🎓 Agentic Graph Advisor

A Graph-First, Multi-Agent Academic Question Answering System

Agentic Graph Advisor is a graph-based, multi-agent question answering system built using Neo4j and local LLMs (Ollama) to answer complex academic advising questions about courses, prerequisites, and degree programs.

Unlike traditional vector-based RAG systems, this project treats the knowledge graph as the source of truth and uses LLMs only for planning, orchestration, and explanation, not fact generation.

---

🚀 What This Project Solves

Academic advising questions are structural, not semantic.

Questions like:

• What do I need before I can take Course X?
• Can I take Course X if I’ve completed Courses Y and Z?
• What courses does Course X unlock?
• What are the core requirements for a degree program?

cannot be answered reliably using embeddings or similarity search alone.

This project solves that by:

• explicitly modeling relationships as a knowledge graph
• performing multi-hop graph traversal
• using deterministic logic where correctness matters
• orchestrating reasoning through LLM-based agents

---

🧠 Core Design Principles

• Graph > Vector Search for structural reasoning
• LLMs are not the source of truth
• Deterministic logic for eligibility checks
• Agentic orchestration with guardrails
• Explainable and debuggable outputs

---

🏗️ System Architecture

User Question
     ↓
Planner Agent
(intent classification + entity extraction)
     ↓
Cypher Agent
(read-only Neo4j queries)
     ↓
Neo4j Knowledge Graph
(courses, prerequisites, programs)
     ↓
Answer Agent
(graph-grounded explanation)
     ↓
Verifier Agent
(completeness & hallucination check)

A Streamlit UI exposes each step of this pipeline for transparency.

---

🧩 Knowledge Graph Schema

Nodes

Course

Course {
  course_code,
  title,
  credits,
  level,
  description,
  department
}

Program

Program {
  program_id,
  program_name,
  degree_type,
  department
}

Relationships

(Course)-[:PREREQUISITE]->(Course)
(Program)-[:REQUIRES {requirement_type}]->(Course)

---

🤖 Agent Responsibilities

1️⃣ Planner Agent

• Classifies user intent (e.g. all_prereqs, eligibility_check)
• Extracts course codes and program IDs
• Uses LLM reasoning + rule-based overrides for reliability

2️⃣ Cypher Agent

• Generates read-only Cypher queries
• Uses deterministic templates for known intents
• Falls back to LLM generation only when necessary

3️⃣ Answer Agent

• Converts graph results into natural language answers
• Uses deterministic formatting for structured outputs
• Never outputs Cypher, code, or hallucinated facts

4️⃣ Verifier Agent

• Ensures answers are supported by graph results
• Detects incomplete responses
• Triggers a follow-up query when needed
• Hardened against malformed LLM output

---

🔁 Agentic Reasoning Loop

The system supports a multi-step agent loop:

1. Initial query execution
2. Verifier evaluates completeness
3. Optional follow-up query with verifier guidance

---

🧪 Supported Question Types

• Course descriptions
• Direct prerequisites (1-hop)
• All prerequisites (transitive closure)
• Shortest prerequisite path
• Program core vs elective requirements
• Eligibility checks (set-difference logic)
• Forward dependencies (“what does this course unlock?”)

---

📊 Example

Question

What do I need before I can take DMS440?

Answer

All prerequisites (transitive closure):

- CSE115: Intro to Programming
- CSE116: Object-Oriented Programming
- CSE191: Intro to Data Structures
- CSE250: Data Structures and Algorithms
- CSE305: Database Systems
- CSE404: Machine Learning Foundations
- CSE440: Graph Databases and Neo4j
- DMS201: Intro to Data Science
- DMS301: Data Mining
- DMS401: Applied Machine Learning
- DMS430: Agentic AI Systems
- MTH101: Calculus I
- MTH102: Calculus II
- MTH201: Linear Algebra
- MTH241: Discrete Mathematics
- MTH301: Probability
- MTH302: Statistics

---

🖥️ Streamlit Interface

The Streamlit UI displays:

• User question
• Planner output (intent + entities)
• Generated Cypher query
• Neo4j result preview
• Final answer
• Verifier verdict

---

🛠️ Tech Stack

• Neo4j – graph database
• Python – core application logic
• Ollama – local LLM inference
• Pydantic – schema validation
• Streamlit – interactive UI
• Docker – Neo4j deployment

---

📁 Repository Structure

agentic-graph-advisor/
├── data/
├── src/
│   ├── agents/
│   ├── db/
│   ├── llm/
│   ├── rag/
│   ├── main.py
│   └── ui_streamlit.py
├── docker-compose.yml
├── requirements.txt
└── README.md

---

🎯 What This Project Demonstrates

• Knowledge graph modeling
• Agentic AI system design
• Safe LLM orchestration
• Multi-hop reasoning
• Production-style defensive engineering
• Knowing when not to use vector RAG

---