KG-Anchored RAG: Retrieval-Augmented Generation for Power System Professional Documents Integrating Topic Modeling and Knowledge Graphs

Abstract

In the power industry, how to efficiently and reliably query relevant documents has always posed a challenge for electrical professionals. Unreliable or inefficient query results can lead to significant inefficiencies and introduce unpredictable errors. Hence, a reliable and efficient knowledge querying system is critical. In practice, the effectiveness of Graph-based Retrieval-Augmented Generation (RAG) systems lies in providing expressive representation of entities and graph structures and this makes it stand out as a widely-used approach for document retrieval. However, typical GraphRAG frameworks encounter challenges such as semantic dilution and topological drift caused by generic technical terminology and granular graph noise especially in professional documents like regulations, etc, which is one of the mostly used type of document in electric industry. Thus, we propose KG-Anchored RAG, a framework that shifts the retrieval paradigm from community-based summarization to precision-guided anchoring. During knowledge construction, our framework employs a topological skeleton refinement and constructs a Knowledge Attachment Matrix using latent topic modeling and one-hot feature injection. During inference, non-linear sharpening and PageRank-based structural resonance are utilized to locate high-density knowledge cells. Evaluation on professional documents in the power industry reveals that our method outperforms localized search baselines in terms of context precision, generative faithfulness, and ranking quality. The proposed framework demonstrates a superior ability to prioritize evidentiary clauses and reduce information redundancy without relying on computationally expensive external re-rankers. Experimental results indicate that KG-Anchored RAG effectively mitigates speculative hallucinations, establishes a reliable architectural paradigm for retrieval-augmented generation in high-stakes, safety-critical vertical industries.

1. Introduction

With the rapid advancement of power systems, the operational paradigms of power systems have become increasingly complex. This complexity is directly manifested in the generation of massive volumes of multi-source, heterogeneous data, including equipment ledgers, regional management regulations, operational logs, and statistical indicators. In such a complex data environment, indexing the data to efficiently query specific information and retrieve associated explanations has become a major challenge [1,2,3]. For example, daily tasks for dispatchers, such as equipment maintenance and decision-making, have become heavily reliant on efficient and accurate knowledge querying. Thus, effective Question Answering (QA) systems have become a core technique for the situation [4]. With the advancement of Natural Language Understanding(NLU) technologies, QA systems have gradually evolved from rule-based approaches, to knowledge-based systems, and finally to contemporary Large Language Models(LLMs)-based frameworks [5]. Nowadays, Large Language Models (LLMs) have become a cornerstone of QA systems, leveraging their powerful ability to bridge the gap between human intent and natural language [6]. Besides, throughout the evolution of NLU technologies, Knowledge Graphs (KGs) have been widely adopted to support the QA systems because of their structured modeling capabilities and superior representation of complex relations [7].

However, existing power domain QA methods face challenges in effectively merging the structural precision of KGs with the generative flexibility of LLMs [8,9]. Traditional KG-based QA systems rely on embedding models for vector-based matching between queries and graph nodes. While these methods utilize the structural advantages of KGs, they often fail to exploit the natural language understanding and semantic modeling depths of modern LLMs, making it difficult to process complex, multi-hop natural language queries [10]. Recently, Retrieval-Augmented Generation (RAG) [11] and specifically GraphRAG [10]—which constructs graph indices via document chunking and community clustering—have emerged to handle unstructured data.

Despite their success in general domains, standard GraphRAG frameworks expose two flaws in the power system vertical [12,13]. The first is a phenomenon we refer to as Semantic Dilution [14], which reflects the reduced discriminative capability of embeddings under high-frequency technical terms. These terms cause document vectors to become overly dense in the embedding space, making cosine similarity ineffective at distinguishing subtle semantic differences [15]. The second is a graph-noise phenomenon we describe as Topological Drift [16], where automatically constructed graphs become saturated with low-salience nodes, which mask the core hierarchical business logic such as “fault type–handling measures” or “equipment category–technical specifications”. As illustrated in Figure 1, these issues often lead GraphRAG to retrieve broad, noisy community summaries, resulting in speculative hallucinations during the LLM generation phase—a risk that is unacceptable in high-stakes operations.

To address these systemic issues, this paper proposes the KG-Anchored RAG framework, specifically designed to adapt to the inherent structural characteristics of knowledge in power system. Our work focuses on designing a problem-driven retrieval framework tailored to the structural characteristics of professional technical documents. The core philosophy of this framework is to establish an integrated “Topological Skeleton–Semantic Attachment” mechanism that uses domain-specific core concepts as anchors to achieve precise mapping from unstructured documents to a structured backbone. Specifically, we first refine the original knowledge graph to extract a topologically representative skeleton of domain concepts [17]. Subsequently, we introduce a BERTopic-based method [18] to construct a hierarchical semantic attachment matrix. By fusing semantic embeddings with discrete skeleton one-hot features, we create augmented document representations that are precisely mapped to the skeleton via a dual-matrix (Document–Topic and Topic–Skeleton) transformation. Finally, we implement a Non-linear Sharpening retrieval process and a PageRank-based structural resonance reranking. This pipeline ensures that only high-density knowledge cells surrounding the activated skeleton nodes are provided to the LLM, reducing information redundancy and hallucination risks.

