Is it possible, for any current LLM, to solve this simple task (a 2K prompt)? Any LLM, open-source or proprietary? Intensive testing revealed, that there is an extreme limited amount of LLMs that can solve my simple task, even with a text redundancy of 2 (!) in the prompt! No reasoning involved, just to find info in a text prompt w/ 2k token length. Happy RAG-time!
My test prompt is 1105 words long, contains 8293 characters and (w/ my tokenizer) close to 2k tokens long. I tested no LLM with a context length below 2k.
The prompt is too long to upload to YouTube.
DIY: Generate any topic specific text and simply insert a paragraph of text (maybe with a task description) twice somewhere in the "lost" middle. Smile.
I tested it for astrophysics, aerodynamic, geography and cooking.
Now you know why professional RAG systems have a text redundancy of 4 to 6 for every alpha-reranked information .... did I mention, that I have a dedicated video on Self-RAG?
https://www.youtube.com/watch?v=i4V9iJcxzZ4
So ... is RAG the solution? Not quite; it primarily addresses the existing shortcomings in our LLMs, sometimes.
#aieducation
#performancetesting
#airesearch
23 Comments