How RAG and LoRA Empower Commercial Banks to Build Intelligent Legal Review Systems

In commercial banking compliance, legal review and popularisation are core risk‑control functions. The Postal Savings Bank processes over 120,000 contracts annually with more than 2 million approval flows, and traditional manual review faces efficiency bottlenecks and error risks.

To address these challenges, the bank’s software R&D centre built a legal‑domain large model on a self‑controlled base LLM, applying fine‑tuning, Retrieval‑Augmented Generation (RAG) and Low‑Rank Adaptation (LoRA) to achieve professional, accurate and up‑to‑date knowledge.

Fine‑tuning data preparation followed three steps:

Collect internal legal rule libraries to ensure correct interpretation of statutes.

Gather diverse contract texts (commercial, labour, lease, etc.) together with corresponding legal clauses and risk points.

Organise historical review records and common audit questions with expert‑verified opinions.

These high‑quality datasets were used to fine‑tune base models such as DeepSeek R1 and ChatGLM‑4, enabling the model to understand banking‑specific documents and suggest risk‑mitigation advice.

RAG‑enhanced retrieval solves hallucination by grounding answers in a structured knowledge base. The pipeline first preprocesses legal documents (OCR, layout analysis, semantic segmentation, metadata extraction, dual keyword‑and‑vector indexing) and then performs multi‑path recall: keyword search (e.g., Elasticsearch/BM25) for exact term matches and semantic vector search for intent matching. Retrieved fragments from statutes, case libraries and template repositories are merged into a “enhanced prompt” that guides the LLM to generate answers with citations.

LoRA incremental pre‑training updates the model’s legal knowledge with minimal compute. By freezing the base model and training only low‑rank adapters, the bank quickly incorporated the 2024 revision of the National Secrets Law , allowing the model to answer new regulatory questions accurately.

Contract review deployment integrates the model into the bank’s knowledge‑management and legal‑affairs systems, forming a closed loop: upload → model analysis → opinion generation → human verification. The workflow reduced average review time from 1.6 days to 0.73 days (over 50 % speedup) and unified risk‑assessment language across branches.

Intelligent popularisation tool extends the model to an employee‑facing Q&A service. It delivers precise legal answers (e.g., electronic‑signature conditions) with clause citations within 10 seconds, and automatically triggers LoRA updates when new regulations arrive. By the time of writing, the tool answered more than 3 000 queries, reducing reliance on on‑site legal staff.

The project demonstrates direct cost savings, risk mitigation and industry leadership. Future work will focus on model distillation and data augmentation to lower deployment resources, expanding scenarios to litigation assistance and complaint risk assessment, and enhancing system compatibility, security and usability.