Embeddings Aren’t Magic: The Predictable Failure Modes of RAG Retrieval

# Vector Search Has Predictable Blind Spots That Could Break Your AI System The AI retrieval systems many companies are building to search through documents have a major weakness: they're surprisingly bad at handling negations (like "do NOT approve"), exact codes or IDs, and company-specific jargon—even though they're great at understanding similar words and paraphrasing. If your business relies on these systems to pull accurate information from documents, you need to know when they'll silently give you wrong answers and what backup methods to use instead. The good news is these failures are predictable enough to plan around.