A comprehensive exploration of legal AI systems tailored for the Sri Lankan legal context, combining SLMs, RAG frameworks, and agentic workflows.
Background & Literature Survey
The legal system plays a critical role in maintaining justice, fairness, and social order. In Sri Lanka, legal knowledge is confined to professionals or documented in complex texts, creating significant barriers for ordinary citizens. Labour disputes, property transfers, family disputes, and criminal litigation all require specialized understanding that most citizens lack.
Recent advancements in AI, particularly Natural Language Processing (NLP), offer promising solutions. The digitization of court judgments and legal documents has created opportunities for computational analysis. Research in Legal NLP has evolved from keyword-based systems to sophisticated transformer-based models like BERT, LEGAL-BERT, and domain-specific LLMs.
However, most existing systems — LawLLM (US), LawGPT (China), Swiss-BERT variants — are jurisdiction-specific and computationally expensive, limiting their applicability to Sri Lanka's unique legal ecosystem, which blends Roman-Dutch law, English common law, and customary traditions.
Key Findings from Literature
LLM-based approaches demonstrate strong reasoning but suffer from jurisdictional overfitting, high computational cost, and hallucination risks. They are not directly transferable to Sri Lanka.
RAG-based systems improve factual grounding but often lack structured output generation, validation mechanisms, and user-friendly interfaces essential for non-expert users.
Small Language Models offer a compelling balance — lower computational overhead, efficient fine-tuning via LoRA/QLoRA, and strong domain adaptation capabilities when trained on curated legal datasets.
| Feature | Quick Check | LawRec | Legal Query RAG | Our System |
| Transformer Models | No | Yes (BERT) | Yes | Yes |
| RAG Integration | No | No | Yes | Yes |
| Sri Lankan Focus | No | No | No | Yes |
| Natural Language Queries | Partial | Partial | Yes | Yes |
| Structured Legal Output | No | No | Partial | Yes |
| Scalability | Medium | Medium | Medium | High |
Identified Research Gaps
Despite significant advances in legal AI globally, critical gaps remain for the Sri Lankan context:
- No fine-tuned transformer models specifically for Sri Lankan Labour, Property, Family, or Criminal Law
- Absence of RAG frameworks adapted to Sri Lankan legal datasets
- No structured criminal outcome classification studies for Sri Lanka
- Inability of existing systems to process natural language queries with high precision in local context
- Lack of structured legal output (Act name, Section, Year, Case references)
- No scenario-based or step-by-step legal explanation for non-experts
- Absence of deed verification systems tailored to Sri Lankan property documents
- Limited accessibility for non-expert users due to complex interfaces
Jurisdictional Concentration Problem
Most Legal NLP research concentrates on the United States Supreme Court, European Court of Human Rights, Chinese criminal courts, and Swiss Federal Supreme Court. These systems benefit from well-digitized databases and large labeled datasets. Sri Lanka, with its hybrid Roman-Dutch and English common law system, represents a significantly underexplored jurisdiction with unique challenges: limited digitized data, multilingual content (Sinhala, Tamil, English), inconsistent document formats, and no standard benchmark datasets.
Research Problem Statement
Despite the increasing need for efficient and accessible legal information systems, Sri Lanka currently lacks AI-driven legal frameworks that integrate modern NLP, transformer-based models, and Retrieval-Augmented Generation — specifically tailored for its unique legal domains.
This absence creates significant barriers to legal accessibility, reduces efficiency in legal research and decision-making, and contributes to inequality in access to legal knowledge among citizens and professionals. A survey of 40 participants (lawyers, law students, general public) revealed:
- 90%+ identified the need for an intelligent legal decision-support system
- Significant delays in accessing relevant legal documents
- Difficulty interpreting fragmented and technical legal language
- Lack of user-friendly systems supporting natural language queries
Main Objective
To develop specialized AI-based legal assistance systems for Sri Lankan law domains that provide reliable, context-aware, and structured legal guidance — combining fine-tuned Small Language Models with retrieval-augmented mechanisms grounded in authoritative legal sources.
Specific Objectives
- Transform unstructured Sri Lankan legal resources into structured, machine-readable training datasets
- Fine-tune transformer-based SLMs (Qwen3-8B, LEGAL-BERT-SMALL) on domain-specific datasets
- Design consistent structured output formats ensuring legal elements (Act, Section, Year) are always present
- Integrate RAG frameworks grounding outputs in verified legal documents via FAISS vector search
- Enable practical recommendations for real-world workplace, property, family, and criminal disputes
- Develop scalable deployment pipelines ensuring low-latency, production-level accessibility
- Build user-friendly web interfaces accessible to non-expert users
System Methodology Overview
All four research components follow a unified, multi-layered methodology that integrates legal data engineering, model adaptation, retrieval design, system integration, and rigorous evaluation. The Agile development framework enables iterative improvement with measurable artifacts at each stage.
Phase 1 — Data Collection & OCR Processing
Weeks 1–4
Collection of legal materials from digital repositories, law books, and physical archives. OCR-based digitization of scanned documents with quality scoring. Multilingual handling (Sinhala, Tamil, English).
Phase 2 — Dataset Construction & Governance
Weeks 5–8
Cleaning, normalization, and JSONL formatting. Schema validation ensuring consistent instruction-context-output structure. Train/validation/test splitting with leakage prevention.
Phase 3 — Model Fine-Tuning
Weeks 9–14
LoRA/QLoRA-based fine-tuning using Unsloth. Domain adaptation for Qwen3-8B (legal recommendation), LEGAL-BERT-SMALL (criminal prediction). Structured output alignment training.
Phase 4 — RAG & Vector Indexing
Weeks 13–18
FAISS index construction from legal document embeddings. Document-diverse reranking. Agentic RAG with LangGraph orchestration: classify → retrieve → grade → generate → validate.
Phase 5 — System Integration & Evaluation
Weeks 19–24
FastAPI backend with modular microservices. React frontend. Multi-layer evaluation (model-level, retrieval-level, system-level). End-to-end testing and iterative refinement.
Core Technologies
The research employs a carefully selected technology stack balancing capability, efficiency, and deployability.
Qwen3-8BLEGAL-BERT-SMALL
LoRA / QLoRAUnsloth
FAISS Vector DBRAG Framework
LangGraphFastAPI
PostgreSQLReact 18
Sentence TransformersPyTorch
HuggingFaceModal (GPU)
OllamaOCR Pipeline
Gemini EmbeddingsAdamW Optimizer
TanStack QueryTailwind CSS
PydanticLangChain
