A Retrieval-Augmented Generation (RAG) system combining a vector-based document retriever with a generative language model (Microsoft Phi). Optimized for knowledge-intensive tasks such as question answering, summarization, and conversational agents.
- Although Microsoft Phi is highly optimized, like most generative LLMs, it is prone to hallucinations due to its training on synthetic or filtered datasets.
- The original dataset used for RAG retrieval has removed labels such as "fake news" or "misinformation," meaning the model may retrieve and generate content based on unreliable sources.
- Detecting hallucinations is particularly challenging in this setup, as the RAG system generates responses conditioned on retrieved documents that may themselves contain errors or omissions.
- Users should exercise caution when using the system for critical decision-making, and consider additional validation steps on the retrieved content.