Implementing a Knowledge Management System with GraphRAG: A Physical Internet Example

Abstract

The rapid expansion and interdisciplinary nature of Physical Internet (PI) research have resulted in fragmented knowledge, limiting the ability of stakeholders to identify emerging trends, actionable insights and genuine research gaps. This study introduces a novel knowledge management approach that uses Graph Retrieval-Augmented Generation (GraphRAG) to systematically organize and integrate PI-related literature. A comprehensive knowledge graph was constructed by extracting and semantically modeling entities and relationships from 2835 academic papers, conference proceedings and international roadmaps. The developed system incorporates fuzzy semantic search and multiple retrieval strategies, including local, global and hybrid approaches, enabling nuanced, context-aware access to information. Stakeholder-specific prompts, tailored to the needs of industry, government and academia, demonstrate how GraphRAG can support the discovery of business model innovations, policy design and underexplored research areas. A comparative evaluation using cosine similarity and BERTScore confirms that graph-based strategies outperform standard LLM retrieval in providing relevant and comprehensive answers while also revealing connections that would be missed in manual reviews. The results demonstrate that the proposed GraphRAG model is a scalable and extensible framework for addressing knowledge gaps and promoting collaboration in PI research synthesis for sustainable logistics. The model also shows promise for application in other complex domains.

1. Introduction

The logistics industry is facing increasing challenges, particularly in addressing its significant contribution to global greenhouse gas (GHG) emissions. These emissions are often exacerbated by operational inefficiencies such as suboptimal routing, inefficient warehouse layouts, and poor inventory management [1]. In response to these challenges, the concept of the PI has emerged as a transformative framework. Introduced by Montreuil et al. [2], PI reimagines global logistics networks by applying design principles from the digital internet to the transportation and distribution of physical goods. By leveraging information and communication technologies (ICT) and standardized logistics infrastructures, PI aims to enhance efficiency, flexibility, and sustainability in freight transportation systems. Over the past decade, international conferences like International Physical Internet Conference (IPIC) and government-led initiatives in regions such as Europe and Japan have accelerated discussions and roadmap developments aimed at realizing the PI vision.

Despite its potential, advancing research in the PI domain requires systematic methods for organizing and integrating diverse knowledge sources. The growing body of literature on PI—spanning academic research, government reports, and industry white papers—presents both opportunities and challenges for researchers seeking to identify trends, gaps, and actionable insights. Addressing this complexity calls for innovative tools that can manage and synthesize large volumes of information effectively. However, manual review articles on PI have reviewed less than 700 articles, e.g., 663 articles in [3], yet there are more than 2000 scientific papers already published.

Despite there being several systematic reviews and bibliometric studies, the synthesis of current knowledge in the PI field is severely hindered by fragmentation, methodological heterogeneity and the rate at which new research is produced. As noted by Tallaki et al., previous manual reviews only analyzed a subset of available publications (less than 700), leaving significant gaps due to resource constraints and limited semantic scope [4]. To address this challenge, we present the Graph Retrieval-Augmented Generation (GraphRAG) method. By building a comprehensive, semantically rich knowledge graph, GraphRAG enables information to be integrated and retrieved from thousands of diverse sources. This scalable, flexible framework for knowledge organization fills a critical gap left by traditional review methods.

GraphRAG is an emerging Artificial intelligence (AI)-driven approach that combines natural language understanding with graph-based relational structures to enhance information retrieval and generation. Unlike traditional methods, GraphRAG excels at uncovering relationships among data points, retrieving highly relevant information, and generating contextually rich outputs. These capabilities make it particularly well-suited for addressing complex problems with interconnected data structures, such as those found in PI research.

This study explores the application of GraphRAG as a novel method for advancing knowledge management in the PI domain. Specifically, it constructs a knowledge graph using GPT-4o mini and Neo4j to systematically analyze PI-related literature. By integrating fragmented research findings, visualizing emerging trends, and identifying knowledge gaps, this approach aims to deepen understanding and foster innovation in sustainable logistics systems.

This study makes three key contributions to the literature on knowledge management and sustainable logistics innovation. First, it represents the first known application of GraphRAG to construct a comprehensive knowledge graph for the PI domain. This graph integrates 2835 documents including academic articles, IPIC proceedings, and strategic roadmaps from the EU and Japan, enabling a structured and scalable analysis of emerging trends, research gaps, and innovation opportunities.

Second, it introduces a fuzzy matching mechanism to improve semantic search within the knowledge graph. By applying cosine similarity with a threshold-based retrieval method, the system improves performance in handling ambiguous queries, such as “sustainable logistics policies”, achieving a reported precision of 87%.

Third, the study designs and evaluates stakeholder-specific prompts tailored to the needs of industry practitioners, academic researchers, and policymakers. These prompts are assessed across four retrieval strategies, namely GPT-only, local, global, and hybrid, demonstrating the model’s ability to deliver actionable insights grounded in both node-level detail and community-level context.

The remainder of this paper is structured as follows: Section 2 reviews relevant literature on PI and generative AI technologies like GraphRAG. Section 3 describes the data sources, and Section 4 presents the methodology used to construct the knowledge graph. Section 5 presents results from various search strategies and evaluates their performance. Finally, Section 6 concludes with key findings and identifies future research directions.

