Korean Tax AI with Graph-RAG
Production-ready Python library for automated Korean tax law Q&A with mandatory legal citations and complete audit trails. Trusted by enterprises for compliance, tax consulting, and knowledge base automation.
Core Capabilities
Enterprise-ready features for legal AI automation
Graph-RAG Architecture
Neo4j 5.x powered semantic graph with Cypher queries for precise legal relationship mapping, hierarchical reasoning, and multi-hop inference across 10,000+ tax articles.
Vector Search
Qdrant 1.7+ with 1536-dimensional embeddings (OpenAI text-embedding-3-small) for semantic retrieval, achieving 95%+ relevance with HNSW indexing and cosine similarity.
Mandatory Citations
Automatic legal reference attachment ensuring compliance and auditability for all responses.
Audit Trail
Complete reasoning transparency with JSON-serialized decision tracking, retrieval scores, LLM prompts/responses, and inference path documentation for compliance audits.
Multi-Tax Coverage
Corporate, value-added, and income tax support with 10+ years of historical legislation data.
Multiple Interfaces
CLI, REST API, and Python library for seamless integration into existing workflows.
Technology Stack
Built with modern AI and enterprise infrastructure
Python Documentation
Getting Started (5 Minutes)
NEW! Zero to first query in 5 minutes. Demo mode, no setup required!
Introduction
Project overview, architecture principles, and key concepts for Korean legal AI.
Quick Start
Installation, three operation modes, and production setup guide.
Real-World Use Cases
NEW! 5 production examples: chatbots, knowledge bases, compliance tools.
Configuration & Costs
Complete settings, cost breakdown ($42-$4,400/mo), and performance metrics.
Troubleshooting
NEW! Self-service solutions for common errors and performance issues.
Code Examples
16 production-grade examples for tax queries, batch processing, and integrations.
Architecture
System design, data flow diagrams, and component interaction documentation.
Engine API
Complete TaxiaEngine API reference with all methods, parameters, and return types.
REST API
HTTP endpoints, request/response schemas, and authentication documentation.
CLI Tool
Command-line interface reference with all available commands and options.
PyPI Package
Install taxia-core from PyPI, view version history, and check dependencies.
GitHub Source
Complete source code, examples, and contribution guidelines.
Data Sources
Tax law data structure, sources, and distribution information (2015-2025).
JavaScript SDK
Switch to JavaScript documentation for React, Vue, Next.js integrations.
Complete Usage Example
from taxia import TaxiaEngine, TaxiaConfig
from taxia.types import QueryResponse
# Configure with custom settings
config = TaxiaConfig(
llm_provider="claude", # or "openai"
model_name="claude-3-5-sonnet-20241022",
temperature=0.0, # Deterministic for legal
max_tokens=4096,
retrieval_k=5, # Top-5 similar documents
)
# Initialize engine
engine = TaxiaEngine(config)
# Execute semantic query
response: QueryResponse = engine.query(
"2024년 법인세 세율은 과세표준에 따라 어떻게 달라지나요?"
)
# Access structured response
print(f"Answer: {response.answer}")
print(f"Citations: {response.citations}") # List[Citation]
print(f"Confidence: {response.metadata.confidence}")
print(f"Retrieval Time: {response.metadata.retrieval_time_ms}ms")
# Error handling
try:
result = engine.query("복잡한 세법 질의")
except ValidationError as e:
print(f"Invalid query: {e}")
except TimeoutError:
print("LLM timeout (>30s)")
⚡ Performance Metrics
- Query Latency: 2-5s average (includes LLM inference)
- Vector Search: <150ms for top-5 retrieval from 10K+ docs
- Graph Traversal: <100ms for multi-hop reasoning
- Memory Usage: ~500MB base + 2GB for vector index
- Concurrency: Async support for 100+ concurrent queries
System Requirements & Deployment
🐍 Python Environment
- ✓ Python 3.9, 3.10, 3.11, 3.12
- ✓ pip 23.0+ or Poetry 1.7+
- ✓ Linux/macOS/Windows (x64)
- ✓ Virtual environment recommended
🗄️ Infrastructure
- ✓ Qdrant: 2GB RAM, 10GB disk
- ✓ Neo4j: 4GB RAM, 20GB disk
- ✓ Optional: Docker Compose setup
- ✓ OpenAI/Claude API key required
⚡ Performance
- → 4-core CPU (8+ recommended)
- → 8GB RAM minimum, 16GB ideal
- → SSD for vector index
- → 1Gbps network for LLM API
⚠️ Important Limitations
- Legal Scope: Korean tax law only (Corporate, VAT, Income Tax 2015-2025)
- Language: Korean queries required (English answers via translation)
- Rate Limits: Dependent on LLM provider (OpenAI: 10K TPM, Claude: 40K TPM)
- Latency: 2-5s per query (network + LLM inference time)
- Compliance: AI-generated responses require human legal review
Ready to Automate Tax Compliance?
Join tax professionals using TAXIA for accurate, auditable, and automated Korean tax law analysis. Start in 5 minutes with our demo mode.