The author developed an evaluation harness for their personal AI digital twin, which answers questions based on their profile. The harness revealed that the AI hallucinated in 25% of test cases, despite a prompt designed to prevent such behavior. The system uses a retrieval-augmented generation (RAG) approach with a simple JSON file as a vector store, employing Amazon Bedrock's Titan v2 model for embeddings and cosine similarity for retrieval. The evaluation process separately assesses retrieval accuracy using metrics like recall@k and nDCG@k, and answer quality via an LLM judge. AI
IMPACT Highlights the challenges in ensuring factual accuracy and preventing hallucinations in RAG systems, even with specific guardrails.
RANK_REASON The item describes the creation and use of a custom evaluation tool for a personal AI application.
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
