With the advancement of governance modernization, inspection and supervision have evolved from manual processes to data-driven and intelligent paradigms. In this context, integrating large language models (LLMs) into inspection workflows offers significant potential for improving policy retrieval, semantic understanding, issue identification, and report generation.

However, the direct deployment of general-purpose LLMs remains constrained by high computational cost, limited domain adaptation, and insufficient reliability in regulated environments. Meanwhile, existing inspection systems suffer from inefficient knowledge retrieval, inconsistent standards, heterogeneous multimodal data, and underutilization of historical information, which hinder accurate issue identification and effective rectification tracking.

Therefore, there is a pressing need for lightweight, domain-adaptive, and trustworthy LLM frameworks to support efficient, standardized, and compliant inspection and supervision processes.

This study aims to develop a lightweight and deployable LLM framework tailored to power enterprise inspection and supervision. The key objectives are as follows:

Efficient domain adaptation: leverage LoRA to achieve parameter-efficient fine-tuning, significantly reducing computational and storage requirements while maintaining performance.

Although the framework builds upon established components (LoRA, RAG, OCR, and prompting), its primary contribution lies in their unified integration into a lightweight, multimodal, and security-aware pipeline, specifically designed for real-world inspection scenarios.

Recent advances in large language models (LLMs) have enabled new applications in governance, compliance auditing, and enterprise supervision [1]. Pre-trained models such as GPT, PaLM, and LLaMA demonstrate strong capabilities in natural language understanding and generation [2], and achieve effective performance in tasks such as question answering and policy interpretation when combined with retrieval-augmented generation (RAG) and fine-tuning techniques [3].

However, their deployment in industrial settings remains constrained by high computational cost and limited adaptability to domain-specific and regulated scenarios. To address efficiency constraints, parameter-efficient fine-tuning (PEFT) methods, including Adapter, Prefix Tuning, and LoRA, have been proposed [4]. Among these, LoRA reduces trainable parameters via low-rank decomposition while preserving performance [5], and has shown effectiveness in several application domains [6]. Nevertheless, existing studies largely overlook requirements such as traceability, consistency, and compliance in regulated environments [7,8].

LLMs have also been applied to domain-specific tasks such as legal analysis, medical document generation, and smart grid forecasting [4,9,10]. While these approaches demonstrate strong task performance [5], they typically rely on single-modal inputs [6] and weakly constrained generation, limiting their suitability for inspection scenarios that require multimodal integration, standardized outputs, and evidence-grounded reasoning [7].

Overall, current research lacks a lightweight, deployable, and compliance-aware LLM framework for inspection and supervision. To address this gap, this study proposes a unified framework integrating LoRA, RAG, and structured prompting, together with a Sensitive Information Control Gate to ensure trustworthy and compliant generation in power enterprise inspection.

Large language models (LLMs) have been increasingly applied to governance-related tasks such as policy analysis, compliance monitoring, and audit assistance [11], demonstrating strong capabilities in information extraction, question answering, and document generation [12].

However, their direct application to inspection scenarios remains limited [13]. Inspection data are typically multimodal and highly domain-specific, including scanned documents, tables, and handwritten records. Moreover, inspection tasks require high factual reliability and auditability, while generative models are prone to hallucinations [14]. These challenges necessitate domain-adaptive and evidence-grounded frameworks for practical deployment.

As LLMs scale, full-parameter fine-tuning becomes computationally prohibitive. Parameter-efficient fine-tuning (PEFT) methods—such as adapters, prefix tuning, and LoRA—provide practical alternatives [15]. Among them, LoRA introduces low-rank decomposition to update a small subset of parameters while keeping the pretrained model frozen [14], significantly reducing memory and training cost. This makes LoRA particularly suitable for resource-constrained deployment scenarios.

However, PEFT alone does not address the need for evidence grounding and task-level consistency in inspection workflows, requiring integration with retrieval and prompting mechanisms.

Retrieval-augmented generation (RAG) enhances LLM performance by incorporating external knowledge during inference, improving factual accuracy and reducing hallucinations [15]. By grounding responses in retrieved evidence, RAG enables more reliable and verifiable outputs in domain-specific tasks [16,17].

Nevertheless, RAG introduces new challenges, including document segmentation, retrieval consistency, and evidence management, especially when dealing with multimodal inspection data. These limitations highlight the need for structured and controllable retrieval mechanisms.

