Learning

How to produce consistent generative AI answers?

Jay Chiarella 2 min read
#learning#generative AI#consistent-answers

Intended audience: Folks who are curious how WiselyDesk produces consistent answers while leveraging “generative AI”.

Before we dive in, let’s cover some jargon.

Jargon explanation:

Semantically Similar: When two sentences/phrases/word have similar meaning. They don’t need to be identical in words, but the meaning is essentially the same.

Categorizing pairs of sentences as semantically similar or not.
Semantic Similarity Explanation

Question

If I ask the bot a quesiton 100 times, I’d like generative AI to produce 100 identical answers. Can you guarantee this?

Answer

You can’t guarantee all 100 answers will be identical. However, by using sources and tuning the bot instructions to your business, the 100 answers should be semantically similar.

WiselyDesk’s secret sauce?

We run a script that asks the chatbot the same questions a large number of times. We then check if the answers are semantically similar or not.

In fact, before we launch our chatbot infront of our customer’s users, we make sure to test the top questions at least 50 to 100 (sometimes 1000) times to ensure each of the answers are semantically similar.

If the answers are all semantically similar, then we test the other top questions. If the answers are not semantically similar, then we invesigate by looking at the sources and the bot instructions. More on this in later articles.

Outro

Ensuring that WiselyDesk does not produce hallucinated or false answers is our top priority. It should be the top priority of any user of a generative AI, and one that we take very seriously. As always, any questions or comments - contact us!

← Back to Blog