The main contributions of this work are summarized as follows:

1.

We propose the KG-Anchored RAG framework tailored for the power domain. By creating a closed-loop design of topological refinement, hierarchical attachment, and structural resonance, we systematically resolve the semantic dilution and topological drift issues, providing a practical retrieval framework for structurally complex technical corpora.

2.

We design a Topological Skeleton Refinement for core node selection and a Knowledge Attachment Matrix mechanism that fuses semantic embeddings with one-hot structural features. Furthermore, we introduce a Non-linear Sharpening mechanism and a PageRank-based strategy, enhancing the discriminative power for technical terms and the specificity of knowledge anchoring.

3.

We conduct a comprehensive evaluation using a proprietary power system dataset with professional-grade test queries. Through Ragas, BERTScore, and ROUGE metrics, we demonstrate the superiority of our method over SOTA baselines. Ablation and sensitivity analyses provide a quantitative basis for the engineering deployment of the proposed framework by identifying optimal configurations for retrieval scope and sharpening parameters.

2. Methodology

2.1. Problem Formulation

This work is motivated by recurring structural characteristics commonly observed in professional technical corpora. In specialized vertical domains such as power system regulation, a knowledge base is typically represented as a collection of n unstructured document fragments (or text units), denoted by $\mathcal{D}=\{d_1,d_2,\dots ,d_n\}$. Through preliminary entity extraction and relation discovery, an initial knowledge graph $\mathcal{G}=(\mathcal{V},\mathcal{E})$ is constructed, where $\mathcal{V}$ represents the set of all identified entities and $\mathcal{E}\subseteq \mathcal{V}\times \mathcal{V}$ denotes their inter-relations.

Traditional Retrieval-Augmented Generation (RAG) and GraphRAG frameworks typically rely on a retrieval function $f(q,\mathcal{D},\mathcal{G})\to \mathcal{C}$, which selects a subset of fragments $\mathcal{C}\subset \mathcal{D}$ via vector similarity or graph neighborhood expansion. However, in professional contexts, this approach encounters two fundamental mathematical barriers that degrade performance:

1.

Semantic Dilution in Latent Space: Domain-specific terminology (e.g., “load reserve”, “capacity”) often exhibits disproportionately high Document Frequency ($DF$) [19]. In the high-dimensional embedding space ${\mathbb{R}}^d$, these terms produce an overly dense and low-variance vector distribution. This phenomenon, which we define as Semantic Dilution [14], collapses the discriminative power of cosine similarity, causing relevant technical clauses to be buried under generic operational noise.

2.

Topological Drift in Automated Graphs: KGs automatically generated from technical documents often suffer from Topological Drift [16], where the graph becomes saturated with granular, low-salience nodes ($\left|\mathcal{V}\right|\gg {10}^3$). These noisy nodes (e.g., redundant equipment IDs, auxiliary descriptions) obstruct the core hierarchical logic (e.g., Safety Standard → Operation Limit → Threshold), leading to imprecise multi-hop reasoning.

To resolve these issues, we formulate the retrieval problem as a Cross-Space Mapping task between the unstructured fragment space and a structured concept space. We define a Topological Skeleton $\mathcal{P}=\{p_1,p_2,\dots ,p_m\}$ as a refined set of m domain anchors ($m\ll \left|\mathcal{V}\right|$) representing the domain’s invariant backbone. Our primary objective is to construct a Knowledge Attachment Matrix $\mathbf{B}$$\in {\mathbb{R}}^{n\times m}$, where each element $b_{i,j}$ quantifies the relative semantic association between document fragment $d_i$ and skeleton concept $p_j$.

Formally, given a user query q, the task is to transform the unstructured query into a sharpened intent vector ${\mathbf{p}}_q\in {\mathbb{R}}^m$ within the skeleton space. The retrieval process is then redefined as a coordinate-driven navigation:

$${\mathcal{C}}^{*}=\arg {\mathrm{top}}_k({\mathbf{p}}_q·{\mathbf{B}}^{\top })$$

(1)

where ${\mathcal{C}}^{*}$ represents the optimal set of high-density “Knowledge Cells” that provide precise evidentiary support for high-fidelity generative responses.

Intuitively, Equation (1) reformulates retrieval as a coordinate-based matching process in the skeleton concept space. Instead of directly retrieving document chunks through neighborhood expansion or raw vector similarity, the query is first projected into a structured concept representation ${\mathbf{p}}_q$, where each dimension corresponds to a refined skeleton concept.

Meanwhile, the matrix $\mathbf{B}$ stores the attachment strength between document fragments and skeleton concepts. Therefore, the matrix product ${\mathbf{p}}_q·{\mathbf{B}}^{\top }$ computes the relevance between the query intent and all document fragments in the same structural space. The retrieved set ${\mathcal{C}}^{*}$ can thus be interpreted as the collection of document units most strongly aligned with the activated structural concepts.

2.2. KG-Anchored Retrieval

As illustrated in Figure 1, our framework consists of an offline knowledge anchoring phase and an online precision navigation phase. By transitioning from standard neighborhood expansion to coordinate-based mapping, we achieve higher signal density for professional QA tasks.