Prompt engineering guides LLM behavior through structured instructions [18], including role assignment, templates, and reasoning strategies such as chain-of-thought [19]. While explicit prompts improve interpretability, they may be sensitive to wording and less stable in domain-specific tasks [20]; implicit methods offer robustness but lack transparency [21].

In inspection scenarios, prompts must ensure consistency, traceability, and alignment with retrieved evidence. Therefore, hybrid strategies that combine structured templates, role constraints, and evidence grounding are essential for reliable multi-task execution [22].

The workflow of LoRA-based adaptation is illustrated in Fig. 2, which consists of three stages: data preparation, parameter-efficient fine-tuning, and model compression.

The workflow of the RAG module is illustrated in Fig. 4, consisting of data processing, knowledge construction, and optimized retrieval. RAG enhances generation quality by grounding model outputs in externally retrieved evidence, thereby improving factual accuracy and reducing hallucinations in domain-specific inspection tasks.

To ensure structured, consistent, and evidence-grounded outputs, we design a task-oriented prompt formulation built on top of the RAG framework. The prompt is formalized as a four-element tuple:(14)𝒫=(ℛ,𝒯,E,ℱ)where ℛ is role specification, 𝒯 is task instruction, E is retrieved evidence from RAG, ℱ output format constraints.

Given a user query q, the model generates:(15)y=LLM(q,𝒫)

This formulation enforces role-aware reasoning, task consistency, and structured output generation.

To further reduce hallucination, explicit constraints are introduced:

evidence-grounded reasoning (outputs must be supported by E);

These constraints significantly improve factual consistency, as validated by the Faithfulness metric reported in Section 5.4.

To enforce security and regulatory compliance, we introduce a Sensitive Information Control Gate (SICG), positioned between the retrieval module and prompt construction.

Given a retrieved evidence chunk x, its sensitivity score is defined as:(16)s(chunk)=∑i=1Nwi⋅fi(chunk)where wi are non-negative weights satisfying ∑i=1Nwi=1 (calibrated by domain experts in accordance with enterprise security policies), and fi(chunk) denotes the individual sensitivity feature functions. These functions include normalized binary indicators for the presence of personal identifiers, financial or operational confidentiality markers, and policy-restricted keywords, each scaled to the interval [0, 1].

Each user role is assigned a clearance level c(user)∈[0,1], with higher values reflecting greater access privileges (e.g., 0.3 for field inspectors, 0.7 for supervisors, and 1.0 for administrators). The filtering decision for each chunk is governed by the following threshold rule:(17)output_chunk={redacted(chunk)ifs(chunk)>θ⋅c(user)chunkotherwisewhere θ∈[0.5,1.0] is a tunable security threshold (default value θ=0.75) that balances protection and usability. The redaction operation replaces sensitive spans with the placeholder “[REDACTED]” while preserving contextual coherence and source traceability metadata.

By combining sensitivity scoring, role-based thresholds, and structured redaction, SICG effectively mitigates information leakage risks while preserving traceability and usability in regulated inspection environments.

To quantitatively evaluate the effectiveness of LoRA fine-tuning, we compared the base Qwen3-8B model with the domain-adapted model on twelve representative inspection and supervision tasks. As shown in Fig. 6a,b, the base model tends to produce generic or outdated responses, whereas the LoRA fine-tuned model exhibits markedly improved domain alignment, policy accuracy, and standardization of inspection outputs.

All reported metrics were recomputed using bootstrap resampling (1000 iterations) to obtain 95% confidence intervals (95% CI). Comparisons between configurations were performed using paired t-test on question-level differences, with p < 0.01 considered statistically significant. These visual demonstrations are further supported by the multi-dimensional evaluation results in Section 5.4 (Figs. 7 and 8) and the full held-out test set performance in Table 1.

The data ingestion process converts heterogeneous inspection and supervision data—including policy documents, reports, and rectification records—into structured knowledge using OCR and text preprocessing. The processed content is embedded and stored in a vector database for efficient retrieval.

During query execution, the RAG module retrieves relevant evidence and generates responses grounded in source documents. As illustrated in Fig. 6c, each retrieved result is accompanied by metadata and evidence tuples, ensuring full auditability and traceability. This mechanism significantly enhances transparency and supports compliant inspection and supervision workflows by linking generated outputs directly to authoritative sources.

In summary, integrating data ingestion with RAG-based querying establishes an efficient, transparent, and fully verifiable intelligent data processing pipeline. Further validation of the RAG configuration is provided through ablation experiments in Section 5.7.

