NeuroTIC: An Autonomous Exploration Engine for Reshaping Knowledge Graphs via Multi-Agent Ontology

The NeuroTIC project aims to address issues in traditional knowledge graph construction such as difficulty in scaling, lack of dynamic updates, and insufficient semantic depth. By combining a dual-agent architecture with a graph database (CozoDB) and locally deployed large language models, it autonomously builds deeply interconnected knowledge graphs from a single concept seed. The system focuses on active exploration, driving the transformation of knowledge graphs from static storage to dynamic evolution.

Knowledge graphs are crucial in fields like search engines and recommendation systems, but traditional construction methods face three major challenges: reliance on expert manual construction makes scaling difficult; static graphs cannot adapt to dynamic knowledge changes; and shallow entity relationships lack deep semantic connections. As an active knowledge exploration engine, NeuroTIC autonomously expands knowledge networks from simple concept seeds to address these pain points.

NeuroTIC's tech stack includes three key components:

1. CozoDB Graph Database: Optimizes complex graph queries and reasoning, supporting recursive queries, path analysis, and Datalog language reasoning;
2. Local Large Language Model Interface: Deployed via the Ollama framework to ensure data privacy and control costs, responsible for tasks like entity recognition and relation extraction;
3. Dual-Agent Architecture: The Explorer (actively discovers new knowledge) and Integrator (verifies and integrates into the existing graph) work collaboratively—this is the core innovation.

The Explorer agent analyzes boundary nodes of the graph, generates exploration queries, and evaluates their value (relevance, reliability, knowledge gain); the Integrator agent is responsible for quality control, resolving entity disambiguation and relation conflicts, verifying new knowledge through rule-based reasoning and semantic similarity calculation, and initiating a traceability mechanism to assist decision-making when necessary. Together, they form an efficient collaborative closed loop.

The system operation consists of four stages:

1. Initialization: Parse input (concepts, text, or structured data) to create an initial knowledge core;
2. Exploration: The Explorer continuously proposes expansion directions and generates candidate knowledge;
3. Integration: The Integrator verifies and integrates new knowledge, maintaining graph consistency;
4. Convergence: When diminishing returns occur or a preset boundary is reached, a complete knowledge structure is formed.

NeuroTIC is suitable for multiple scenarios:

• Academic Research: Quickly build domain knowledge overviews and assist in exploring research topics;
• Enterprise Knowledge Management: Extract knowledge from scattered documents, build a unified enterprise graph, and support training and knowledge sharing;
• Content Recommendation: Deep graphs capture detailed user interests, enabling precise and explainable recommendations.

Current challenges include knowledge quality control (autonomous exploration may introduce errors) and computational efficiency (performance optimization of graph queries and model reasoning). Future directions: introduce domain-specific agents; develop human-machine collaboration interfaces; establish automatic knowledge maintenance mechanisms to handle outdated information.

NeuroTIC represents a paradigm shift in knowledge graph construction from manual-dominated static maintenance to autonomous dynamic evolution. By combining graph databases, large language models, and multi-agent collaboration, it opens a new path for knowledge management and is expected to help humans organize and utilize knowledge assets in more scenarios in the future.