If you've ever asked ChatGPT a question and gotten a confident, plausible, completely wrong answer — you've experienced what AI researchers call a hallucination. Now imagine that same problem in an HR chatbot, where an employee asks about parental leave and gets a fabricated policy back. That's not just embarrassing. That's a legal risk.
The solution is a technique called Retrieval-Augmented Generation, or RAG. It's the architecture that makes Mosaic AI's HR Q&A assistant safe to put in front of employees.
Large language models like GPT-4 are trained on enormous amounts of public text. They know a lot about HR policies in general. They know nothing about your policies specifically. So if you ask one "How much PTO do new employees get?" it will happily invent a reasonable-sounding answer based on what's typical — and that answer may have nothing to do with what your handbook actually says.
For HR, that's a non-starter. Employees need answers grounded in your real policies, not industry averages.
RAG works in two stages: retrieval, then generation.
1. Retrieval. When an employee asks a question, Mosaic AI first searches your handbook for the most relevant passages. This uses the same semantic search technology that powers our resume matching — finding policy sections by meaning, not just keyword. Ask about "time off for a new baby" and it pulls the parental leave section, even though your handbook never uses the word "baby."
2. Generation. Those retrieved passages get fed to the LLM along with the original question and a strict instruction: answer only using the text provided. The model writes a natural, conversational response, but the facts come from your handbook. Not its training data. Not its imagination. Your handbook.
Three things become possible with a properly built RAG system:
Any HR chatbot that doesn't use retrieval grounding is fundamentally guessing. It might guess well most of the time, but "most of the time" is not a standard you want governing benefits, leave, or compliance questions.
Mosaic AI's HR Policy Q&A Assistant is built on RAG from the ground up, with vector search over your actual handbook content and explicit grounding instructions to the LLM. The result is a chatbot HR can trust to put in front of every employee. Book a demo to see it answer questions about a handbook you upload.