The domain-specific model, enhanced by LoRA fine-tuning, RAG, and structured prompting, accurately retrieves up-to-date regulations and domain-specific requirements. It provides superior support for policy interpretation, issue identification, and supervisory decision-making in power enterprise inspection and supervision scenarios. Quantitative evaluation of the complete framework is presented in the subsequent sections.

To further evaluate the framework, twelve representative types of problems covering different task categories (policy interpretation, issue identification, rectification recommendations, and compliance tracking) and difficulty levels (easy, medium, and hard) were used for automated testing. All quantitative comparisons, including variance, confidence intervals, significance tests, and repeated-run stability, are reported in Section 5.4.

To ensure statistical robustness and generalizability, the proposed framework was additionally evaluated on the full held-out test split of the instruction-tuning dataset, which contains 687 unseen questions. Using the same LoRA-adapted model with the complete RAG + four-element prompt engineering pipeline, we automatically computed key semantic metrics (Relevance, Correctness, Faithfulness, and Retrieval F1) across all 687 test samples. Results on this larger, independent test set are summarized in Table 1. The 12 representative types of problems were retained for in-depth multi-dimensional analysis, including generation quality (BLEU/ROUGE), response efficiency, task-specific breakdowns, and direct comparison with ablation studies (Sections 5.7–5.9). This dual-evaluation strategy—broad automatic assessment on the full 687-question held-out set combined with detailed qualitative and multi-metric analysis on the 12 representative types of problems—provides strong evidence of both generalization capability and practical robustness in real-world power enterprise inspection and supervision scenarios. All metrics were computed with bootstrap resampling (1000 iterations) to obtain 95% confidence intervals; comparisons between configurations were performed using paired t-test (p < 0.01), ensuring reported variance, statistical significance, and repeated-run stability. Screenshots in Fig. 6 serve solely as supplementary visual aids; the primary evidence consists of the tables and statistical tests presented herein. Each input was processed end-to-end using the complete framework (LoRA fine-tuned model + optimized RAG retrieval + four-element prompt).

The evaluation combines generation metrics (BLEU, ROUGE-1/2/L) and semantic metrics (Retrieval F1, Relevance, Correctness, and Faithfulness). The semantic metrics are defined as:(18)si=cos⁡(emb(y^i),emb(yi)),s^i=si+12 (19)ki=|K(y^i)∩K(yi)||K(yi)| (20)Relevancei=s^i,Correctnessi=0.6s^i+0.4ki,Faithfulnessi=ki

All raw metrics were min–max normalized for comparability:(21)m~i=mi−minjmjmaxjmj−minjmj,M=112∑i=112m~i

The model achieves scores between 0.72 and 0.87, with Relevance 0.87 (95% CI: 0.84–0.90) and Faithfulness 0.86 (95% CI: 0.83–0.89) reaching the excellent range. Retrieval F1 0.81 (0.78–0.84) and Correctness 0.83 (0.80–0.86) confirm strong factual alignment (all bootstrap 1000 iterations). These results are statistically robust and significantly outperform the zero-shot baseline (paired t-test, p < 0.01, see Section 5.1). These results are summarized in Fig. 7, and the multi-dimensional evaluation is shown in Fig. 8.

It is worth noting that the Faithfulness metric was computed under the complete framework (LoRA + RAG + four-element prompt), where the model is explicitly constrained to rely only on provided evidence. This built-in safeguard addresses potential robustness issues arising from imperfect retrieval, which is particularly important for regulated power enterprise inspection tasks. Although a dedicated adversarial benchmark was not conducted due to deployment-environment constraints (see Section 5.6), the achieved Faithfulness of 0.86 demonstrates effective mitigation of hallucination risks in practice [20,21].

These outcomes demonstrate that the proposed LoRA-enhanced lightweight LLM with RAG and prompt optimization mechanisms achieves balanced excellence between generation quality, semantic faithfulness, and computational efficiency. The consistent performance across all twelve evaluation items verifies the model’s robustness and practical value for real-world power enterprise inspection and supervision tasks.

To demonstrate the practical impact of the Sensitive Information Control Gate, comparative experiments were conducted on representative inspection and supervision queries containing sensitive data (e.g., personnel records, internal audit findings, and proprietary operational details). Performance was evaluated using three key metrics: leakage rate (percentage of sensitive entities exposed in generated outputs), compliance score (0–100 scale, assessed by domain experts against enterprise data security policies), and average response time.

As summarized in Table 2, the gate reduces the leakage rate from 12.5% to 0.8% while improving the compliance score from 78 to 96, with only a negligible overhead of 0.2 s in response time.