2. Literature Review

2.1. Physical Internet

The concept of PI was first introduced by Montreuil et al. as a transformative framework for sustainable logistics [2]. By applying the design principles of the digital Internet, PI aims to create an interoperable, collaborative and efficient freight transport system. The core idea is to route modular containers through transit hubs, called PI hubs, to maximize consolidation opportunities and minimize inefficiencies in transportation networks [4]. Over the past decade, PI has received considerable attention from researchers and practitioners alike for its potential to address critical challenges such as greenhouse gas emissions and resource inefficiencies in global logistics systems [5].

The growing body of literature on PI reflects its increasing importance. The growing academic interest in the PI is reflected in the increasing volume of peer-reviewed publications and systematic literature reviews over the past decade, as documented by recent studies [6,7,8]. Foundational works by Ballot et al., Aron and Sgarbossa, and Wu et al. have systematically reviewed PI research, summarized its progress, and identified future research directions.

Existing review papers on PI have only surveyed a limited number of studies. For example, Tallaki et al. surveyed 663 publications using bibliometric methods [4]. Nguyen et al. employed a bibliometric knowledge mapping approach to assess research landscape in PI, digital twin and supply chain management areas [9]. In this study, 467 papers were reviewed. These examples highlight the need for a more comprehensive review to a larger or complete set of relevant literature to achieve a thorough and holistic synthesis of the current state of the art.

Despite these efforts, the rapid expansion of PI-related research has created a fragmented knowledge base. This fragmentation stems not merely from an information overload due to the sheer increase in the number of publications but from the interplay of multiple factors. First, methodological inconsistencies are a key factor. Existing review studies exhibit significant differences in the literature databases they utilize, their analytical methodologies (e.g., systematic reviews, bibliometric analysis, content analysis), and the metrics they employ. Consequently, the comparability of research findings is low, and the insights gained tend to be partial and limited. Second, the dispersion and diversification of research subjects also contribute to fragmentation. PI research has developed while intersecting with diverse fields such as logistics system design, information and communication technology, sustainability, policy and regulation, and business models. Consequently, individual researchers focus on different themes and analytical units, hindering sufficient integration of cross-disciplinary knowledge. Third, insufficient theoretical integration is a significant factor. Much existing PI research relies on individual case studies or technical approaches, with insufficient systematization of theoretical models or conceptual frameworks as an academic foundation. This makes it difficult to understand research findings in relation to each other, leading to a tendency for insights to remain scattered.

Considering these challenges, this study aims to reconstruct the overall landscape of PI research from a more comprehensive and integrated perspective, clarifying the structure and developmental trends of the research field.

2.2. Generative AI

AI has attracted significant attention in a variety of fields and industries. Beginning with Alan Turing’s foundational ideas in the 1950s, AI has undergone multiple phases of development, including periods of stagnation famously referred to as AI winters [10]. Generative AI technologies have revolutionized information processing across various domains. Since the release of OpenAI’s ChatGPT-3.5 in 2022, large language models (LLMs) have exhibited exceptional capabilities to understand context and generate text that closely resembles human language [11]. However, LLMs face limitations such as knowledge rigidity, reliance on static training data, and susceptibility to generating inaccurate or misleading outputs when confronted with unfamiliarity [12,13,14]. These challenges highlight the need for hybrid approaches that combine LLMs with external databases to enhance reliability and contextual accuracy.

2.3. GraphRAG

GraphRAG is an emerging technology that addresses the limitations of traditional LLMs by integrating natural language understanding with graph-based relational structures (see Figure 1). Introduced by Microsoft researchers in early 2024, GraphRAG improves information retrieval by linking LLMs to external knowledge graphs, enabling more accurate and contextually relevant responses [15,16]. Unlike similar technologies such as HybridRAG or LightRAG, GraphRAG excels at capturing indirect relationships between data points while maintaining high retrieval accuracy.

This diagram outlines an example where each plot related to PI is derived from the paper authored by Montreuil et al. [2].

Recent studies have begun to explore GraphRAG’s applications in logistics and supply chain management. For example, Li et al. demonstrated its utility in transforming unstructured supplier data into knowledge graphs to improve supplier discovery processes [17]. Similarly, Ojima et al. proposed a GraphRAG-based knowledge management system for automotive failure analysis [18]. In the context of PI, Naganawa and Hirata presented a case study showing how GraphRAG can enhance sustainability efforts by enhancing policy generation [19].

Given its ability to integrate diverse data sources and uncover complex relationships, GraphRAG is a promising tool to address the fragmented knowledge landscape in PI research. This study builds on these advances by using GraphRAG to construct a comprehensive knowledge graph for PI-related literature, enabling systematic analysis of trends, gaps, and opportunities.

Technologies similar to GraphRAG are HybridRAG and LightRAG. We summarize the differences in these technologies in

Table 1. Comparison of GraphRAG, HybridRAG, and LightRAG.

Feature	GraphRAG	HybridRAG	LightRAG
Retrieval Method	Graph-based plus semantic search	Combination of vector search and keyword search	Pure vector-based retrieval
Structure Awareness	Captures direct and indirect relationships via graph topology	Limited. Sometimes captures superficial relevance	Low. Relationships are not structurally modeled
Response Quality	High contextual depth with relational grounding	Moderate. May miss deeper or indirect context	Fast but may include fragmented or shallow responses
Scalability	Moderate. Depends on graph complexity	High. Simple integration of vector and keyword methods	Very high. Minimal processing overhead
Semantic Matching	Strong by fuzzy matching, cosine similarity	Moderate	High when using dense embeddings
Best Suited For	Integrated knowledge management, tailored stakeholder insights	General-purpose retrieval tasks	Quick. High-speed applications with simple queries

.

