Curious to hear your experience. I built a simple RAG using GPT4-Turbo some weeks ago. Only used it for a few hours but was mostly satisfied. I did notice if I sent it too many documents, it would not find the (one) doc I was looking for.
GPT4 Turbo is top of the class it does RAG very well, it is important to provide good context with help of Vector DB but if you cannot provide relevant document it cannot do much.
All the opensource models are super bad at this and mostly i want to blame the fine tuning to get to the leaderboard is affecting the quality
Curious to hear your experience. I built a simple RAG using GPT4-Turbo some weeks ago. Only used it for a few hours but was mostly satisfied. I did notice if I sent it too many documents, it would not find the (one) doc I was looking for.
GPT4 Turbo is top of the class it does RAG very well, it is important to provide good context with help of Vector DB but if you cannot provide relevant document it cannot do much. All the opensource models are super bad at this and mostly i want to blame the fine tuning to get to the leaderboard is affecting the quality