We present metis, an AI agent platform that assists users with semantic modeling, search, and discovery across knowledge graphs. It combines the conversational capabilities of Large Language Models with the precision of semantic knowledge graphs, integrating seamlessly with the functionalities of the metaphactory platform, such as semantic search and visualization. metis delivers AI agents that provide generative power, semantic precision & contextual, explainable insights.
At the heart of metis lies a semantic model, the trust and context layer, that informs and scopes every AI interaction. This model guides the agents, ensuring not only relevancy of the results but also explainability and alignment with enterprise semantics.
The generic architecture of metis 1, is structured into three main layers: the User Interface, the metis agent layer, and the underlying metaphactory knowledge graph platform. The interaction process begins when a user sends a Request through the User Interface, which can include a message and files in natural language. This request is ingested by the Conversational Agent in the metis layer. The agent’s Context Manager enriches the request by combining it with predefined instructions, dialogue history, and the relevant semantic model, which is retrieved from metaphactory. This enriched prompt is then passed to an LLM-powered Orchestrator that plans and executes a workflow to fulfill the request. If necessary, the Orchestrator performs tool calls, sending specific input parameters to one or more available tools. These tools are built on top of metaphactory APIs and interact with the data layer, which is managed by a Repository Manager. The tool returns a Tool Result, consisting of either the requested output or an error message. The Orchestrator may iteratively call tools until the necessary information is retrieved and uses the results for Response Formulation. Finally, the resulting Response, which can be a combination of natural language and various visualizations, such as tables or charts, is sent back to the User Interface.
Based on the generic architecture of the metis platform, diferent kind of agents can be configured. We here focus on two types of agents equipped with a set of tools to support specific tasks: a Semantic Modeling Assistant to collaboratively develop and refine ontologies 2, and a Search & Discovery Agent with access to the knowledge graph to resolve information needs of its users (see the architecture in Appendix A).
Semantic Modeling Assistant. The Semantic Modeling Assistant is designed to support users
in the collaborative development and refinement of ontologies. This agent guides users through a
structured modeling process based on the established Linked Open Terms (LOT) methodology3 [
The initial steps of the LOT methodology, such as Ontology Requirements Specification , are driven by metis. The assistant helps users formulate competency questions and define use cases, which are stored as explicit modeling artifacts in the ontology. For the Ontology Implementation phase, metis assists with the crucial sub-activities of reuse, conceptualization, and evaluation. It facilitates ontology reuse by treating the metaphacts Ontology Repository5 as a primary resource hub. The agent distinguishes between conceptualization and encoding, presenting modeling concepts like classes and properties in a human-interpretable format while handling the formal encoding in the background. Ontology evaluation is supported through both formal SPARQL-based checks and by applying the LLM’s reasoning capabilities to critique modeling artifacts. The subsequent steps of ontology publication and maintenance are handled by the robust, built-in functionalities of the metaphactory platform, such as automated documentation generation and editorial workflows. The assistant’s architecture (Figure ??) features pre-configured tools tightly integrated with metaphactory to create and persist these modeling artifacts. Search & Discovery Agent. The Search & Discovery Agent enables users to perform intuitive exploration and targeted search across knowledge graphs using natural language. This agent is highly customizable; users can adapt its instructions, enable or disable specific tools, and tune LLM parameters to fit their domain and use case. Its architecture (cf. Figure A) gives it access to three categories of tools: • Structured Querying tools perform tasks like entity lookup and linking, as well as translation of natural language to SPARQL for precise data aggregation and analysis. • Unstructured Retrieval tools leverage metaphactory’s retrieval functionality to find information in textual content that cannot be easily mapped to a graph pattern. • Visualization tools render the retrieved data in various formats, such as tables, charts, and graphs, to present it to the user in the most appropriate way. 2See the architecture in the platform documentation: https://help.metaphacts.com/resource/Help:SemanticModelingAssistant. 3https://lot.linkeddata.es/ 4https://metaphacts.com/images/PDFs/metaphacts_Semantic_Modeling_Guidelines_2.0.pdf 5https://ontologies.metaphacts.com/
Through this combination of tools, the agent supports multi-turn dialogues, allowing users to iteratively refine their information needs and build on previous results to gain deep, context-aware insights from their data.
creating ontologies from scratch by guiding the user through the LOT methodology or refining and evaluation existing ontologies. The following scenarios are demonstrated in our supplemental material and partially shown in Figure 1: 1. From specification to implementation . In this scenario the ontology is developed from scratch. metis guides the user through the ontology requirements specification, ontology formalization, ontology implementation, ontology evaluation, and ontology publication activities. The user can provided additional input by uploading files to support the ontology development process. 2. Reusing ontological resources. Ontology developers both reuse and re-engineer ontological resources. metis identifies existing ontology resources that can be reused facilitating the integration, adaptation, and extension of these resources to meet specific domain requirements. 3. Evaluate and restructure ontological resources. Ontology engineers systematically evaluate and restructure (e.g., modularize, prune, extend, specialize) ontological resources. metis incorporates requirements specification and built-in evaluation mechanisms to assist in assessing and refining ontological resources.
Figure 1 shows an example how metis facilitates the construction of a Personal Information ontology by building a controlled vocabulary for the Topic class, which is derived from a spreadsheet of predefined topics. During the modeling process metis ensures that each ontology element is explicitly defined with appropriate metadata. In the example the Task class is enriched with a human-readable label and description. Furthermore, metis tracks semantic relations and links the Task class to the Resource class through the newly defined relation uses.
Agent across several knowledge graphs, covering diferent domains. Notably, SemOpenAlex 6 [
1. Natural language request. The user formulates an information need to retrieve the number of publications addressing Alzheimer Disease from the year 2014 to 2024. 2. Transparent Multi-step Reasoning & Tool Orchestration. The agent transparently uses a sequence of tools to fulfill the user request. First the entity Alzheimer Disease is identified in the 6https://semopenalex.org 7https://metaphacts.com/solutions/products/dimensions-knowledge-graph request and linked to its corresponding IRI in the Knowledge Graph. Then a structured SPARQL query is generated and executed to retrieve the number of relevant publications. 3. Result visualization. After the data is retrieved by executing the SPARQL query the agent uses its visualization capabilities to present the number of publications per year as a bar chart. 4. Explanations. Along with the presented results, explanations about how the result was generated are provided to the user. Upon user request, detailed information about all tools used and the entire data lineage are presented. 5. Multi turn dialogue. The user sends a follow-up request to rank all publications from the previous result by their altmetric score8. 6. Context-aware answer. The agent retrieves the 10 Alzheimer Disease publication with the highest altmetric score from 2014 to 2024 and presents the results in a table for clear visualization.
PLIseUuCw9-OAlmJxFDwNfmsRVPVGRHzNN. https://metaphacts.com/solutions/products/metis. via https://www.youtube.com/playlist?list= Further resources on metis are available at