Overall, hybridRAG combines vector search and keyword search, enabling it to obtain data with strong superficial relevance, but it is difficult to capture indirect relationships. In addition, LightRAG is basically a vector search based on simple similarity and does not consider relationships, so although it is fast in generation, there is a high possibility that fragmentary answers will be mixed in the generated results. Therefore, we considered GraphRAG to be the best choice for this study.

The fragmentation of PI literature is primarily due to the rapid growth in publications, interdisciplinary cross-linkages (e.g., between logistics, ICT and policy) and the lack of a standardized taxonomy or ontology. As topics, terminology and research agendas diverge, siloed knowledge clusters emerge, impeding cumulative learning. Our graph-based approach overcomes this issue by mapping explicit and latent relationships across domains, which traditional manual reviews are unable to achieve on a large scale.

3. Data

To construct a Knowledge Graph (KG) specifically tailored to the PI, this study aggregated data from four key sources (

Table 2. List of Contents.

Category	Year	Number of Documents	Collection Method
Research Paper	2006–2024	2537	Semantic scholar API
Proceedings of IPIC	2016–2024	296	Download from conference repository
PI Roadmap (ALICE)	2020	1	Download from ALICE website
PI Roadmap (Japan)	2022	1	Download from METI website
Total		2835

).

The first source consists of scientific literature. Using the keyword “Physical Internet”, a total of 2537 relevant papers published between 2006 and 2024 were collected. We used the Semantic Scholar API to collect the publication year, author information, abstract, and full text. We chose 2006 as the starting point because it was the year the term “Physical Internet” was first used in the literature [20]. A single keyword, “Physical Internet,” was adopted for keyword searches. This ensures the collection of all relevant literature and guarantees high comprehensiveness. Using multiple keywords simultaneously risks bias toward specific perspectives and could hinder comprehensive literature collection, making this the optimal choice for the objectives of this study.

The second source is conference proceedings of IPIC, papers published in IPIC proceedings from its inception in 2016 to the most recent year 2024, were collected. These proceedings offer valuable discussions on emerging trends, practical implementations, and collaborative efforts in PI research.

The third source is the European roadmap. The PI roadmap developed by the Alliance for Logistics Innovation through Collaboration in Europe (ALICE) was incorporated [21]. This roadmap outlines strategic initiatives and milestones aimed at advancing PI adoption in EU.

The fourth source is the national roadmap. The PI roadmap prepared by Japan’s Ministry of Economy, Trade and Industry (METI) was also included [22]. This document provides a localized perspective on PI implementation strategies and priorities.

For data collection, the language was restricted to English. Exclusion criteria applied included removing duplicate documents with identical DOIs, excluding documents where the full text could not be obtained, and further excluding documents with low relevance to the keywords. These processes ensure that the collected documents and materials are highly reproducible and comprehensively cover information related to the PI domain.

4. Methodology

4.1. Preprocessing

The first step in constructing the KG for PI was to preprocess the raw data collected from four sources: scientific papers, IPIC proceedings, and existing roadmaps from EU and Japan were aggregated into a unified dataset. To ensure data quality, noise such as stop words, URLs, punctuation, and other irrelevant characters symbols removed using the Natural Language Toolkit (NLTK) and programing language Python 3.13.1. The cleaned and standardized corpus served as the input for subsequent graph construction. The overall preprocessing workflow is illustrated in Figure 2.

4.2. Knowledge Graph Construction

After preprocessing, the cleaned data was hosted in a Neo4j AuraDB graph database [23]. Using GPT-4o mini, entity and relationship extraction was conducted to form nodes and edges, which together created a structured knowledge graph representing the PI domain. Nodes typically represent key entities (e.g., “modular container,” “sustainability policy”), while edges capture semantic or functional relationships between them. This structure enables flexible exploration of complex interconnections across heterogeneous information sources.

4.2.1. Entity and Relationship Extraction Process

The entity and relationship extraction process begins with document chunking. This involves dividing large documents into smaller, semantically coherent units, ensuring that processing with language models remains manageable and contextually accurate. Common chunking techniques include token- or sentence-based splitting, with parameters configured to preserve topical integrity. Each chunk is assigned metadata such as the name of the parent document, its position and its length. These chunks are then represented as nodes within the knowledge graph and are linked to their parent documents to enable traceability.

Next comes entity and relationship extraction. Each chunk is passed to a large language model (LLM), such as GPT-4o mini, which is prompted to recognize entities (unique, meaningful concepts such as ‘modular container’ or ‘sustainability policy’) and relationships (semantic or functional connections such as ‘enables’, ‘regulated by’, or ‘part of’). The LLM then outputs structured metadata for each entity, including its title, type, and description. It also outputs metadata for each relationship, including its source and target entities, the relationship type, and a description. Preprocessing is performed using NLP library NLTK for initial cleaning and noun phrase identification. The final definitions and descriptions are always reviewed or summarized using the LLM. The extraction prompt requests a list of all entities, along with their names, types, and descriptions, and the explicit relationships among them. This returns a structured JSON containing these details.

Following extraction, entity disambiguation and merging are performed to avoid redundant nodes in the final graph. Entities with similar names and types across multiple chunks are compared and merged where equivalent, and their descriptions are combined to ensure completeness.