2.2.1. Topological Skeleton Refinement

The process begins with the extraction of a domain-specific backbone from the initial knowledge graph $\mathcal{G}$. Unlike standard GraphRAG, which treats all extracted entities equally, we perform Topological Skeleton Refinement to filter out high-frequency generic terms that contribute to topological drift. For each node v, we measure its frequency based on its degree centrality ($Deg\left(v\right)$) and its inverse document frequency ($IDF\left(v\right)$) [19].

As visualized in Figure 2, our refinement strategy effectively partitions the entity space. Entities belonging to core categories (e.g., equipment, standards) or those with high connectivity but moderate-to-low global frequency are retained in the Skeleton Zone to form the Skeleton Word List $\mathcal{P}$. Conversely, terms exhibiting high semantic dilution are isolated in the Noise Zone and pruned. This step ensures that the navigational “landmarks” of the system are representative of the actual business logic and possess high discriminative power for precision retrieval.

2.2.2. Semantic Attachment

To link unstructured text units to the refined skeleton, we implement a two-step mapping process:

1.

One-hot Injection: For each document chunk $d_i$, we generate a composite representation ${\mathbf{E}}_i=[{\mathbf{v}}_{sem}\oplus \lambda ·{\mathbf{v}}_{oh}]$, where ${\mathbf{v}}_{sem}$ is the semantic embedding and ${\mathbf{v}}_{oh}$ is the One-hot Encoding relative to the skeleton $\mathcal{P}$. This forces the topic modeling process to be “aware” of the structural backbone.

2.

Matrix Computation: We perform latent topic modeling on the augmented space to derive the Document-Topic matrix $\mathbf{H}$ and the Topic-Skeleton matrix ${\mathbf{T}}^{\prime }$. The final Knowledge Attachment Matrix $\mathbf{B}$ is computed as:

$$\mathbf{B}=\mathbf{H}\times {\mathbf{T}}^{\prime }$$

(2)

Here, $\mathbf{H}\in {\mathbb{R}}^{n\times t}$ represents the document-topic association matrix obtained from latent topic modeling, where each row describes the distribution of a document fragment over t latent topics. The matrix ${\mathbf{T}}^{\prime }\in {\mathbb{R}}^{t\times m}$ denotes the topic-to-skeleton mapping matrix, which captures the association between latent topics and refined skeleton concepts. The resulting matrix $\mathbf{B}$ provides a unified coordinate system in which each document fragment is softly attached to multiple structural concepts according to semantic relevance. This matrix serves as a pre-computed coordinate system where each document fragment is “attached” to its corresponding skeleton nodes via latent semantic bridges.

As shown in Figure 3, the topic space exhibits high resolution with 19 distinct clusters. This visual evidence supports our hypothesis that the Knowledge Attachment Matrix $\mathbf{B}$ can overcome Semantic Dilution by providing a granular coordinate system. The tight grouping of technical segments indicates that fragments discussing similar regulatory requirements are successfully mapped to the same skeletal anchors.

2.2.3. Precision Navigation and Structural Resonance (Online Inference)

During inference, our method follows a “search-then-resonate” logic to outperform the broad expansion used in baselines:

1.

Non-linear Sharpening: The user query q is transformed into a query vector ${\mathbf{p}}_q$. We apply a non-linear sharpening function (Softmax with temperature $\tau =3$) to polarize the weights. This effectively suppresses irrelevant nodes and creates a “laser beam” intent that points to specific skeleton concepts (e.g., “Load Reserve Capacity”).

2.

Coordinate Mapping: The system performs a matrix product ${\mathbf{p}}_q·{\mathbf{B}}^{\top }$ to perform Document Navigation. This allows the system to skip generic community summaries and directly identify specific regulatory clauses.

3.

Structural Resonance: To refine the ranking, we construct a local heterogeneous graph of the top-15 candidates and apply PageRank-based approach. This identifies the “semantic center” among retrieved fragments, ensuring that the most evidentiary document outranks background descriptions.

3. Discussion

3.1. Empirical Analysis of Graph and Semantic Pathologies

Before evaluating the downstream QA performance, we first analyze the structural and semantic characteristics of the automatically constructed knowledge graph. Our analysis reveals that power-system regulatory corpora exhibit severe graph sparsity and semantic redundancy, which directly challenge the assumptions underlying standard GraphRAG retrieval.

3.1.1. Graph Structural Degeneration

Table 1. Structural Statistics of the Constructed Knowledge Graph.

Metric	Value
Total Nodes	2556
Total Edges	272
Graph Density	$4.16\times {10}^{-5}$
Connected Components	2309
Largest Component Size	50
Largest Component Ratio	1.96%
Degree-0 Nodes	2246
Degree-0 Ratio	87.87%
Low-value Nodes ($deg\le 1$)	96.87%
Near-uniform PageRank Ratio	92.95%

summarizes the structural statistics of the generated knowledge graph. The graph contains 2556 nodes but only 272 edges, resulting in an extremely sparse topology with a density of $4.16\times {10}^{-5}$. More critically, 2246 nodes (87.87%) are completely isolated, while 96.87% of all nodes have degree less than or equal to one.

