Automap: An Automated Knowledge Graph Generation System Based on Multi-Agent Architecture

Core Overview of Automap

Automap is an open-source project maintained by ProyectoPIONERA (GitHub link: https://github.com/ProyectoPIONERA/automap), designed to address the complexity and time-consuming nature of traditional knowledge graph construction. Based on a multi-agent architecture, it uses large language models (LLMs) and the LangGraph framework to automate the entire process from CSV data to knowledge graphs, including CSV schema analysis, ontology search, competency question generation, iterative optimization of YARRRML mappings, and knowledge graph materialization.

Challenges of Traditional Construction

In the data-driven era, converting structured data into semantic knowledge graphs requires domain experts to manually write mapping rules, define ontology relationships, and perform multiple rounds of verification, making the process complex and time-consuming.

The Emergence of Automap

Automap emerged as a solution, transforming manually dependent tasks into standardized processes through an automated agent pipeline, significantly improving efficiency.

Decentralized Multi-Agent Architecture

Automap adopts a decentralized multi-agent architecture, breaking down KG construction into key steps such as schema analysis, ontology search, semantic mapping, competency question generation, YARRRML mapping generation, and verification. Each step is handled by a dedicated agent, with collaborative work driven by a state machine.

Advantages of Modular Design

The modular design enhances system maintainability; agents communicate via well-defined interfaces to ensure data consistency and integrity, and each component can be independently optimized and extended.

Collaborative Generation by Three Agents

YARRRML mapping generation is collaboratively completed by three agents: the Prefix Agent (manages namespace prefixes), the Entity Agent (maps data columns to ontology classes), and the Relationship Agent (establishes entity associations).

Performance Optimization

Parallel execution of the Prefix and Entity Agents reduces time, while the KV cache mechanism reuses prefix declarations, minimizing redundant generation and lowering computational overhead for large datasets.

Competency Question-Driven Verification

Automatically generated or user-defined competency questions are converted into SPARQL ASK queries and verified via pyoxigraph in-memory storage; failure triggers re-generation of mappings.

Multi-Level SHACL Verification

A three-level strategy generates ontology-derived shape constraints: Astrea REST API → local rdflib → structural correctness shapes, maintaining verification capabilities even under network restrictions.

Self-Correction Mechanism

It supports up to 10 syntax retries and 6 logical retries; the schema alignment module automatically detects and fixes missing columns, prevents broken mappings, and reduces manual intervention.

Technology Stack

It uses LangGraph for workflow orchestration and pyoxigraph for RDF storage and querying; supports command-line/environment variable configuration, and native Docker deployment simplifies the process; terminal observability displays agent status and reasoning processes in real time.

Application Scenarios

It helps enterprises/research institutions quickly convert CSV data into queryable and inferable knowledge graphs; the multi-agent architecture provides a reference paradigm for other automated data processing tasks.

Project Value

Automap demonstrates the potential of LLMs and multi-agent architectures in automating complex data processing, significantly improving KG construction efficiency while maintaining high quality.

Future Directions

As LLM capabilities improve and multi-agent collaboration matures, Automap will play an important role in scenarios such as enterprise knowledge management, scientific research data integration, and open data publishing, making it an ideal solution for teams lacking professional RDF engineers.