Advanced-Graph-RAG

An Advanced Retrieval-Augmented Generation (RAG) system powered by a Knowledge Graph + Neo4j + Cypher-based retrieval + Review loop architecture.

This project:

• Converts unstructured data into a Knowledge Graph using a Cloud LLM
• Loads the graph into Neo4j AuraDB
• Generates and executes Cypher queries using an LLM
• Applies fallback strategies
• Review & validation loop
• Combines:
  • User query entity extraction
  • Neighbour node traversal
  • Vector similarity search
• ✅ Workflow orchestration using LangGraph
• ✅ Tracing & observability using LangSmith

---

Architecture Flow

The entire pipeline is implemented as a structured graph workflow using LangGraph.

User Query
→ Cypher Generation (LLM)
→ Cypher Execution (Neo4j)
→ Review
    • If valid → Final Answer
    • If invalid → Graph RAG → Final Answer
→ If Cypher fails → Graph RAG → Final Answer

---

Setup

1️⃣ Install Dependencies

pip install -r requirements.txt

2️⃣ Set Environment Variables

Create a .env file in the root directory:

# Cloud LLM API Key
GROQ_API_KEY=your_groq_api_key  # Provide your cloud provider's API key

# Neo4j Credentials
NEO4J_URI=neo4j+s://<...>.neo4j.io
NEO4J_USERNAME=neo4j
NEO4J_PASSWORD=your_password

# LangSmith Tracing (Optional)
LANGSMITH_API_KEY=your_langsmith_api_key

---

Steps to Run

1️⃣ Create Knowledge Graph

Run:

python 1_knowledge_graph_conversion_using_cloud_llm.py

This generates: graph.pkl

2️⃣ Load Graph into Neo4j Aura DB

Open and run:

2_load_graph_to_neo4j.ipynb

This loads nodes and relationships into Neo4j.

3️⃣ Run Advanced Graph RAG

Open and run:

3_advanced_graph_rag.ipynb

---

🔍 Tracing & Observability (LangSmith)

Tracing is powered by LangSmith, enabling full visibility into the LangGraph workflow.

With tracing enabled, you can:

• 🔎 See the exact execution path (Cypher → Review → Fallback)
• 📊 Track token usage per LLM call
• ⏱ Measure latency at each step
• 🔁 Inspect retries and validation loops

🔬 LangSmith Trace Screenshot

---

Notes

• graph.pkl is required for Step 2 and Step 3.
• Ensure Neo4j is running before executing notebooks.
• The current implementation uses OpenAI GPT-OSS-20B via Groq Provider. You can replace it with any LLM provider (OpenAI, Azure, Anthropic, etc.) as per your requirements.

---