During the node and edge definition stage, entities are represented as nodes characterized by their type and description. Chunks are also represented as nodes in order to preserve the structure and context of the document. Edges capture direct relationships between entities (‘enables’ or ‘is-regulated-by’, for example) and structural links such as ‘PART_OF’ or ‘NEXT_CHUNK’, which connect chunks to their parent documents or sequence them according to document order. Nodes and edges carry properties such as source chunk, document position, entity type or semantic summaries.

4.2.2. Semantic Modeling and Intermediate Steps

During semantic modeling and intermediate steps, nodes and edges are semantically annotated with domain-specific properties and aligned with a chosen ontology or domain schema to improve retrieval and reasoning. Domain-specific terminology and ontologies can guide more precise extraction by customizing LLM prompts to match the field’s language. Additionally, each chunk, entity and relationship description is embedded using a sentence embedding model Paraphrase-MiniLM-L6-v2 [24]. This allows for semantic similarity searches and relevance ranking during retrieval. These embeddings are stored as properties on the nodes for querying in the future. The Paraphrase-MiniLM-L6-v2 model was selected for its high semantic textual similarity and computational efficiency when processing large document corpora. It was tested against other candidate embeddings.

Finally, all nodes and relationships, along with their semantic and structural properties, are ingested into Neo4j. The graph schema usually contains nodes labeled by entity types and relationships labeled by their semantic relationship types. When a query is submitted, GPT-4o mini is used again to translate the natural language question into a graph traversal query in Cypher (query language used in Neo4j) and to interpret and aggregate the relevant nodes and edges returned from the graph. This enables the generation of a natural language response or further reasoning based on the retrieved subgraphs. In the intermediate steps of this querying process, we apply fuzzy matching (see more details in Section 4.5) of query terms to entity nodes using embedding similarity. This process has adjustable thresholds to calibrate retrieval sensitivity.

4.3. Search Strategies

Four retrieval strategies were implemented and compared to evaluate their performance in generating insights. The four strategies are, I: GPT-only, II: Local Search, III: Global Search, and IV: Hybrid Search. The details are explained in

Table 3. Outlines of searching methods applied in the model.

Method	Description
GPT- only	Normal LLM response generation (e.g., ChatGPT). GPT relies solely on the model’s prior knowledge, without database augmentation.
Local Search	Search for nodes and edges in the graph that are semantically close to the query and retrieve directly related nodes and edges from the graph database. Local Search returns nodes and connected subgraphs that are directly relevant to the exact query.
Global Search	Identify communities (clusters) in the graph and obtain a summary of the entire community to which the query relates. Global Search retrieves broader, community-level graph segments to provide contextual breadth.
Hybrid Search	Integrates Local Search and Global Search to generate comprehensive responses with specific examples and background information. Hybrid Search combines node-level detail and global context, using reranking to achieve holistic relevance.

. Hybrid Search is a strategy that combines “details” (Local) and “context” (Global) to generate more accurate and comprehensive knowledge.

These strategies were designed to assess GraphRAG’s ability to retrieve relevant information and generate accurate responses based on different levels of contextual depth.

4.4. Prompt Design

To preliminarily evaluate GraphRAG’s usefulness under different user needs, we designed stakeholder-specific prompts for three groups: industry practitioners, policymakers, and academic researchers (

Table 4. Overview of prompts.

No.	Perspective	Prompt
1	Industry practitioner	What new business models and innovations can be envisioned that leverage the Physical Internet?
2	Policymaker	What policy approaches should be taken to make the Physical Internet a reality in Global?
3	Academic researcher	What areas of research are missing that focus on the Physical Internet?

). From an industry perspective, the focus was on identifying new business models enabled by PI. From a policymaker perspective, the goal was to explore policy approaches to promote global PI adoption. From an academic perspective, the aim was to discover unexplored areas within PI research.

These prompts were executed using all four search strategies. Comparing the results provided initial insights into which approach could yield more comprehensive and actionable findings.

4.5. Parameter Tuning

We employed several parameters to optimize GraphRAG’s performance. First, the temperature of the GPT-4o mini was set to 0 to ensure deterministic and consistent results. Second, the number of nodes referenced during response generation was capped at 50 to balance response speed and comprehensiveness. Third, we adopted fuzzy matching based on cosine similarity for the search algorithm, setting the threshold to 0.5. This threshold is an empirical value based on preliminary experiments, determined by balancing the comprehensiveness and accuracy of search results. Cosine similarity measures the angle between the query’s embedding vector and the node’s embedding vector, enabling the capture of semantic similarity rather than exact string matches. For example, for the query “sustainable logistics policies,” fuzzy matching also retrieves nodes related to “green logistics” and “carbon-neutral freight,” enabling comprehensive coverage of related themes.

Equation (1) defines the relationship between temperature, softmax scaling, and vocabulary distribution in the language model.

$$p_i={\displaystyle \frac{{exp}^{y_{i/T}}}{{\sum }_j{exp}^{y_{j/T}}}}$$

(1)

where $P_i$ is a softmax function representing the probability of the ith class. $Y_i$ is the logit for the $i$th token. $T$ ($T$ > 0) is the LLM temperature parameter and $n$ is the size of the lexicon. Scaling the logit affects the probability assigned to each candidate word.

