In large-scale maintenance organizations, identifying subject matter experts and managing communications across complex entities relationships poses significant challenges - including information overload and longer response times - that traditional communication approaches fail to address efectively. We propose a novel framework that combines RDF graph databases with LLMs to process natural language queries for precise audience targeting, while providing transparent reasoning through a planning-orchestration architecture. Our solution enables communication owners to formulate intuitive queries combining concepts such as equipment, manufacturers, maintenance engineers, and facilities, delivering explainable results that maintain trust in the system while improving communication eficiency across the organization.
In this paper, we propose a novel framework that addresses these challenges. Our solution empowers
communication owners to find any audience groups relevant to their maintenance issue through simple,
natural language queries, while providing transparent reasoning for the retrieved results. Our framework
eliminates the need for complex database operations or specialized query languages by integrating
RDF graph databases with full-text search capabilities, enabling eficient connections between user
terms and known entities, as well as facilitating complex queries across various entity relationships
(e.g., equipment, people, and sites). To bridge the gap between natural language and structured queries,
we implement dual LLM-driven workflows that transform user inputs into graph queries. Additionally,
we adopt a planning-orchestration architecture [
The Figure 1 illustrates the system’s architecture of our solution for Amazon RME is composed of four main components: Customer UI, AI Engines, Knowledge Base, and the proposed framework. We briefly describe the first three components before detailing the proposed framework.
The Customer UI provides three core functionalities: natural language query processing, construction
of formal queries, and visualization of results with reasoning explanations. The AI Engines infrastructure
layer leverages foundation models available at Amazon Bedrock (Nova Pro [
Two key aspects of translating Natural Language (NL) to any query language are interpreting and understanding the key terms – NL representations of existing entities contained in the NL statement – and the relationships among them. The proposed framework relies on two architectural components responsible for these tasks. The Named Entity Recognition (NER) workflow component identifies the key terms and aligns them with their corresponding existing entities. The Formal Query Formulation (FQF) workflow formally establishes the relationships among these entities as expressed in the NL statement.
For the NER workflow, we adopt a prompt-based approach for identifying key terms using Amazon Nova Pro as the foundation model. The prompt uses few-shot examples and specialized tags to mark the key terms to be processed at a later stage, and requests the model to output the original statement augmented with the marked key terms. This approach incorporates several lessons learned in recognizing named entities using LLMs. We highlight two key findings: 1) constraining the key terms search space helps avoid hallucinations, and 2) using consistent markers ensures output reliability. After obtaining the initial statement with marked key terms, we parse it using a regex expression to extract the terms. For each term, we then proceed to find the best matching entity in the Knowledge Base ( i.e., the top-ranked entity label with the highest text matching score).
The FQF workflow follows a similar architecture, using Nova Pro as its based model and few-shot
examples prompt to express the natural language statement in a formal logical formulation. Diferently
from the NER workflow, in this workflow, we use a structured output to retrieve the answer. The
output comprises of the formulation of the statement as Boolean algebra expression and a reasoning
breakdown of the formulation process. The latter output was a key finding during our research work,
where we discovered that a Chain-of-Thought [
This expression is then converted to JSON Visual Query Language (jVQL) – to visually express the query in the UI – and into SPARQL – to query Amazon Neptune and retrieve the corresponding data.
While LLMs can efectively process natural language, their interpretations might not always align with the users’ intent. For instance, in a query to “find all technicians working with vendor X in LATAM ”, users need to verify that the system correctly identified each entity ( technicians as job title, X as vendor, LATAM as region). In RME’s complex structure, even simple queries like “managers at European sites” require clarity on whether the system includes only managers or also senior leadership positions. Such validation becomes crucial in high-stakes environments where communications must reach the right audience in a timely manner, particularly for operational updates, network actions and safety-related messages.
To address this need for validation, our framework implements an explainability feature based on
multi-step reasoning [
This paper presented a novel framework that addresses audience targeting and communication challenges in large-scale maintenance organizations. By combining RDF graph databases with LLM-driven workflows, our solution enables natural language querying while maintaining transparency through explainable AI techniques. The framework bridges the gap between user intent and formal queries, demonstrating how enterprises can leverage AI capabilities without compromising explainability in critical operational communications. Future work could extend this approach to other domains and explore additional explainability techniques.
During the preparation of this work, the author(s) used Claude family models and Amazon Nova family models to grammar and spelling check. After using these tool(s)/service(s), the author(s) reviewed and edited the content as needed and take(s) full responsibility for the publication’s content.