GraphRAG-MarineMind

A lightweight GraphRAG-based ecological reasoning system for analyzing marine ecosystem dynamics using structured knowledge graphs, statistical ecological indicators, and local LLM reasoning.

---

Overview

GraphRAG-MarineMind is a multi-layer retrieval-augmented generation system designed for marine ecology analysis, combining:

• Knowledge Graph (structured ecological relationships)
• Statistical Layer (ecological trends & drivers)
• Local LLM reasoning (Phi-3 via Ollama)

The system enables grounded ecological Q&A using structured environmental knowledge.

---

System Architecture

User Query ↓ Graph Retrieval (Tier 1) ↓ Statistical Lookup (Tier 2) ↓ Context Builder (Fusion Layer) ↓ Phi-3 Mini (Local LLM) ↓ Scientific Answer

---

System Components

Tier 1 — Knowledge Graph

Captures ecological relationships such as:

• Sentinel-2 → detects → seagrass_extent
• Seagrass_loss → reduces → biodiversity
• Satellite_imagery → estimates → benthic_cover

---

Tier 2 — Ecological Statistics Layer

Represents ecological trends and drivers:

• seagrass_extent → declining
• drivers: temperature increase, eutrophication, human activity
• indicators: NDVI change, habitat fragmentation

---

LLM Layer (Phi-3 via Ollama)

Used to:

• interpret retrieved ecological knowledge
• generate structured scientific explanations
• avoid hallucination using context grounding

---

Features

• Fully local LLM inference (Ollama + Phi-3)
• Graph-based ecological reasoning
• Statistical ecological indicator layer
• Context fusion between structured data sources
• Lightweight CLI query engine

---

Example Query

Input:

What methods estimate seagrass extent?

Output (example):

• Sentinel-2 satellite imagery
• Multitemporal remote sensing analysis
• Benthic cover estimation from satellite data

---

Tech Stack

• Python
• Ollama (Phi-3 Mini)
• JSON-based Knowledge Graph
• Rule-based statistical inference layer

---

Purpose

This project demonstrates:

• Hybrid Graph + Statistical RAG design
• Ecological knowledge representation
• Local LLM integration for scientific reasoning
• Early-stage research prototype for marine AI systems

---

Future Improvements

• Vector-based document retrieval (Tier 3)
• FastAPI deployment
• Evaluation framework for retrieval accuracy
• Integration with real Sentinel-2 datasets

---

Status

This is an active research prototype, designed for experimentation in ecological AI systems and GraphRAG architectures.

Architecture

                    ┌────────────────────────────┐
                    │        User Query          │
                    └────────────┬───────────────┘
                                 │
                                 ▼
                    ┌────────────────────────────┐
                    │      Query Router          │
                    │ (Intent Detection Layer)   │
                    └────────────┬───────────────┘
                                 │
        ┌────────────────────────┼────────────────────────┐
        │                        │                        │
        ▼                        ▼                        ▼
    ┌────────────────────┐   ┌────────────────────┐   ┌────────────────────┐
    │   Tier 1: Graph    │   │   Tier 2: Stats    │   │  Tier 3:Documents
    │  Knowledge Engine  │   │  Statistical Layer │   │  (Future Vector DB)│
    │                    │   │                    │   │                    │
    │ - Relationships    │   │ - Trends           │   │ - PDFs             │
    │ - Methods          │   │ - Drivers          │   │ - Literature       │
    │ - Ecological Links │   │ - Indicators       │   │ - Embeddings       │
    └─────────┬──────────┘   └─────────┬──────────┘   └─────────┬──────────┘
              │                        │                        │
              └───────────────┬────────┴──────────────┬─────────┘
                              ▼                       ▼
                    ┌────────────────────────────────────┐
                    │     Context Aggregation Layer      │
                    │ (Unified Ecological Knowledge)     │
                    └────────────────────────────────────┘
                                       │
                                       ▼
                    ┌────────────────────────────────────┐
                    │     Local LLM (Phi-3 / Ollama)     │
                    │  Grounded Scientific Explanation   │
                    └────────────────────────────────────┘
                                       │
                                       ▼
                        ┌────────────────────────────┐
                        │       Final Answer         │
                        │  (Structured + Grounded)   │
                        └────────────────────────────┘