HA-MOACO: A Structure-Aware Graph-RAG System for Small Language Models

HA-MOACO is a structure-aware Graph-RAG system designed for small language models (SLMs). Its core goal is to address the pain points of SLMs, such as hallucinations and insufficient multi-step reasoning capabilities, and provide an efficient and reliable solution for professional domain applications. By modeling knowledge with graph structures, the system supports deep reasoning, enabling SLMs to achieve professional task performance close to that of large models while retaining their lightweight and efficient advantages.

Large language models perform well in general tasks, but their deployment in professional domains faces challenges such as high costs and large latency. Small language models (SLMs) are lightweight and efficient, but they are prone to hallucinations (generating incorrect content) and have weak multi-step reasoning capabilities. Traditional Retrieval-Augmented Generation (RAG) technology treats knowledge as flat text fragments, ignoring structured relationships and limiting deep reasoning capabilities.

The core architecture of HA-MOACO consists of four key components:

1. Knowledge Graph Construction Module: Converts unstructured documents into structured graph representations, extracting entities, relationships, and attributes
2. Structure-Aware Retriever: Retrieves relevant text fragments while obtaining associated graph structure subgraphs
3. Multi-Step Reasoning Engine: Uses graph traversal algorithms to support chain reasoning and gradually construct answers
4. Hallucination Detection and Correction Layer: Identifies and corrects potential erroneous generations through cross-validation and consistency checks

HA-MOACO's optimization measures for SLMs include:

• Context Compression: Uses graph structure summarization technology to deliver more information within a limited context window
• Reasoning Guidance: Uses graph structures to guide the model to think in a logical order, making up for SLMs' insufficient reasoning capabilities
• Domain Specialization: Supports customization of knowledge graphs for specific professional domains such as healthcare, law, and finance

HA-MOACO is suitable for the following scenarios:

• Enterprise Knowledge Management: Deploy small models locally to process internal documents and protect data privacy
• Professional Consulting Systems: Provide reliable Q&A services for fields such as law, healthcare, and engineering
• Edge Intelligent Devices: Implement trustworthy AI-assisted decision-making in resource-constrained environments

HA-MOACO is open-sourced to provide a complete Graph-RAG implementation reference, proving that through architectural design, SLMs can achieve professional task performance close to that of large models while maintaining cost-efficiency advantages. The codebase includes knowledge graph construction tools, retrieval optimization algorithms, and evaluation benchmarks, providing a starting point for researchers and developers. The structured RAG direction is expected to become an industry-standard practice for trustworthy AI.