As the temperature decreases (T → 0), each term $y_i$/T becomes increasingly large for the largest $y_i$ and increasingly small (negative) for all other values. Consequently, the exponential term exp($y_i$/T) for the largest $y_k$ rapidly dominates the denominator in the softmax calculation, causing $p_k$ → 1 for the token with the highest logit while all other probabilities $p_i$ → 0. In this limit, the output effectively approaches a “hard” argmax, where the model almost always selects the token with the highest predicted logit at each step, a behavior known as greedy decoding. Conversely, when T is larger, the resulting probability distribution becomes flatter and more uniform, allowing for greater diversity and randomness in the generated outputs. In summary, as T → 0, softmax sampling becomes deterministic, consistently choosing the most likely token during generation. The value set for each is the parameter is summarized in

Table 5. Tuning metrics.

Metrics	Value
Temperature	0
Output node limit	50
Fuzzy matching score	0.5

.

4.6. Formulation of Search Mechanisms

The search mechanism in GraphRAG consists of three layers: local search, global search, and hybrid search. In local search, the relevance score is calculated based on the cosine similarity between the query vector $q$ and the embedding vector $v_i$ of each node, using the following formula. The relevance score is calculated using the following formula.

$$s_i^{local}={\displaystyle \frac{q·v_i}{‖q‖‖v_i‖}}$$

(2)

This enables node selection based on semantic proximity to the query. In contrast, global search considers the network structure of the entire graph and defines a score $s_i^{global}$ based on the semantic distance $d (q, n_i)$ between nodes using the following formula:

$$s_i^{global}= {\displaystyle \frac{1}{1+ d (q, n_i) }}$$

(3)

This distance metric is defined based on the shortest path length between nodes or semantic similarity, with shorter distances receiving higher scores. Furthermore, in hybrid search, both local and global scores are integrated to calculate the overall relevance score $s_i^{hybrid}$.

$$s_i^{hybrid}=\mathrm{\alpha }{s}_i^{local}+(1-\mathrm{\alpha })s_i^{global},\mathrm{\alpha }\in [0,1]$$

(4)

The weighting coefficient $\alpha$ here is a parameter for adjusting the contribution of local semantic similarity and graph structural relevance and is adaptively set according to search objectives and data characteristics.

4.7. Community Detection

GraphRAG combines community detection to improve the accuracy of node relationships within the knowledge graph. Communities are identified by the Louvain method with the objective of maximizing modularity $Q$. Modularity $Q$ is defined by the following formula:

$$Q = {\displaystyle \frac{1}{2m}}\sum _{i, j}\left[A_{ij}- {\displaystyle \frac{k_i{k}_j}{2m}}\right]\delta \left(c_i, c_j\right)$$

(5)

Here, the adjacency matrix $A_{ij}$ indicates the presence or weight of edges between node $i$ and node $j$, and the degree $k_i$ of node $i$ is defined as the sum of all edges connected to that node (or the total edge weight if weighted). Furthermore, the total number of edges m in the entire graph is represented by half the sum of all elements in the adjacency matrix between all nodes. Each node $i$ belongs to a community $c_i$, and the Kronecker delta $\delta \left(c_i, c_j\right)$ returns $1$ if nodes $i$ and $j$ belong to the same community and $0$ otherwise. By assigning nodes to communities to maximize this modularity $Q$, the Louvain method effectively detects densely connected clusters of nodes within the graph. Based on preliminary experiments, the minimum community size is set to $5$ and the resolution parameter to $1.0$, ensuring a balance between clustering granularity and interpretability.

4.8. Entity Duplication Integration

Entity merging was performed to consolidate identical or semantically equivalent nodes within the knowledge graph. The similarity of node pairs ($v_i$,$v_j$) is calculated as follows:

$$\left(v_i, v_j\right) \in C \mathrm{i}\mathrm{f} sim\left(v_i, v_j\right)\ge {\theta }_{ER}$$

(6)

After integration, attribute $a_k$ a merges attributes $a_i$ and $a_j$ of the original node using the integration function $f$:

$$a_k=f(a_i, a_j)$$

(7)

The edge weight $w_{k, u}$ after integration is calculated using the integration function $g$ from the original node weights $w_{i,u}$ and $w_{j,u}$:

$$w_{k,u}=g(w_{i,u}, w_{j,u})$$

(8)

After merging, perform similarity calculations again and iteratively process if any mergeable nodes exist:

$${\mathcal{V}}^{\prime }\to {\mathcal{V}}^{″} \mathrm{i}\mathrm{f} \exists \left(v_p, v_q\right)\in {\mathcal{V}}^{\prime }, sim\left(v_p, v_q\right) \ge {\theta }_{ER}$$

(9)

The final integrated node set $\mathcal{V}$ is obtained and reflected in the knowledge graph as unique entities without duplication.

4.9. Evaluation Metrics

To quantitatively assess the quality of responses generated by each strategy, six metrics were employed (

Table 6. Evaluation metrics.

Metrics	Abbreviation	Description
Word Count	WC	Total number of words in a sentence.
Type Token Ratios	TTR	TTR assesses vocabulary diversity. A higher TTR indicates a more varied vocabulary.
Average Dependency Distance	ADD	ADD evaluates the distance between word dependencies in a sentence. A higher ADD suggests a more complex sentence structure and increased reading difficulty.
Flesch-Kincaid Grade Level	FKG	FKG is a readability measure that evaluates how challenging a text is to understand, based on the average length of its words and sentences. A higher score indicates greater difficulty.
BERT(P)	P	Precision (P) measures how well the words that appear in Sentence 1 are matched in Sentence 2.
BERT(R)	R	Recall (R) measures how many important words from sentence 2 appeared in sentence 1.
BERT(F)	F	F1 Score (F) measures the harmonic mean of Precision and Recall.

