Adaptive-AtomRAG ⚡️

Sub-millisecond query routing & cost-aware execution for GraphRAG architectures.

Adaptive-AtomRAG is a highly optimized conditional routing gateway designed to resolve the latency and cost bottlenecks of heavy Atom-Entity Graph (AEG) RAG systems. By integrating a lightweight TF-IDF + SVM classifier, the system dynamically routes queries, reducing average latency by 47% and token costs by 48% on simple tasks.

🚀 Key Features

• Sub-millisecond Routing: Classification takes < 1 ms, preventing unnecessary LLM planning overhead.
• Cost-Aware Execution (CA-RAG): Maximizes utility using a customizable scalar function balancing quality, latency, and token cost.
• Anti-Noise Context: Prevents context pollution on simple factual queries by bypassing deep graph traversal.
• Production Ready: Built on FastAPI with asynchronous endpoints, ready for Prometheus metrics and Kubernetes deployment.

📊 Benchmarks & System Impact

Metric	Naive RAG / Direct	AtomRAG (Unconditional)	Adaptive-AtomRAG (Routed)	Impact
P95 Latency	~4.0s	~22.3s	Dynamic (52.9% of base)	-47.1%
Token Cost	Fixed	Maximum	Optimized	-48.5%
Routing F1 (Macro)	N/A	N/A	0.928	TF-IDF + SVM

🧠 Mathematical Foundation

The router maximizes the query utility $U_b$ for a given bundle $b$:

$$U_b = w_Q \hat{Q}_b(q) - w_L \hat{L}_b^{\text{norm}} - w_C \hat{C}_b^{\text{norm}}$$

To prevent router collapse during token-level routing, sparse regularization penalties $\lambda$ are applied to the loss gradient:

$$\frac{\partial d_t}{\partial L} = \left(\frac{\partial o_t}{\partial L_{\text{NTP}}}\right)^T a_t + 2\lambda d_t$$

🛠 Quick Start

# Clone the repository
git clone [https://github.com/YourUser/Adaptive-AtomRAG.git](https://github.com/YourUser/Adaptive-AtomRAG.git)
cd Adaptive-AtomRAG

# Spin up the router
docker-compose up -d --build


Adaptive-AtomRAG/
├── README.md                  # Витрина проекта
├── requirements.txt           # Зависимости
├── docker-compose.yml         # Быстрый запуск
├── docs/
│   └── whitepaper.md          # Твоя статья с формулами и бенчмарками
└── router/
    ├── __init__.py
    ├── classifier.py          # Логика TF-IDF + SVM маршрутизатора
    └── main.py                # FastAPI шлюз