A Knowledge Service System for Foundation Pit Engineering Based on LoRA and RAG
Система предоставления знаний для проектирования и строительства котлованов на основе LoRA и RAG
2026-01-01
SCID: 54.1/nguwv2g5
Discuss with AI
Foundation pit engineeringLarge language modelsLow-rank adaptation (LoRA)Regulatory knowledge serviceRetrieval-augmented generation (RAG)
Figures from the paper
Abstract (AI)
Foundation pit engineering involves a complex system of technical standards. Yet practitioners often struggle to retrieve relevant provisions and interpret them efficiently during construction decision‐making. To address these challenges, this paper presents a knowledge service system that integrates low‐rank adaptation (LoRA) with retrieval‐augmented generation (RAG). Based on 35 national and industry standards, we constructed a dataset of more than 8000 high‐quality question–answer pairs. LoRA was adopted for parameter‐efficient domain fine‐tuning, and RAG was incorporated to support real‐time retrieval and source attribution of regulatory provisions. In a comprehensive evaluation on 487 test items (including both subjective and objective questions), the system achieved an overall accuracy of 82.93%, gaining 16.38 and 11.15 percentage points over the pure fine‐tuned model and the pure RAG model, respectively. The average response time was 1.98 s, compared with 3.65 s for the pure RAG model, thereby satisfying the real‐time response requirements for on‐site decision support. This study developed a prototype system using open‐source frameworks, which supports offline inference and multiformat document processing in an experimental setting, thus demonstrating the feasibility and practical viability of the proposed architecture. This work offers a scalable knowledge service solution for foundation pit engineering and provides a generalizable framework for deploying large language models (LLMs) in other knowledge‐intensive engineering domains.
Key Findings
1
A foundation pit engineering knowledge service system integrates LoRA domain adaptation with RAG-based retrieval and regulatory source attribution.
2
Across 487 subjective and objective test items, the integrated system achieved 82.93% accuracy—16.38 and 11.15 percentage points higher than pure fine-tuning and pure RAG, respectively.
3
An open-source prototype supports offline inference and multiformat document processing experimentally, demonstrating feasibility and potential scalability for other knowledge-intensive engineering domains.
4
The authors constructed a dataset of more than 8,000 high-quality question–answer pairs from 35 national and industry standards.
5
The system reduced average response time to 1.98 seconds, compared with 3.65 seconds for pure RAG, meeting stated real-time on-site decision-support requirements.
Research Object
Knowledge service system for foundation pit engineering
Research Subject
Accuracy, retrieval effectiveness, response time, and practical viability of LoRA- and RAG-integrated regulatory knowledge support for construction decision-making
Publication Details
Publication Date
2026-01-01
Journal
Publisher
ISSN
Cited by
0
Open access PDF
Access Type
Author Information
Download PDF
Subscribe to digest
References available in scid.ai5
Large language models encode clinical knowledge2023
Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks2020
On the Opportunities and Risks of Foundation Models2021
A Survey on Trustworthy LLM Agents: Threats and Countermeasures2025
Domain adaptation of large language models for geotechnical applications2026