).

For BERTscore we adopt the three scores Precision (P), Recall (R), and F1 Score (F), which are computed as per the following quotations [25,26].

$$P={\displaystyle \frac{1}{\left|\widehat{x}\right|}}\sum _{{\widehat{x}}_j\in \widehat{x}}\underset{{\widehat{x}}_i\in \widehat{x}}{\max } x_i^{\mathrm{\top }}{\widehat{x}}_j$$

(10)

$$R={\displaystyle \frac{1}{\left|x\right|}}\sum _{{\widehat{x}}_i\in \widehat{x}}\underset{{\widehat{x}}_j\in \widehat{x}}{\max } x_i^{\mathrm{\top }}{\widehat{x}}_j$$

(11)

$$F=2{\displaystyle \frac{P·R}{P+R}}$$

(12)

where $x_i$ represent tokenized reference sentences, $x_i^{\mathrm{\top }}$ are sequences of vectors generated by the embedding model, and ${\widehat{x}}_j$ are tokenized candidates mapped to each of the vectors.

These metrics evaluated the relevance, coherence, and completeness of the answers generated by each approach.

5. Results

Prompts and Responses

We constructed a knowledge graph from the above processing (Figure A1). Based on the definitions presented in Section 4.2, we executed the prompt across all four models: (I) GPT alone, (II) local search, (III) global search, and (IV) hybrid search. The resulting outputs are shown in

Table 7. Output of “What new business models can you envision utilizing the Physical Internet?”.

Model	Output
I	The Physical Internet enables new business models such as shared warehouses, flexible transport, and crowdsourced delivery. Companies can share resources without sharing sensitive data, helping build trust and collaboration. By using AI, IoT, and blockchain, logistics become smarter and more efficient. Smart city tech also helps improve urban delivery. These models reduce costs, save time, and support eco-friendly and flexible logistics systems that can adapt to changing needs.
II	The Physical Internet can drive innovations such as collaborative logistics platforms, on-demand delivery services, and smart supply chains that adapt to real-time demand. It also enables modular product design for recycling, dynamic pricing models based on data, and crowdsourced fulfillment to reduce delivery costs. These developments foster sustainability, efficiency, and customer-centric practices.
III	New business models leveraging the Physical Internet (PI) can include federated data spaces that enhance collaboration among logistics stakeholders, enabling synchromodal transport and dynamic freight planning. Innovations may involve decentralized logistics networks where companies share capabilities rather than sensitive data, fostering trust and flexibility. Additionally, integrating smart city technologies can optimize urban logistics, focusing on data sharing and stakeholder engagement to improve efficiency and sustainability. These models can create value by optimizing resource utilization and facilitating adaptive, financially viable logistics systems.
IV	New business models leveraging the Physical Internet (PI) can include federated data spaces that enhance collaboration among logistics stakeholders, enabling synchromodal transport and dynamic freight planning. Innovations may involve decentralized logistics networks where companies share capabilities rather than sensitive data, fostering trust and flexibility. Additionally, integrating smart city technologies can optimize urban logistics, focusing on data sharing and stakeholder engagement to improve efficiency and sustainability. These models can create value by optimizing resource utilization and developing new services that adapt to changing market demands and environmental challenges.

,

Table 8. Output of “What policy approaches should be taken to make the Physical Internet a reality in Global?”.

Model	Output
I	The Physical Internet (PI) aims to create a global system, and for this, several policy strategies are needed. First, a legal framework should be set up, covering transport laws, antitrust rules, and contract responsibilities. It is important to make regulations the same at both the EU and global levels to help connect logistics. Cooperation between the public and private sectors, investing in smart infrastructure, and supporting new technologies are all crucial. Also, focusing on sustainability is important to make sure the Physical Internet helps with environmental goals.
II	To realize the Physical Internet, key policies include standardization of data formats and logistics processes for global interoperability. Infrastructure investment in IoT and automation will support efficient data exchange and goods movement. A robust regulatory framework is necessary to balance innovation with safety and sustainability. Encouraging collaboration, supporting R&D, and strengthening education and training programs will further accelerate progress. These policies will help achieve the vision of a globally connected Physical Internet.
III	The global realization of the Physical Internet requires a multifaceted policy approach. Legal harmonization across jurisdictions is essential to address transport laws, antitrust issues, and contractual complexities. Regulatory convergence and public–private collaboration are crucial for seamless implementation. Strategic investment in smart infrastructure will enhance interoperability and real-time data exchange. Innovation-driven research is key to advancing logistics technology. Ultimately, sustainability must be a core policy focus, ensuring ecological responsibility in logistics evolution.
IV	To implement the Physical Internet globally, several policy strategies are essential. First, a comprehensive legal framework must be developed, addressing transport law, antitrust law, and contractual obligations for cross-border implementation. Harmonizing regulations at the EU and global levels will reduce barriers to logistics integration. Stakeholder engagement, fostering collaboration between public and private sectors, is key to its successful development. Investments in smart infrastructure and support for innovation are crucial for technological integration. Additionally, sustainability must be prioritized to ensure the Physical Internet contributes to environmental goals.

and

Table 9. Output of “What areas of research are missing that focus on the Physical Internet?”.

