1. Abstract

Despite the public availability of Ethiopia’s Constitution and core legal codes, their complexity and formality make them difficult to understand and apply in real-life situations. This project presents a conversational Legal AI Assistant that simplifies and contextualizes Ethiopian law through semantic search, article-cited reasoning, and user-friendly interaction. Powered by LangChain, ChromaDB, HuggingFace embeddings, Groq-hosted LLaMA 3, and Streamlit, the assistant explains legal provisions in plain language, interprets case scenarios, and offers grounded insights into potential consequences—bridging the gap between legal text and practical understanding.

---

2. Problem Statement

While Ethiopia’s legal codes are public and searchable, applying them meaningfully to specific real-world situations remains complex. For example, determining the potential penalty for a hit-and-run while intoxicated requires cross-referencing the Criminal Code’s traffic, homicide, and sentencing provisions. Few tools exist today that connect a case narrative—its context, evidence, and relevant laws—to a grounded legal interpretation.

Traditional databases provide raw text search but lack the ability to process full case context or offer structured reasoning. Meanwhile, generic AI models may hallucinate answers or misrepresent punitive statutes, undermining user trust in legal applications. There’s a gap between legal literacy and legal applicability—especially when users seek not just what the law says, but what it might mean for them in practice.

This project addresses that gap by delivering a context-sensitive legal assistant that reasons from Ethiopia’s laws to real-world outcomes. Whether asked to explain due process rights or to assess potential sentences from a case description, the assistant provides grounded, article-cited responses. In doing so, it transforms passive access into proactive understanding—where the law begins to answer back.

---

3. Methodology

This project combines retrieval-augmented generation (RAG), structured prompt engineering, and vector-based search to deliver grounded legal reasoning through natural-language interaction. The assistant is trained on the Ethiopian Constitution, Civil Code, and Criminal Code, and is engineered to interpret user queries and case descriptions while maintaining legal accuracy and relevance.

3.1 Architecture Overview

The system is composed of four main layers:

• Document Store

  • sentence-transformers/all-MiniLM-L6-v2 for semantic embeddings
  • ChromaDB as the vector store
  • Legal documents chunked and tagged by article and section

• LLM Reasoning Engine

  • Groq-hosted LLaMA 3 (8B) via langchain_groq.ChatGroq
  • Deterministic prompt execution and fast response

• Prompt Layer (Chain of Thought for Law)

  • Structured to guide article-cited, step-by-step legal reasoning
  • Detects case-based inputs and infers plausible legal consequences
  • Includes memory context and retrieval context in every call

• Interfaces

  • Streamlit UI for end-user interaction
  • FastAPI backend for endpoint routing
  • Command-Line Interface (CLI): for offline or developer testing

3.2 Data Preparation

• Source: publicly available government and legal repositories
• Cleaned for OCR and encoding artifacts
• Chunked using article-aware logic with summary_title metadata
• Embedded and stored in ChromaDB for semantic retrieval

3.3 Legal QA Flow

1. User submits question or case
2. Query embedded and relevant chunks retrieved
3. Prompt constructed with:
   • Retrieved context
   • User question
   • Chat history (if available)
4. LLM generates grounded answer with citations
5. UI renders explanation + referenced articles

3.4 System Architecture Diagram

The diagram below illustrates how the assistant is structured across four layers—data ingestion, retrieval, reasoning, and interaction:

3.5 Installation & Usage

To run the Legal AI Assistant locally:

Installation

git clone https://github.com/Abenezer-Baheru/Legal_AI_Assistant.git
cd Legal_AI_Assistant
pip install -r requirements.txt

Create a .env file with the following variables:

API_KEY=your_groq_key

Usage

You can interact with the assistant in three modes:

A. Command Line Interface (CLI)
Run the assistant directly in your terminal:

python app/query_and_answer.py

B. User Interface (UI)
FastAPI Backend
Start the API server for LLM-powered legal QA:

uvicorn app.main:app --reload

This launches the backend at http://localhost:8000, which powers the frontend.

Streamlit UI
In a separate terminal, launch the web interface:

streamlit run frontend/streamlit_app.py

This opens the assistant in your browser at http://localhost:8501, where you can chat using a clean, interactive interface.

4. Results & Insights

The Legal AI Assistant was tested locally on five foundational use cases, confirming its ability to reason accurately across key areas of Ethiopian law. While formal deployment and benchmarking are still in progress, early results indicate strong alignment between user input and relevant legal responses.

4.1 Local Functionality Tests

Core capabilities were validated:

1. Constitution-Based Queries
   Accurately responds to questions involving constitution.
   

2. Civil Code-Based Queries
   Successfully interprets obligations and property law.
   

3. Criminal Code-Based Queries
   Maps case scenarios to relevant offenses and penal provisions.
   

4. Out-of-Context Questions
   Detects when legal context is insufficient and responds appropriately.
   

5. Case-Based Legal Reasoning
   Synthesizes law and facts to infer potential punishments with proper grounding.
   

4.2 Observations

• Maintains high legal fidelity and avoids hallucination
• Prompt design yielded conversational and reasoned responses
• Local CLI testing proved smooth and responsive, with fast iteration

---

5. Scope of Legal Queries and Future Directions

5.1 Scope of Legal Queries

The Legal AI Assistant is designed to answer questions grounded in Ethiopia’s Constitution, Civil Code, and Criminal Code. Below is a breakdown of supported and unsupported query types, along with examples.

5.2 Future Directions

The Legal AI Assistant for Ethiopia shows strong promise as a foundational tool for civic education and legal understanding. Planned enhancements include:

---

6. References

• FDRE Constitution of Ethiopia (1995)
• Ethiopian Civil Code (1960)
• Ethiopian Criminal Code (2004)
• Sources retrieved from lawethiopia.com and publicly accessible government databases

---

7. Project Repository

Explore the full source code, CLI tools, UI, and model configuration on GitHub:
github.com/Abenezer-Baheru/Legal_AI_Assistant

Maintenance Status: Active — under ongoing development and refinement
Support: For issues or feature requests, please open a GitHub issue or contact via ReadyTensor profile.