RAG-Enhanced Energy Advisor
The RAG-Enhanced Energy Advisor explores how retrieval-augmented generation (RAG) frameworks can improve decision-making and control strategies in energy management systems — while also examining their potential security vulnerabilities.
This project simulates a scenario where an attacker attempts to trick an LLM into generating inappropriate or unsafe outputs, such as fabricated or misleading control actions, or even false medical diagnostics within smart building health-energy systems.
The system aims to demonstrate defensive prompting, retrieval filtering, and trust calibration mechanisms to ensure that LLM-based advisory systems remain robust, interpretable, and safe.
Research Context
- Integrates RAG pipelines for real-time adaptive learning in multi-building environments.
- Examines prompt-injection attacks that can mislead models into unsafe or irrelevant outputs.
- Introduces trust-aware retrieval weighting to dynamically filter retrieved documents based on domain relevance and safety metrics.
Core Technologies
- LangChain for retrieval orchestration
- FAISS / ChromaDB for vector-based semantic search
- OpenAI GPT / Llama 3 as the base reasoning model
- CityLearn environment for multi-building simulation and energy optimization
Key Insights
The project highlights the dual nature of RAG systems — powerful for enhancing reasoning and grounding, yet susceptible to data poisoning and adversarial instructions.
By incorporating safety filters and reinforcement-based trust weighting, this framework helps move toward secure, reliable LLM-driven control in energy and cyber-physical systems.