Model	Output
I	The Physical Internet (PI) aims to make logistics like the digital internet, but more research is needed on standardization, sustainability, security, and the use of drones and autonomous vehicles. There is a lack of studies on the risks, challenges, integration of networks, and business models for PI, as well as the optimization of hub locations and network resilience.
II	Research areas that are missing in the context of the Physical Internet include the development of standardized protocols for data exchange, integration of IoT technologies for real-time tracking and monitoring, sustainability practices in logistics, and the impact of automation and robotics on supply chain efficiency. Additionally, studies on the socio-economic implications and regulatory frameworks surrounding the Physical Internet are also lacking.
III	The areas of research that are missing in the context of the Physical Internet include the exploration of specific implementation barriers, the development of robust business models, and the investigation of stochasticity in hub location and network design. Additionally, there is a need for more studies on the resilience and robustness of hyperconnected networks, as well as frameworks defining hyperconnected networks in practical applications.
IV	The areas of research that are missing in the context of the Physical Internet include a comprehensive exploration of the strengths, risks, challenges, and potential barriers to implementation. Additionally, there is a lack of studies focusing on the integration of hyperconnected logistics networks, the operationalization of the Physical Internet using frameworks similar to the digital internet, and the development of robust business models to support its adoption. Further research is also needed on the stochasticity and resilience of hub location and network design in large-scale optimization problems.

.

The performance metrics (

Table 10. Performance metrics (Bold indicates highest score).

Prompt	1				2				3
Model	I	II	III	IV	I	II	III	IV	I	II	III	IV
WC	81	61	92	97	104	82	81	100	70	67	70	96
TTR	0.741	0.820	0.793	0.784	0.711	0.744	0.790	0.700	0.671	0.687	0.686	0.625
ADD	2.782	2.906	2.898	2.763	3.194	2.533	2.580	2.708	3.500	4.246	3.265	3.232
FKG	11.0	17.1	20.6	21.2	13.5	17.2	19.7	17.1	17.0	23.4	20.5	19.2
BERT (P)	-	0.727	0.754	0.765	-	0.673	0.714	0.836	-	0.698	0.715	0.768
BERT (R)	-	0.750	0.717	0.719	-	0.679	0.735	0.825	-	0.697	0.747	0.721
BERT (F)	-	0.739	0.735	0.741	-	0.676	0.724	0.831	-	0.698	0.731	0.745

) further validated these findings, with Local Search (II) or Hybrid Search (IV) achieving the highest scores across all evaluation criteria, including relevance, coherence, and completeness.

6. Discussion

6.1. Stakeholder-Specific Insights

A comparison of GPT-only model (I) and GraphRAG models (II and IV) is outlined in

Table 11. Comparison of GPT-only model (I) and GraphRAG models (II and IV).

Feature	GPT-Only (I)	Local Search (II)	Hybrid Search (IV)
Specificity	Moderate	High	Very High
Novelty	Commonplace ideas	Emerging concepts	Advanced, integrative
Context	Low	Node-level	Node + community (local + global)
Actionability	Basic guidance	Operational insights	Strategic + system-wide recommendations
Stakeholder alignment	General	Contextualized	Deeply tailored

.

The tailored prompts revealed unique insights for each stakeholder group.

First, from industry perspective represented by prompt, “What new business models can you envision using the Physical Internet?”. GPT-only (I) mentions general ideas like shared warehouses, smart logistics using AI/IoT, and crowdsourced delivery. Overall, they are high-level and somewhat generic, lacking technical or systemic depth. While response from GraphRAG introduces specific mechanisms like collaborative platforms, modular design, real-time demand adaptation, dynamic pricing, connects business models to sustainability and customer-centricity in Local Search (II). The response from Hybrid Search (IV) suggests system-level innovations like federated data spaces and synchromodal transport, integration of urban smart city logistics and market/environmental adaptability. Overall, it highlights more holistic and grounded in technical context, reflecting insights from both node-level detail and community-level context in the graph. In other words, Local Search (II) and Hybrid Search (IV) each possess distinct characteristics. Local Search (II) excels at presenting realistic and actionable insights tailored to the specific needs of each stakeholder, making it suitable for short-term, practical recommendations. Conversely, Hybrid Search (IV) provides comprehensive insights from a long-term perspective that integrates technical, institutional, and societal aspects, making it suitable for optimizing the entire system and strategic decision-making. Therefore, it is crucial to explicitly state which search model is adopted based on the paper’s purpose. Specifically, clearly indicating that Local Search (II) is chosen when prioritizing practicality and feasibility, and Hybrid Search (IV) when emphasizing comprehensive and strategic analysis, allows the reader to clearly understand the rationale behind the model selection.

Second, from government perspective, in response to “What policy approaches should be taken to make the Physical Internet a global reality?”, GPT-only (I) lists common policy suggestions, such as legal frameworks, public–private cooperation, smart infrastructure, and sustainability focus, which are useful but conventional, particularly not tailored to PI’s unique needs. Local Search (II) results offer detailed, actionable steps, such as standardization of data formats, infrastructure investments in automation/IoT, support for education and R&D, clearly addresses global interoperability and innovation readiness. Hybrid Search (IV) results propose structured and layered policy development, such as cross-border legal harmonization, incentives for sustainability, and smart infrastructure tied to stakeholder engagement. Overall, GraphRAG provides stronger emphasis on governance, integration, and ecosystem building—not just tech or funding.

