Abstract

Ethiopia’s tax regulations—including VAT, turnover tax, and income tax—are publicly available but remain inaccessible to non-specialists due to legal jargon, scattered formats, and a lack of contextual guidance.

This project introduces a Retrieval-Augmented Generation (RAG)-based chatbot that transforms natural-language questions into grounded, clause-cited answers from Ethiopian tax documents. Powered by LangChain, ChromaDB, sentence-transformer embeddings, and Groq-hosted LLaMA3, the system delivers precise, legally faithful responses in under a second.

By combining semantic search with generative reasoning, the assistant bridges the gap between complex legal frameworks and everyday business needs—from determining VAT thresholds to clarifying exemptions—empowering SMEs, freelancers, and legal professionals.

Problem Statement

While Ethiopia’s tax codes are technically accessible, navigating them requires legal expertise, cross-referencing multiple documents, and interpreting complex clauses.

Professionals, small business owners, and freelancers often struggle to answer critical questions such as:

• What VAT applies to digital services?
• When is income tax due for freelancers?
• Who qualifies for turnover tax exemptions?

Challenges with existing solutions:

• Standard search engines lack clause-level precision.
• Generic AI models risk hallucinating or misrepresenting laws.
• No existing tool is tailored to Ethiopian tax compliance with verifiable citations.

This project directly addresses these gaps by providing a tax-specific AI assistant that retrieves authoritative sources and generates reliable, cited explanations.

Methodology

Architecture Overview

---

Data Pipeline

1. Document Loading – VAT, turnover tax, and income tax PDFs are cleaned and parsed.
2. Chunking—LangChain’s RecursiveCharacterTextSplitter divides legal text into metadata-rich sections, preserving article references.
3. Embedding & Storage – Chunks embedded via all-MiniLM-L6-v2 and stored in ChromaDB.
4. **RAG Flow – The user query is embedded, relevant chunks are retrieved, and LLaMA3-8B generates grounded answers.
5. Interfaces – FastAPI backend powers /ask; the Streamlit frontend enables user interaction.

Results

Local Testing

Performance:

• Average latency: ~500ms via Groq-hosted LLaMA3-8B.
• No hallucinations observed in controlled tests.
• Citation links increased trust and usability.

Observations

• Fast and grounded responses (~500ms latency via Groq)
• Consistent legal fidelity (no hallucinations observed in controlled prompts)
• Clear citation links improved user trust
• Docker setup enabled reproducible results across machines

Evaluation

• Qualitative Testing – Real-world queries from business forums and SME support channels.
• Accuracy Checks – Cross-referenced with official proclamations.
• Latency Measurements – Consistently under 1 second.
• Robustness Tests – Edge-case handling with safe fallback responses.

Dataset Description

Primary sources include:

• Turnover Tax Proclamation No. 308/2002
• Income Tax Proclamation No. 979/2016
• VAT Proclamation (2023 Draft)
• Commercial Code of Ethiopia
• Customs Guide
• The Rights of Taxpayers under Ethiopian Law

All documents were obtained from public Ethiopian government portals and processed into machine-readable format.

Monitoring & Maintenance

The assistant is designed for long-term legal reliability through:

• Automated re-embedding when new laws are published.
• Planned admin dashboard for query logging and analytics.
• Fallback messaging for low-confidence answers.
• Future integration of user feedback for reinforcement learning.

Comparative Analysis

Limitations

• Currently English-only (Amharic support in progress).
• Static legal sources until updated manually.
• Not a certified legal advisor.
• Limited quantitative accuracy metrics so far.

Future Directions

• Amharic LLM integration for bilingual responses.
• Admin Dashboard for analytics and monitoring.
• User Feedback Loop to refine embeddings and responses.
• Compliance simulation to predict tax outcomes based on user scenarios.
• Public API access for integration into legal portals and SME platforms.

References

Legal Documents Used

• Commercial Code of Ethiopia
• The Right of Taxpayers under Ethiopian Law
• Income Tax Regulation
• Taxation in Ethiopia
• Proclamation No. 979/2016 – Income Tax Proclamation
• Turnover Tax Proclamation No. 308/2002
• VAT Proclamation – Latest Draft (2023, English Version)
• Ethiopia Customs Guide
• Ethiopian Tax Law Summary
• General Tax Law

Conclusion

The Ethiopian Tax Law Assistant demonstrates how retrieval-augmented generation can transform legal accessibility in developing countries. By merging semantic search, clause-level citations, and low-latency generation, the project offers SMEs, freelancers, and legal professionals an accurate, trustworthy, and fast way to navigate complex tax laws.

The early results—near-instant responses with 100% legal fidelity in controlled tests—show its potential as a foundation for broader legal tech innovation in Ethiopia and beyond.

Project Repository
The full source code, UI scripts, Docker configs, and prompts are available here:
https://github.com/Filimon-H/ethiopian-tax-law-assistant

How to Reproduce

Prerequisites

• Python 3.10+
• Groq API key (get one here)
• Docker (optional for containerized setup)

🔧 Local Development Steps

# Step 1: Clone the repository and enter the project directory
git clone https://github.com/your-username/ethiopian-tax-law-assistant.git
cd ethiopian-tax-law-assistant
# Step 2: Add your API key
cp .env.example .env
# Then open .env and add:
# GROQ_API_KEY=your_key_here
# Step 3: Install Python dependencies
pip install -r requirements.txt
# Step 4: Load API keys into environment
python src/config/load_config.py
# Step 5: Load and chunk documents
python src/ingestion/document_loader.py
# Step 6: Embed and store text chunks into vector DB
python src/embedding/embed_store.py
# Step 7: Start FastAPI backend
python src/api/main.py
# Step 8: In a separate terminal, launch the Streamlit UI
streamlit run streamlit_app/app.py
🐳 Dockerized Execution
docker-compose build
docker-compose up
Access the UI at: http://localhost:8501