These observations indicate that the graph does not satisfy the connectivity assumptions required by neighborhood-expansion-based GraphRAG frameworks. Instead of forming semantically navigable communities, the graph degenerates into a large collection of fragmented and weakly connected components. Furthermore, the PageRank distribution is nearly uniform for 92.95% of nodes, suggesting that the graph lacks sufficiently discriminative structural hubs for reliable multi-hop reasoning.

This phenomenon quantitatively validates the existence of the Topological Drift problem discussed in Section 3, where automatically extracted entities fail to preserve the true hierarchical business logic of the power-system domain.

3.1.2. Semantic Redundancy and Embedding Collapse

We further analyze the semantic distribution of document chunks in the embedding space. As shown in

Table 2. Semantic Statistics of the Regulatory Corpus.

Metric	Value
Vocabulary Size	2602
Pairwise Cosine Mean	0.8062
Nearest-neighbor Cosine Mean	0.9255
Top-k Repetition Rate	0.1891
Head Coverage (Top-50 Terms)	0.3500
Zipf Slope	−1.1080

, the average pairwise cosine similarity between document embeddings reaches 0.8062, while the average nearest-neighbor similarity further increases to 0.9255.

The unusually high embedding similarity demonstrates that power-system regulations occupy an overly concentrated latent semantic space. High-frequency operational terms such as “system”, “control”, and “requirement” repeatedly appear across different technical clauses, causing semantically distinct fragments to become densely clustered in the vector space.

This embedding-space crowding substantially weakens the discriminative capability of cosine-similarity-based retrieval and quantitatively confirms the Semantic Dilution phenomenon introduced earlier. Consequently, standard vector retrieval and community summarization methods tend to retrieve broad but weakly relevant context, increasing hallucination risk during generation.

3.2. Experimental Setup

3.2.1. Dataset

The evaluation is conducted on a proprietary dataset of power system professional documents. The dataset consists of 8 professional documents, including the “Technical Guidelines for Power Systems” and the “Electric Power Thesaurus” series. During the indexing phase, the documents were partitioned into 122 granular document units with a maximum chunk size of 1200 tokens to facilitate high-resolution retrieval.

3.2.2. Baseline Models

To evaluate the effectiveness of the proposed framework, we compare our method against the following retrieval-augmented generation (RAG) baselines:

• Naive RAG [5]: A conventional RAG approach that retrieves the top-K document fragments based purely on vector cosine similarity in the embedding space.
• GraphRAG-Local [10]: A localized graph-based search method proposed by Microsoft that identifies relevant entities and retrieves their immediate neighborhood along with pre-computed community summaries.
• GraphRAG-Global [10]: A global search strategy that leverages top-down community reports to provide a macroscopic summary of the entire knowledge graph.
• Hybrid-Search: A sparse-dense hybrid retrieval strategy combining vector similarity with lexical keyword matching.
• Hybrid BM25-Dense: A weighted fusion retrieval method integrating BM25 sparse retrieval and dense embedding similarity.
• BGE-Reranker: A two-stage retrieval pipeline employing dense retrieval followed by cross-encoder neural reranking using the BGE reranker model.

3.2.3. Implementation Details

We utilized Qwen-2.5-32B [20] as the backbone LLM for answer generation while DeepSeek-R1-32B was used exclusively for evaluation and answer quality assessment, and nomic-embed-text:latest [21] was selected for generating textual embeddings for initial retrieval and. The topic mapping was implemented using BERTopic [18]. To handle the medium-scale dataset of 122 units. Our reranker is based on PageRank [22], integrating keyword density, matrix-based topic scores, and vector similarity into a heterogeneous graph.

The final hyper-parameter selections after parameter sensitivity testing are summarized in

Table 3. Hyper-parameter Configurations.

Parameter	Description	Range	Selection
c	Primary document retrieval count	$\{3,5,10\}$	5
g	Auxiliary topic compensation count	$\{1,3,5\}$	3
$\tau$	Softmax sharpening temperature	$\{1,2,3\}$	3
$\delta$	Semantic mapping threshold	$\{0.6,0.75,0.85\}$	0.75

.

3.2.4. Evaluation Metrics

We employ a comprehensive set of metrics to evaluate the system from multiple perspectives. We use the Ragas framework [23] to compute Faithfulness, Answer Relevancy, Context Precision, and Context Recall. BERTScore [24] is utilized to measure the token-level semantic similarity between generated answers and Ground Truth (GT). ROUGE (ROUGE-2 and ROUGE-L) [25] is used to assess the retention of rigid technical terminology. And we implement nDCG@K [26] to evaluate the effectiveness of our native matrix-driven and PageRank-based ranking without external re-rankers.

3.3. Main Results

The evaluation results of our proposed framework in comparison with baseline models are summarized in

Table 4. Comprehensive Performance Comparison Based on RAGAS and Generation Metrics.