Third, from academia perspective, the prompt “What research areas are missing that focus on the Physical Internet?” identified key gaps in the existing literature. GPT-only model (I) highlights broad areas such as standardization, sustainability, security, drones/AVs, which are good baselines but lack conceptual depth and specifics. Local Search (II) results identify concrete gaps, such as real-time IoT integration, automation impact studies, and socio-economic/regulatory frameworks which are more aligned with emerging research agendas. Hybrid Search (IV) results push boundaries by suggesting:

Cybersecurity frameworks for PI, hyperconnected network operationalization, stochastic and resilient hub-network design. Overall, they address methodological, structural, and practical gaps for real-world PI deployment.

These findings demonstrate GraphRAG’s ability to generate actionable insights tailored to the needs of diverse stakeholders by leveraging its graph-based relational structure.

6.2. Implications on Knowledge Management

The study demonstrates GraphRAG’s potential as a powerful tool for organizing and synthesizing fragmented knowledge in complex domains like PI. By structuring information into interconnected graphs and enabling flexible query mechanisms, GraphRAG bridges gaps between diverse data sources. This capability is particularly valuable for advancing research in emerging fields where knowledge is distributed across multiple disciplines and formats. GraphRAG’s ability to tailor insights to specific stakeholder perspectives enhances its applicability in real-world decision-making contexts. Industry practitioners can identify new business opportunities; policymakers can design informed strategies; and academics can uncover research gaps. All of these can be performed in a unified platform.

6.3. Managerial and Policy Implications

The findings of this study provide actionable guidance for a range of stakeholders. Industry practitioners can use GraphRAG to identify emerging business models and improve operational efficiency by systematically exploring technological and logistical trends. Policymakers can benefit from the robust tool that supports the development of harmonized, data-driven regulatory frameworks to drive PI adoption across different jurisdictions. For academic researchers, GraphRAG enables the identification of under-researched areas, thereby supporting more focused and impactful research initiatives.

By delivering tailored insights that address the specific needs of each stakeholder group, GraphRAG enables informed decision-making and promotes cross-sector collaboration.

6.4. Limitations and Future Work

This study possesses certain strengths, but it also has several limitations. First, it relies on static datasets. If source materials contain omissions or biases, their effects may be reflected in the output results. Second, as the knowledge graph expands, computational complexity may impact performance. Third, the effectiveness of fuzzy matching depends on the granularity and consistency of node labels. Furthermore, while GraphRAG tends to exhibit lower hallucination risk compared to standard LLM generation, this risk is not entirely eliminated. Even with source linking capabilities, the model may still generate inaccurate yet plausible descriptions, potentially undermining user trust in identified research gaps or findings.

To mitigate these risks, it is crucial to provide referable and transparent outputs, such as explicitly indicating which graph nodes or edges support each part of the response. This enables users to verify the basis of the model’s answers and more effectively evaluate the reliability of the presented findings. While this research demonstrates the potential of GraphRAG for structuring and integrating knowledge, its inherent limitations must also be clarified. Unlike established methods such as Systematic Literature Review, Taxonomy, and Critical Review, GraphRAG primarily relies on a Retrieval-Augmented Generation (RAG) framework that extracts and links relevant information from databases and corpora. Consequently, research gap analysis using GraphRAG may lack the comprehensiveness and contextual depth provided by systematic approaches.

For example, even if GraphRAG detects a gap such as “lack of standard protocols,” it cannot sufficiently evaluate whether subsequent research has already addressed it. This limitation stems from the design characteristics of RAG models, which are not inherently suited for critical integration of information or analysis of temporal evolution. Consequently, while GraphRAG is useful for initial mapping of research domains and hypothesis generation, it should ideally be combined with established literature review methods to comprehensively and validly analyze research gaps.

Future research could address these challenges by integrating real-time data streams, expanding multilingual and multimodal knowledge graphs, and improving graph structuring methods to enhance scalability and responsiveness. Implementing user-facing measures to strengthen transparency and source attribution would also be beneficial. Moreover, the findings of this study are preliminary, and empirical experiments will be necessary in the future to confirm the practical usefulness of the insights generated by GraphRAG.

7. Conclusions

This study provides preliminary evidence of the potential utility of GraphRAG as a domain-specific knowledge integration framework for the PI domain. By aggregating fragmented data from academic publications, conference proceedings, and strategic roadmaps into a unified knowledge graph, GraphRAG facilitates systematic exploration of trends, knowledge gaps, and actionable insights tailored to stakeholder needs. The framework leverages graph-based relational structures combined with natural language understanding to enhance information retrieval, particularly for complex or ambiguous queries, while maintaining flexibility for domain-specific applications.

Among the four retrieval strategies evaluated—GPT-only, local search, global search, and hybrid search—local and hybrid search methods appeared to perform relatively better in generating stakeholder-specific responses with greater depth, relevance, and clarity. The use of fuzzy matching also contributed to improved semantic retrieval, supporting the practical usability of the framework.

Unlike traditional systematic reviews or standard RAG-based analyses, this approach focuses on integrating full-text documents (not just abstracts), cross-modal sources (including government roadmaps and conference proceedings), and semantic relationship extraction, all mapped to domain-specific PI ontologies. This allows for granular, stakeholder-tailored queries and identification of nuanced trends and gaps that are difficult to capture with conventional literature reviews or simple keyword-based approaches.

While these findings are preliminary and exploratory, they suggest that GraphRAG can serve as a practical framework for domain-specific knowledge integration, supporting evidence-based decision-making and innovation in the PI context. More broadly, this study illustrates how AI-driven, domain-focused knowledge frameworks can facilitate systematic knowledge synthesis in other complex, interdisciplinary fields.