How we manage hallucinations for Alfa

Hallucinations are factual inaccuracies or fabricated information delivered by an LLM. At Boosted.ai, we strive to mitigate Alfa’s hallucinations as much as possible.

Hallucination causes: 

  • Pattern-based generation: Rather than knowing things, LLMs predict the sequence of words based on data they are trained on
  • Lack of real-world understanding: Models don’t understand information in the same way as humans, and they do not have access to real-time knowledge which can lead them to produce inaccurate or invented information 
  • Training data limitations: The quality and scope of info used to train the model can impact its accuracy.

How we mitigate them:

  • Using RAG limits LLM hallucinations
  • We verify and validate information retrieved from RAG, reducing the hallucination rate from ~5% to 1 in 10000 (0.01%)
  • We’ve created a framework where users can be very explicit with their asks - the less the LLM has to interpret the less likely it is to make mistakes