Model	Faith.	C-Pre.	C-Rec.	Ans-Rel.	BS-F1	BS-R	R-1 F1	R-2 F1	R-L F1
Naive RAG	$0.{165}_{\pm 0.021}$	$0.{090}_{\pm 0.008}$	$0.{111}_{\pm 0.010}$	$0.{647}_{\pm 0.018}$	$0.{635}_{\pm 0.006}$	$0.{725}_{\pm 0.014}$	$0.{021}_{\pm 0.005}$	$0.{009}_{\pm 0.003}$	$0.{021}_{\pm 0.005}$
GraphRAG-Global	$0.{179}_{\pm 0.023}$	$0.{154}_{\pm 0.015}$	$0.{238}_{\pm 0.027}$	$0.{653}_{\pm 0.024}$	$0.{645}_{\pm 0.005}$	$0.{738}_{\pm 0.016}$	$0.{058}_{\pm 0.006}$	$0.{0177}_{\pm 0.004}$	$0.{058}_{\pm 0.006}$
GraphRAG-Local	$0.{280}_{\pm 0.020}$	$0.{233}_{\pm 0.025}$	$0.{103}_{\pm 0.013}$	$0.{638}_{\pm 0.011}$	$0.{645}_{\pm 0.005}$	$0.{750}_{\pm 0.016}$	$0.{078}_{\pm 0.004}$	$0.{0310}_{\pm 0.004}$	$0.{078}_{\pm 0.004}$
Hybrid-Search	$0.{188}_{\pm 0.007}$	$0.{171}_{\pm 0.018}$	$0.{291}_{\pm 0.046}$	$0.{624}_{\pm 0.023}$	$0.{632}_{\pm 0.010}$	$0.{703}_{\pm 0.007}$	$0.{018}_{\pm 0.011}$	$0.{0065}_{\pm 0.003}$	$0.{018}_{\pm 0.011}$
Hybrid BM25-Dense	$0.{326}_{\pm 0.029}$	$0.{312}_{\pm 0.038}$	$0.{444}_{\pm 0.021}$	$0.{649}_{\pm 0.024}$	$0.{650}_{\pm 0.004}$	$0.{753}_{\pm 0.016}$	$0.{086}_{\pm 0.014}$	$0.{0380}_{\pm 0.006}$	$0.{086}_{\pm 0.014}$
BGE-Reranker	$0.{335}_{\pm 0.043}$	$0.{362}_{\pm 0.021}$	$\mathbf{0}.{\mathbf{467}}_{\pm \mathbf{0}.\mathbf{029}}$	$0.{664}_{\pm 0.027}$	$0.{666}_{\pm 0.004}$	$\mathbf{0}.{\mathbf{764}}_{\pm \mathbf{0}.\mathbf{016}}$	$0.{091}_{\pm 0.011}$	$0.{0410}_{\pm 0.005}$	$0.{091}_{\pm 0.011}$
Ours (Proposed)	$\mathbf{0}.{\mathbf{386}}_{\pm \mathbf{0}.\mathbf{024}}$	$\mathbf{0}.{\mathbf{440}}_{\pm \mathbf{0}.\mathbf{016}}$	$0.{320}_{\pm 0.025}$	$\mathbf{0}.{\mathbf{669}}_{\pm \mathbf{0}.\mathbf{011}}$	$\mathbf{0}.{\mathbf{674}}_{\pm \mathbf{0}.\mathbf{012}}$	$0.{755}_{\pm 0.007}$	$\mathbf{0}.{\mathbf{117}}_{\pm \mathbf{0}.\mathbf{004}}$	$\mathbf{0}.{\mathbf{048}}_{\pm \mathbf{0}.\mathbf{004}}$	$\mathbf{0}.{\mathbf{114}}_{\pm \mathbf{0}.\mathbf{004}}$

Faith. = Faithfulness; C-Pre. = Context Precision; C-Rec. = Context Recall; Ans-Rel. = Answer Relevancy; BS = BERTScore; R-1/R-2/R-L = ROUGE-1/ROUGE-2/ROUGE-L. All results are reported as mean ± standard deviation over repeated evaluations. Bold values indicate the best performance among all compared methods.

and

Table 5. Ranking Quality Comparison (nDCG@K) under Different Retrieval Budgets (K).

Budget (K)	GraphRAG-Local	Ours (Proposed)
$K=3$	0.3261	0.4664
$K=5$	0.5649	0.6480
$K=10$	0.5498	0.6413

The retrieval budget K refers to the number of original text units (chunks) provided to the LLM.

. Across all key dimensions of the RAG pipeline—retrieval accuracy, generative faithfulness, and semantic alignment—the KG-Anchored RAG exhibits a consistent performance advantage over standard GraphRAG and Hybrid-Search methods.

As observed in Table 4, the proposed method achieves a Context Precision of 0.4400 and a Faithfulness score of 0.3835, demonstrating a stable performance advantage over the baselines. In contrast, GraphRAG-Local scores 0.1209 and 0.1769 in these respective metrics. The lower precision in the baseline models is primarily attributed to the increased semantic entropy within the expanded 122-unit retrieval space. In power system regulations, high-frequency generic terms create dense clusters that traditional community-based summarization fails to differentiate. While the GraphRAG baselines often retrieve broad neighborhood summaries that dilute specific technical evidence, our method utilizes the refined topological skeleton to navigate the Knowledge Attachment Matrix. By establishing precise coordinates between core concepts and document units, this mechanism effectively filters out generic operational noise and anchors the retrieval onto the most relevant evidentiary segments.The improvement in generative integrity is further evidenced by the doubling of the Faithfulness score compared to the baselines. This technical efficiency is most notably highlighted by the ROUGE-2 F1 scores; our method (0.0482) is greatly higher than that of GraphRAG-Local (0.0031). This substantial margin confirms that our proposed approach successfully preserves rigid technical efficiency.

