Hi everyone!
What are strategies used to reduce Hallucination rate. According to sources ChatGPT itself has a hight rate around 25% .
Is anyone developing CHATgpt on marketplace using tools like RAGFix, or end points?
Benefits:
Ensuring all responses are grounded in real-time verified data instead of relying on pre-trained model alone.
Pulling directly from research databases, PDFs, or structured factual repositories before generating responses.
Providing fully formatted citations in real-time to enhance source credibility.
Thank you!!
Hi @cimoolivia and welcome to the community!
I haven’t used RAGFix, but I do have a problem with advertising that states “100% accuracy”.
Regarding significant hallucination reduction - my take is that you can never eliminate it with absolute certainty, but the best strategy to minimizing it is by optimizing your retrieval component. If the answer is in the retrieved chunks with high signal-to-noise ratio, and especially if the retrieved chunks re-inforce each other, this will give you great confidence in reducing hallucinations. And retrieval is really about two parts - ingestion/indexing, and the actual retrieval. So my advice is to focus on getting those to the best quality possible.