KG-RAG AI Tutor

Adaptive AI tutoring system based on Knowledge Graph-enhanced Retrieval-Augmented Generation (KG-RAG).

Paper: How to build an adaptive AI tutor for any course using KG-RAG (Dong et al., ICEIT 2025, IEEE).

Architecture

Student query
    │
    ├─► Chat history reuse (MiniLM / embedding similarity ≥ 0.85)
    │
    ├─► RAG: vector similarity over course chunks
    │
    └─► KGR: knowledge-graph traversal for expanded context
              │
              ▼
        DeepSeek-V3 response synthesis

Pipeline (paper Section III)

1. KG construction — chunk course PDFs (~1,000 tokens), extract [Entity, Relation, Entity] triples with DeepSeek-V3, save to data/knowledge_graph.xlsx.
2. Knowledge-Guided Retrieval — embed the query, match KG nodes, traverse neighbors for interconnected context.
3. Response synthesis — combine similarity context + KG-expanded context and generate the tutoring answer.

Quick start

git clone https://github.com/098765d/AI_Tutor.git
cd AI_Tutor
pip install -r requirements.txt
cp config.example.yaml config.yaml

Edit config.yaml and add your keys:

llm:
  api_key: ""          # DeepSeek API key

embeddings:
  api_key: ""          # DashScope API key (optional — falls back to local MiniLM)

Run the app:

streamlit run app.py

1. Open the sidebar and click Build / Rebuild Course Index (uses PDFs in pdf/).
2. Ask questions in the chat box.
3. Switch between KG-RAG and Standard RAG to compare modes.

Configuration

Key	Purpose
llm.api_key	DeepSeek-V3 chat completions
embeddings.api_key	Alibaba text-embedding-v2 (paper default)
kg_rag.chunk_size	PDF chunk size in words
kg_rag.kg_hop_depth	Graph traversal depth for KGR
kg_rag.chat_history_threshold	Reuse threshold (default 0.85)

If DashScope key is empty, embeddings automatically fall back to sentence-transformers/all-MiniLM-L6-v2.

Project layout

AI_Tutor/
├── app.py                 # Streamlit UI
├── config.yaml            # API keys (gitignored — copy from config.example.yaml)
├── config.example.yaml
├── pdf/                   # Bundled course materials
├── src/
│   ├── tutor.py           # KG-RAG orchestration
│   ├── knowledge_graph.py # KG build + KGR
│   ├── vector_store.py    # RAG retrieval
│   ├── llm.py             # DeepSeek client
│   ├── embeddings.py      # DashScope / local embeddings
│   ├── documents.py       # PDF loading & chunking
│   ├── chat_history.py    # History reuse (Fig. 6 in paper)
│   └── export.py          # HTML chat export
└── data/                  # Generated index files (gitignored)

References

• ICEIT 2025 paper (IEEE Xplore)
• Assessment quiz used in the study
• DeepSeek API
• DashScope embeddings

License

See LICENSE.