A qualitative case study further validates the gate’s effectiveness. For a high-sensitivity query requesting detailed findings from an internal personnel audit, the model without the gate inadvertently exposed employee identifiers and confidential financial figures. With the gate enabled, all sensitive segments were automatically replaced with “[REDACTED]” placeholders, while non-sensitive evidence-supported recommendations remained intact and fully verifiable.

Role-based access control performance is detailed in Table 3, confirming consistent enforcement across user privileges.

These results confirm that the Sensitive Information Control Gate not only closes critical security gaps identified in the literature but also preserves the framework’s efficiency and usability in real-world power enterprise deployment.

Due to the limited computational resources available in our experimental setup (matching the deployment environment of Guangdong Power Grid workstations), systematic comparisons with other PEFT methods (Adapter, Prefix Tuning, QLoRA) and full-parameter fine-tuning were not performed. Nevertheless, our LoRA-adapted model significantly outperforms the base Qwen3-8B, aligning with extensive prior benchmarks. These studies consistently show that LoRA achieves 95%–99% of full fine-tuning performance with <1% trainable parameters and 32%–44% shorter training time [19,20]. This literature-supported efficiency validates our framework as a practical, lightweight solution for power enterprise inspection and supervision.

To address the reviewer’s concern regarding the lack of systematic validation of the RAG module, we conducted additional pilot ablation experiments on a representative subset of the power enterprise inspection corpus (200 policy documents and inspection reports). These experiments evaluate the impact of different retrievers (BM25 vs. dense embeddings), chunking strategies, and top-k values on key end-to-end metrics. The results are summarized in Table 4.

The statistical analysis indicates significantly better than all other configurations (paired t-test, p < 0.01). All CIs from 1000 bootstrap iterations (see Section 5.1).

The current configuration achieves the best overall balance between retrieval quality (highest Relevance and Faithfulness) and practical inference efficiency. BM25, while faster, shows significantly lower semantic understanding on Chinese regulatory and technical documents, consistent with findings in domain-specific Chinese RAG benchmarks. Larger or smaller chunks and extreme top-k values either introduce noise or reduce recall, confirming that the chosen 600-token QA-oriented splitting with hybrid re-ranking is optimal for power enterprise inspection tasks.

This ablation, combined with the strong end-to-end performance already reported in Figs. 7,8 and Table 1, demonstrates that the RAG module is not only effective but also deliberately tuned for the target deployment environment. Comprehensive grid-search over all possible retrievers and hyperparameters was omitted to maintain strict alignment with real-world hardware constraints, as explained in Section 5.6 for the LoRA component.

To address the reviewer’s concern regarding the vagueness of the prompt engineering description, we conducted additional pilot ablation experiments on the same 12 representative inspection and supeervision tasks used in Section 5.4. These experiments compare the proposed four-element template (role-based + structured output + Chain-of-Thought reasoning) against four common alternatives. All tests used the same LoRA-adapted model and RAG configuration. Results are summarized in Table 5.

The proposed four-element template achieves the statistically best balance (paired t-test, p < 0.01) across all semantic metrics while maintaining practical inference speed. Role assignment and structured output format significantly reduce hallucinations and improve traceability, which are critical for regulatory compliance in power enterprise inspection. Pure CoT or few-shot strategies improve reasoning but lack the explicit evidence grounding and output constraints required by the task.

To validate the chosen quantization strategy, we conducted additional pilot ablation experiments on the same 12 representative inspection and supervision tasks used in Section 5.4. These experiments compare the chosen q8_0 quantization against the original bf16 (non-quantized) model and two lower-bit variants (q5_0 and q4_0). All tests used the same LoRA-adapted Qwen3-8B model and RAG configuration on the target 32 GB vGPU hardware. Results are summarized in Table 6.

The q8_0 configuration achieves the best overall balance: it delivers the reported ~50% storage reduction, improves inference speed by 18%, and causes only negligible degradation in semantic metrics (Δ ≤ 0.01–0.02). Full distributional statistics for response times (Mean ± SD, Min–Max, and 95% bootstrap CI) under different quantization levels are visualized in Fig. 9. As shown in the box plots, the q8_0 model not only reduces the median inference time compared with the bf16 baseline but also exhibits substantially lower variance and fewer outliers. This indicates more stable and predictable latency under varying RAG top-k settings, making q8_0 particularly suitable for deployment on resource-constrained hardware while preserving high Relevance, Faithfulness, and Correctness scores. All measurements were performed on the same 32 GB vGPU hardware used in the target deployment environment.