Optimizing AI Models: Key Strategies for Better Outcomes

Deepanshu That really is the big question. You do need to be using a state of the art model GPT-4o, Claude Opus or Gemini otherwise you are starting at a disadvantage. You need to clearly add in instructions to not respond if the context does contain the answer and you need to iterate on those instructions as you find problems. You will find that you need it to be stricter sometimes and others a lot less (all go in instructions). You find those problems using RAG Q&A Eavls and Hallucination Evals, with Explanations. The Explanations on Evals are the KEY guides to improve your templates. Other areas we've seen teams iterate, is improving the quality of search retrieval so you can use less text for generation. Sometimes using more material than needed makes it a harder problem.