I’ve been testing my app for quite some time and in some cases I see that gemini happens to create things that don’t actually exist, for example it said that abbot of canterbury (which is an old poem and quite classical) was actually a novel by some random guy (name not listed here) . Is there some way to fix this? I’ve lowered the temperature but that doesn’t have much of an effect.
1 Like
Great question!
The first and most important thing to remember is that Gemini, and any LLM, is not a source of truth. They are language models and pattern systems. No matter how much is done to eliminate hallucinations, there is always a chance the patterns will set the LLM up to generate something that is false.
There is a lot of work to “ground” responses from an LLM in a set of information. Retrieval Augmented Generation is usually the one most frequently talked about. Under these schemes:
- You will use an LLM like Gemini or an embedding model to “understand” what the user is looking for.
- Using more traditional methods (for example, a vector search or semantic search) you’ll come up with source documents that appear to address what the user is talking about.
- Then, you’ll give these documents, along with the original question, to Gemini with instructions to use the documents to answer or summarize an answer to the question.
You can find quite a few articles, tools, and write-ups discussing various approaches to this. Here’s a presentation I did on the concept: https://youtu.be/E8tL_WL8NTo?si=ZB9ZzHFC6qPaF2Iw
2 Likes
So we have to extend the AI’s capability?