected by co-occurrence across passages, weighted by IDF specificity so that generic hubs (countries, years) don't dominate, this is the multi-hop signal HotpotQA is built on.
3. R-GCN re-ranker (torch_geometric) scores passages over that graph; its score is fused with the dense score (fusion_alpha), so the graph refines dense retrieval rather than replacing it.
4. Reader: a local LLM (qwen3.5:9b via Ollama) reads the top passages and answers with explicit supporting sentences.
5. XAI layer: reasoning paths, an ERASER-style necessity/sufficiency overlay, a dual GNN/Dense attribution signal, and counterfactual entity ablation, all served to the same web UI shown above.

Results

On the HotpotQA dev set, the graph-based re-ranker (gnn_bridge_qwen9b) reaches Joint F1 0.52, clearly above a strong dense-only baseline (baseline_dense, 0.40), with the same reader and prompt for every system so the comparison isolates the retrieval architecture. See Report.pdf for the full methodology, all four baselines (BM25, dense, PPR-over-bridge-graph, R-GCN), ablations, and the XAI evaluation (ERASER, attribution agreement, counterfactual robustness).

Explainability

Every answer in the UI is backed by four complementary explanation layers:

Layer	Question it answers	Level
Reasoning paths	Which passage → bridge entity → passage chain produced the answer?	Retriever
Dual GNN/Dense signal	Where did the graph change the ranking vs. dense alone?	Retriever
ERASER overlay	Which retrieved passages are necessary, does removing them flip the answer?	Reader
Counterfactual ablation	What happens to the answer if a specific bridge entity is removed?	Interactive

attribution recall, the fraction of ERASER-necessary passages the GNN actually surfaced in its top-K, is the project's key diagnostic against explanation mismatch (a retriever-side explanation that doesn't match what the reader used).

---

Quickstart, lightweight demo (no GPU, no Ollama)

The fastest way to see the system: a FastAPI + D3 web app that serves the interface above from a pre-computed cache of example questions, full reasoning graph, all XAI panels, zero setup beyond Python.

git clone https://github.com/fraadap/GraphRAG.git
cd GraphRAG
./run_light.sh

Then open http://localhost:8000. run_light.sh creates a virtual env, installs the ~6 lightweight dependencies (requirements.txt), and starts the server in cache-only mode:

XAI_CACHE_ONLY=1 python -m uvicorn src.api.app:app --host 0.0.0.0 --port 8000

Pick a question from Examples ▾, answer, graph and every explanation panel render instantly. Typing a question outside the cached set returns a clear message: the offline demo only serves pre-computed questions (see below for live answering).

Running the full pipeline (GPU + Ollama)

The offline demo is a thin shell over the real, trainable pipeline in src/. To answer arbitrary questions live, reproduce the results, or retrain the re-ranker:

# the committed requirements.txt is the lightweight demo's, the full pipeline# additionally needs the GPU/ML stack:
uv pip install torch --index-url https://download.pytorch.org/whl/cu124
uv pip install torch-geometric scikit-learn sentence-transformers transformers accelerate rank-bm25
uv pip install "git+https://github.com/amazon-science/ReFinED.git" --no-deps
uv pip install ujson unidecode nltk boto3 lmdb fastcoref
# 1. download HotpotQA
python scripts/download_hotpot.py
# 2. build the knowledge graph + bridge graph
python scripts/build_kg_tier1.py --config configs/kg_tier1.yaml
python scripts/build_bridge_graph.py
# 3. train the R-GCN re-ranker
python scripts/train_gnn.py --config configs/gnn_train.yaml
# 4. evaluate a system end-to-end (retriever + reader + official HotpotQA metrics)
python scripts/run_eval.py --config configs/gnn_bridge_qwen9b.yaml
# 5. serve the live UI (GPU model + Ollama reader instead of the cache)
uvicorn src.api.app:app --host 0.0.0.0 --port 8000

Requires an Ollama server (ollama pull qwen3.5:9b) reachable at http://localhost:11434, and a CUDA GPU for GNN training/inference. Every retriever (baseline_bm25, baseline_dense, ppr_qwen9b, bridge_qwen9b, gnn_bridge_qwen9b) shares the same reader and prompt, so configs/*.yaml is the single source of truth for a reproducible comparison.

---

Project structure

src/
├─ api/ FastAPI backend (cache-only + live serving) + static D3 frontend
├─ data/ HotpotQA loader + official eval wrapper
├─ kg_build/ coref, entity linking (ReFinED), relation extraction, bridge graph
├─ gnn/ R-GCN dataset / features / model / training loop
├─ retrieval/ BM25, dense (BGE), PPR, GNN re-ranker, share one Retriever protocol
├─ reader/ Ollama-backed reader + mock reader for GPU-less testing
└─ xai/ ERASER, GNNExplainer, reasoning paths, attribution agreement,
XAI payload builder consumed by the API
scripts/ download / KG build / bridge graph / GNN training / eval / XAI demo
configs/ one YAML per run, baselines + the main R-GCN system
outputs/xai_cache/ pre-computed payloads powering the lightweight demo (tracked in git)
Report.pdf full write-up: methodology, results, ablations, XAI evaluation

Tech stack

Layer	Tool
Entity linking	ReFinED (Wikipedia QIDs)
Embeddings	BGE-large-en-v1.5
Knowledge graph	NetworkX (in-memory, no external DB)
Re-ranker	R-GCN (torch_geometric)
Retriever-side XAI	GNNExplainer + reasoning-path extraction
Reader-side XAI	ERASER (sufficiency / comprehensiveness / necessity)
LLM reader	Qwen (served locally via Ollama)
Backend	FastAPI
Frontend	Vanilla JS + D3 v7 (force-directed graph)

---

Built as the final project for Current Trends in Machine Learning. See Report.pdf for the full academic write-up.