How do you use LLMs to verify databases with minimal hallucinations?
I have a medium sized database of around 150 entries each with 10-15 parameters.
It was put together by Claude. But the amount of hallucinated data is extraordinary!
Trying to fix it using another llm like chat gpt or gemini hasn’t worked since they balk at looking for data for >50 data points.
Gemini actually deleted 100 entries from the database while analysing it!
So the question is-
Is there a suitable way to analyze the database for inaccuracies/hallucinations and fix them, apart from manually going through each data point personally? First of all why give write permission to an LLM ? Especially in this case of analysis. Secondly, isn't it more simple and accurate to write a good old code, even with the help of AI, to just verify algorithmically the data points ? So you don't have to wrestle with hallucinations.