The impact of the retrieval budget (K) on the ranking quality, as summarized in Table 5, reveals a consistent performance advantage for our proposed framework over the GraphRAG-Local baseline. Across all tested budgets, the native matrix-driven ranking—further refined by Personalized PageRank (PPR)—demonstrates superior ability in prioritizing relevant technical evidence. As shown in Table 5, the ranking quality for both models improves as K increases from 3 to 5. This suggests that with the increased complexity of 122 document chunks, a very restricted budget ($K=3$) may occasionally exclude primary evidence from the candidate pool. However, even at this small scale, our method achieves an nDCG of 0.4664, which is 43.0% higher than the 0.3261 recorded by GraphRAG-Local. This margin confirms that our precision-guided anchoring is more efficient at identifying core technical clauses with minimal data compared to the summarization-heavy approach of the baseline. The highest ranking performance for our method is achieved at $K=5$, reaching an nDCG of 0.6480. Beyond this point, as seen at $K=10$, the score for both models begins to decline. This trend supports the observation that in specific domains, increasing the retrieval volume beyond a certain threshold introduces secondary, less relevant document units. These document units create semantic interference, which slightly complicates the ranking process. Nevertheless, our framework maintains a steady lead at $K=10$, outperforming the baseline by 16.6%. This robustness is attributed to the structural resonance mechanism, which effectively filters out the additional noise and maintains the priority of the most evidentiary knowledge cells at the top of the context list.

Statistical Significance Analysis

To evaluate the robustness of the observed improvements, we further conducted statistical significance testing across all evaluation metrics. Each experiment was repeated multiple times under different random seeds, and paired significance tests were performed between the proposed method and competing baselines.

Table 6. Statistical Significance Analysis Against Baselines.

Metric	Local	Global	Hybrid	BM25-Dense	Rerank
Faithfulness	$1.3\times {10}^{-5}$	$5.3\times {10}^{-7}$	$5.4\times {10}^{-7}$	$9.5\times {10}^{-6}$	$1.3\times {10}^{-5}$
Answer Relevancy	$4.2\times {10}^{-4}$	$0.005$	$0.002$	$0.002$	$0.004$
Context Precision	$2.9\times {10}^{-6}$	$4.3\times {10}^{-8}$	$4.4\times {10}^{-7}$	$2.2\times {10}^{-6}$	$1.3\times {10}^{-6}$
Context Recall	$5.2\times {10}^{-6}$	$2.7\times {10}^{-6}$	$5.3\times {10}^{-6}$	$1.2\times {10}^{-5}$	$1.7\times {10}^{-5}$
BERTScore-F1	$0.002$	$0.001$	$4.8\times {10}^{-4}$	$5.4\times {10}^{-4}$	$0.006$
ROUGE	$4.1\times {10}^{-4}$	$1.6\times {10}^{-4}$	$4.6\times {10}^{-5}$	$0.012$	$0.048$
nDCG	$0.022$	$0.012$	$7.1\times {10}^{-6}$	$0.019$	$0.017$

reports the corresponding p-values against the strongest baseline methods. Across all major metrics, the proposed KG-Anchored RAG achieves improvements ($p<0.05$), with most metrics reaching stronger significance levels.

The statistical results confirm that the observed performance gains are not caused by random fluctuations, but instead arise from the consistent effectiveness of the proposed topology-guided retrieval framework.

3.4. Ablation Study and Hyperparameter Sensitivity Analysis

To verify the scientific necessity of each module within the KG-Anchored RAG and to evaluate its robustness under varying configurations, we conducted a series of ablation and sensitivity experiments.

The comprehensive ablation analysis presented in

Table 7. Comprehensive Ablation Results Using All Metrics.

Module Variant	Faith.	C-Pre.	C-Rec.	Ans-Rel.	BS-F1	BS-R	Avg. Len ↓
Full (Ours)	0.3835	0.4400	0.3200	0.6686	0.6740	0.7549	2144
w/o Skeleton Refinement	0.2218	0.3000	0.1000	0.5746	0.6432	0.7240	2427
w/o Semantic Bridge	0.2985	0.1015	0.1420	0.5687	0.6708	0.7279	2447
w/o Non-linear Sharpening	0.1709	0.3000	0.1032	0.5698	0.6545	0.6997	2442
w/o PageRank	0.1074	0.1165	0.1815	0.5321	0.6617	0.7345	2132

↓ indicates that lower values are better. Bold values indicate the best performance among all compared variants.

confirms that each modular component of the KG-Anchored RAG framework is essential for maintaining both retrieval accuracy and generative integrity. The data indicates that the removal of any single module leads to a measurable decline across all performance indicators, with different modules serving distinct roles in the Anchoring–Navigation–Generation pipeline.

