I’m trying to teach the assistant to use the retrieval knowledge primarily, but it completely ignores the instructions (randomly, but in 80%). It uses its own knowledge and hallucinates much more often then using the uploaded files. Tried to switch the models, and, surprisingly, 3.5 turbo is more reliable in using retrieval. But still not reliable.
How to force it to use the retrieval knowledge?
Try adding in Instruction like “Do Not cite any information that is not present in the file” or something similar. Assistant V2 are giving me better results than V1.
Are you uploading files manually or via code REST API Hit?
No, prompting doesn’t help. It seems that’s about RAG itself.
So, the solution was pretty simple and elegant: embeddings with pinecone. Now even 3.5t provides pretty accurate results for 0.5$ per 1m tokens 
Thanks for your feedback @bersus! Would you be open to sharing some of your prompts and examples where things didn’t meet your expectations? We’d love to improve things in the next version of the tool. Feel free to drop me a quick email at nikunj@openai.com and we can set something up.
