Reducing Hallucinations

We continue to improve the fundamental performance of AI agents in Pleep. This month we introduced architectural changes that significantly increased the accuracy and safety of conversations.

First, let's clarify what a hallucination is. Sometimes AI can generate plausible but false responses — for example, providing incorrect product names or prices. This is a fundamental limitation of LLM models. AI doesn't "know" facts the way a human does. It is trained to predict the most plausible continuation of text, not necessarily the most truthful one. That is why AI always tries to give some answer, even when the true answer is unknown to it.

We implemented a sophisticated anti-hallucination system to virtually eliminate them, and have already raised response accuracy to 99.71% (RAGBench Evals, 1000 shots).

This makes Pleep the most reliable AI sales agent on the market.

In just the last 7 days during testing, the new system helped prevent 17,589 potentially bad responses that could have contained fabricated information or simply mixed up prices for different products.

If the AI cannot verify information (for example, a customer asks an unusual question), it does not make things up — it transfers the conversation to a human operator.