For 2 months after it's release, I struggled to figure out how to get the Gemini Pro API to read context documents sent in a RAG prompt. Finally got it to work a week ago. Then, just a few days ago, it stopped working again. My format is basically this:
"text": "<documents>context document list in xml format</documents>
System Message: blah blah
User Message: blah blah
"
And Gemini follows the system and user messages, but simply refuses to read or comprehend or remember what is in the document list.
This is true in the API as well as the playground:
https://makersuite.google.com/app/prompts?state=%7B%22ids%22%3A%5B%221CAkGmB7lvuRYqjFGr5tkVmFh0fUQXZ...
I have posted this on the Discord channel, but myself and one other person appear to be the only people using Gemini Pro for this sort of text intense application. No one ever has an answer, or even a suggestion, if we even get a response.
When pro was working, it was far superior to gpt-3.5-turbo-16k. Now it's useless. There has been a clear change somewhere. What's going on?
LinkedIn
Twitter