As evidenced by the results, the PageRank-based module is a key factor for generative reliability. Its removal (Variant w/o PageRank) causes the most dramatic collapse in faithfulness, dropping from 0.3835 to a mere 0.1074, and the lowest recorded answer relevancy. This confirms that even if relevant documents are present in the candidate pool, the lack of structural consensus provided by the PageRank-based algorithm results in the LLM receiving conflicting or noisy signals, which triggers speculative hallucinations and logical incoherence.

The Semantic Bridge ($t^{\prime }$) serves as the primary anchor for retrieval specificity. Excluding this vector-space mapping (Variant w/o Semantic Bridge) leads to the lowest context precision. Without the alignment provided by $t^{\prime }$, the system fails to bridge the gap between abstract topics and the structured skeleton, effectively reverting to a sparse literal search that is unable to handle specific terminology or technical abbreviations. This disconnect is also reflected in the drop in BERTScore recall to 0.7279, indicating that a portion of the evidentiary chain is lost when the semantic bridge is broken.

The impact of noise suppression is illustrated by the Skeleton Refinement ablation. Removing the skeleton refinement results in an increase in average context length to 2427 characters. This suggests that without a refined skeleton, the navigational space is overwhelmed by generic high-frequency terms, causing the “feature drowning” effect where core technical evidence is buried under operational noise.

Finally, the Non-linear Sharpening module is shown to be vital for retrieval focus. Without the high-temperature Softmax polarization (Variant w/o Non-linear Sharpening), the system exhibits a sharp decline in context recall and faithfulness. This data validates the necessity of our approach; without it, the retrieval signals remain too diffuse to consistently prioritize the correct regulatory clauses within the Knowledge Attachment Matrix. Overall, the ablation study demonstrates that the superior performance of KG-Anchored RAG is derived from the synergistic integration of topological filtering, latent semantic mapping, and structural graph resonance.

The sensitivity of the retrieval scope is analyzed through Top-c and Top-g in

Table 8. Sensitivity Analysis of Top-c (with Top-g = 3).

Top-c	BS-F1	BS-R	Faith.	C-Pre.	C-Rec.	Ans-Rel.
$c=3$	0.6600	0.7676	0.2175	0.4420	0.1000	0.6097
$c=5$	0.6740	0.7549	0.3835	0.4400	0.3200	0.6686
$c=10$	0.6709	0.7701	0.1726	0.3380	0.3290	0.7413

and

Table 9. Sensitivity Analysis of Top-g (with Top-c = 5).

Top-g	BS-F1	BS-R	Faith.	C-Pre.	C-Rec.	Ans-Rel.
$g=1$	0.6799	0.7844	0.4050	0.4910	0.2940	0.6588
$g=3$	0.6740	0.7549	0.3835	0.4400	0.3200	0.6686
$g=5$	0.6563	0.7349	0.1811	0.3720	0.4135	0.6740

. The results indicate that performance peaks at moderate retrieval volumes, following a “less-is-more” principle regarding information density. In Table 8, $c=5$ represents the optimal balance, achieving the highest Faithfulness (0.3835) and BERTScore F1 (0.6740). Similarly, Table 9 shows that the most precise signal is captured at $g=1$, yielding peak Context Precision (0.4910) and Faithfulness (0.4050).

As the retrieval budget expands to $c=10$ or $g=5$, while Context Recall expectedly improves, Faithfulness suffers a sharp decline (dropping below 0.20). This collapse suggests that larger retrieval windows introduce redundant, weakly-related segments that create “semantic interference.” In professional domains, these noisy distractors overwhelm the LLM’s reasoning process, proving that maintaining a high signal-to-noise ratio within a compact context is more effective for high-fidelity answering than broad information gathering.

The impact of the internal mapping and activation parameters is detailed in

Table 10. Impact of Similarity Threshold $\delta$.

$\mathit{\delta }$	BS-F1	BS-R	Faith.	C-Pre.	C-Rec.	Ans-Rel.
$\delta =0.65$	0.6646	0.7557	0.1035	0.2000	0.3710	0.6796
$\delta =0.75$	0.6740	0.7549	0.3835	0.4400	0.3200	0.6686
$\delta =0.85$	0.6711	0.7633	0.3036	0.2000	0.1000	0.6339

and

Table 11. Impact of Sharpening Temperature $\tau$.

$\mathit{\tau }$	BS-F1	BS-R	Faith.	C-Pre.	C-Rec.	Ans-Rel.
$\tau =1$	0.6623	0.7374	0.1565	0.2000	0.2000	0.6456
$\tau =2$	0.6649	0.7396	0.1939	0.3000	0.1620	0.6299
$\tau =3$	0.6740	0.7549	0.3835	0.4400	0.3200	0.6686

. As shown in Table 10, the Similarity Threshold ($\delta$) exhibits an optimal value at 0.75, yielding the highest Faithfulness (0.3835) and Context Precision (0.4400). Lowering the threshold to 0.65 increases the Context Recall to 0.3710 by allowing more permissive mapping between topics and the skeleton, but it introduces noise that degrades the Faithfulness score to a minimum of 0.1035. Conversely, a threshold of 0.85 is overly restrictive, resulting in broken navigational paths and a minimal recall of 0.1000.

