MiniNihonAI

Lightweight AI Infrastructure for the Japanese-Language Applications, Built for Extensible Multimodal Agents

---

Overview

MiniNihonAI is a modular, lightweight AI infrastructure framework tailored for Japanese-language applications. It is optimized for local deployment (e.g., RTX 4090) and cloud elasticity, and serves as a robust backend engine for building retrieval-augmented generation (RAG) systems, multimodal assistants, and future autonomous agent applications.

This repository focuses solely on the Infra layer—apps like tourism assistants or autonomous driving agents will be built on top of this infrastructure using its modular APIs.

---

Core Features

• ✅ Local-first deployment with optional cloud extensions
• ✅ Japanese LLM support (Mistral 7B, Phi-2, Elyza-Japanese)
• ✅ Embedding & RAG engine with FAISS/Qdrant
• ✅ Optional multimodal processing (CLIP, BLIP2)
• ✅ Modular FastAPI endpoints (e.g., /ask, /embed, /image_caption)
• ✅ PEFT-friendly architecture (LoRA/QLoRA compatible)
• ✅ Extensible for downstream Agents: tourism bots, autonomous vehicle assistants, etc.

---

Directory Structure (Full)

MiniNihonAI/
├── api/                        # FastAPI server and route logic
│   └── main.py                # Entry point for API
│   └── routers/               # Route modules
│       └── qa_router.py       # QA endpoint logic
│
├── configs/                   # Configurations
│   └── config.yaml            # All model and service configs
│
├── data/                      # Data storage and examples
│   ├── documents/             # Raw and cleaned knowledge docs
│   ├── images/                # Sample tourist images
│   └── index/                 # FAISS / Qdrant vector indexes
│
├── deploy/                    # Deployment scripts
│   ├── docker-compose.yml     # Docker infra
│   └── aws_launcher.sh        # AWS EC2 setup script
│
├── models/                    # Model loaders
│   ├── llm_loader.py          # Load and run LLMs
│   ├── embed_loader.py        # Load embedding models
│   ├── clip_loader.py         # For multimodal (image-text)
│   └── translator.py          # Optional multilingual tools
│
├── rag/                       # Retrieval-Augmented Generation
│   ├── retriever.py           # Vector search logic
│   └── rag_pipeline.py        # QA pipeline with generator
│
├── agent/                     # (Experimental) Agent logic
│   ├── intent_router.py       # Convert natural language into structured control commands
│   └── motion_planner.py      # Motion planning module for autonomous navigation
│
├── ui/                        # (Optional) Gradio/Streamlit UI
│   ├── gradio_ui.py
│   └── streamlit_ui.py
│
├── utils/                     # Utilities
│   ├── logger.py              # Logging utility
│   ├── splitter.py            # Text splitter for RAG
│   └── file_loader.py         # File and doc loader
│
├── vector_store/             # Vector DB interfaces
│   ├── faiss_engine.py        # FAISS wrapper
│   └── qdrant_engine.py       # Qdrant wrapper
│
├── logs/                     # Logging outputs
│
├── .env                      # Environment variables (API keys, etc)
├── requirements.txt          # Python dependencies
└── README.md

---

Use Cases Enabled by This Infra

• 🗺️ Tourism Agents – Japanese Q&A over local spots, routes, food
• 🖼️ Multimodal Search – Upload images and retrieve nearby location info
• 💬 RAG QA Bots – Retrieval-augmented answering over structured or unstructured Japanese knowledge
• 🈳 On-Device LLM Inference – Low-latency GPT-style response generation
• 🚘 Autonomous Driving Agents (Prototype) – Natural language intent → driving task interpretation pipeline
• 🛣️ Motion Planning for Autonomous Agents – Waypoint-based path planning and behavioral routing (WIP)

---

Agent-Oriented Extensions

This infra also supports building autonomous or semi-autonomous agents, such as:

🧳 "Japanese Tourism Guide Agent"

A cultural and location-aware assistant that handles tourist queries like “推荐京都春天拍照的景点有哪些？” and:

• Answers using location-tagged documents, embedding search, and optional image input
• Recommends spots, routes, and real-time considerations like weather or crowding
• Can power mobile apps, web guides, or LINE bots

🧭 "Japanese Autonomous Driving Agent"

A prototype that interprets voice or text commands like “代々木公園まで連れてって” and:

• Translates → extracts intent → queries embedded location docs
• Returns planned route or task
• Optionally connects to real/simulated vehicle SDKs

🤖 "Japanese-English Translation Agent"

A service agent capable of performing lightweight real-time or batch translation using a combination of LLMs, NLLB models, and a retrieval base.

• Translates and rewrites based on user intent and tone
• Can integrate with browser plugins or chat interfaces

🍱 "Hotel Concierge Robot Agent"

A multimodal assistant that receives text or image inputs to:

• Recommend nearby restaurants or services
• Answer common questions about bookings or hotel rules
• Generate suggested replies to Japanese-language guest queries

Modules used:

• llm_loader.py, translator.py, vector_store/
• agent/intent_router.py, agent/motion_planner.py
• Future extensions: speech_interface.py, vehicle_api.py, task_manager.py, tts_engine.pyFuture extensions: speech_interface.py, vehicle_api.py

---

---

Deployment Targets

• ✅ Local machine (Linux with CUDA, Windows WSL2 supported)
• ✅ Docker Compose setup
• 🔜 AWS EC2 GPU spot instance launcher

---

License

MIT License (c) 2025 Miao Jiang.