The Sharpening Temperature ($\tau$) analysis in Table 11 demonstrates that $\tau =3$ provides the most effective discrimination. At $\tau =1$ (linear weighting), the system struggles with poor selectivity across the document space, reflected in the lowest scores for precision and faithfulness. In contrast, $\tau =3$ effectively polarizes the query vector, ensuring that relevant knowledge nodes are prioritized to provide high-density context, which ultimately results in the highest BERTScore F1 (0.6740) and Answer Relevancy (0.6686). These results confirm that the selected hyper-parameters represent a stable high-performance equilibrium for the KG-Anchored RAG framework.

3.5. Case Study and Failure Analysis

To further understand the behavior of different retrieval strategies under professional QA settings, we conducted a qualitative analysis on representative failure cases.

3.5.1. Failure of Community-Based Retrieval

A representative example is the query:

“What is the definition of step voltage in electrical safety terminology?”

The correct answer explicitly requires the regulatory definition:

“the voltage between two points on the ground separated by 1 m (0.8 m in Chinese standards).”

However, both GraphRAG-Local and Hybrid retrieval fail to recover the exact clause containing the numerical constraint. Instead, they retrieve broad conceptual explanations related to grounding safety and electrical shock hazards.

This failure occurs because the relevant clause is embedded within a highly repetitive safety-regulation context. Under severe semantic redundancy, community-based retrieval tends to prioritize globally central descriptions rather than numerically precise technical definitions.

In contrast, the proposed KG-Anchored retrieval successfully activates the corresponding skeleton concepts associated with “electrical safety” and “voltage definition”, allowing the system to directly navigate toward the evidentiary clause instead of relying on neighborhood expansion.

3.5.2. Observed Limitations

Although the proposed framework substantially improves retrieval precision, several limitations remain. In terminology-definition tasks where the exact clause is absent from the retrieval pool, the model may still generate partially generalized explanations. This issue is particularly evident for highly ambiguous terms such as “real-time database” or “reclosing success”, where semantic overlap exists across multiple engineering subdomains.

These observations suggest that the effectiveness of the framework still depends on the completeness of the underlying skeleton construction and the coverage quality of the initial document chunking process.

4. Limitations

Although the proposed KG-Anchored RAG demonstrates strong performance in professional power-system QA tasks, several limitations remain. First, the current evaluation is conducted on a medium-scale proprietary dataset consisting of 8 regulatory documents and 122 document chunks. While the dataset contains highly specialized professional terminology, larger-scale validation across broader industrial corpora is still necessary. Second, the framework depends on the quality of the initial knowledge graph extraction process. Severe entity fragmentation or missing relations may negatively affect the construction of the topological skeleton and weaken the effectiveness of coordinate-based navigation. Third, the skeleton refinement strategy currently relies on manually designed statistical indicators such as degree centrality and inverse document frequency. More adaptive graph denoising mechanisms may further improve robustness under heterogeneous document distributions. Finally, the current study focuses primarily on Chinese power-system regulations translated into English prompts. The generalization capability of the framework under multilingual or cross-domain settings remains an important direction for future investigation.

5. Conclusions

In this paper, we presented KG-Anchored RAG, a problem-driven retrieval framework designed for structurally complex technical documents in the power domain. By identifying and addressing the limitations of current GraphRAG methods—specifically the issues of “Semantic Dilution” and “Topological Drift”—we successfully shifted the retrieval paradigm from coarse-grained community summarization to a precision-guided anchoring mechanism.

Our framework integrates a multi-stage optimization pipeline comprising Topological Skeleton Refinement, Semantic Bridge, Non-linear Sharpening, and PageRank-based Reranking. This architecture ensures that unstructured regulatory documents are systematically mapped onto a structured domain backbone through a latent topic-based Knowledge Attachment Matrix ($\mathbf{B}$). The introduction of the Softmax sharpening mechanism and PageRank-based algorithm further enhances the system’s ability to polarize navigational signals and identify semantic centers, effectively suppressing the noise generated by high-frequency generic technical terms.

Experimental results on a proprietary dataset of 8 power system professional documents (comprising 122 granular document units) demonstrate that the proposed method outperforms standard GraphRAG baselines. Our approach achieved a Context Precision of 0.4400 and a Faithfulness score of 0.3835, representing a more than twofold improvement in factual consistency and retrieval accuracy compared to GraphRAG-Local. Notably, the framework attained a ROUGE-2 F1 score of 0.0482, which is approximately 15.5 times higher than the baseline, confirming its superior ability to preserve rigid technical terminology. Furthermore, the system reached an nDCG@5 of 0.6480 using native matrix-driven navigation, proving that structural resonance can effectively prioritize evidentiary clauses without the need for computationally expensive external re-rankers.

The industrial implications of this work are important for safety-critical environments. By providing a high-fidelity and noise-resistant QA mechanism, the KG-Anchored RAG framework offers a reliable technical support tool for power system operators, reducing the risks associated with speculative hallucinations. Future research will focus on the scalability of this anchoring mechanism to cross-modal data, such as integrating equipment ledgers and real-time operational telemetry into the topological skeleton, as well as developing dynamic update capabilities for the attachment matrix.