Hello, I recently started developing such systems but I wonder what they should be like. What principles should I follow to increase efficiency? Should I prefer long answers or short and concise ones? Also, how can I verify the given responses? In your opinion, which approaches work?
How are ChatGPT-like tools used?
👁️ 4 views💬 2 replies❤️ 0 likes
2 Replies
When I first started using Buck with similar applications, I faced the same challenges. For responses, I recommend starting with practical application in "prompt engineering." For example, in tasks like prompt engineering, try breaking the task into smaller parts and asking the model to answer step by step. For phrasing, I initially used unstructured long paragraphs, but after some experiments, I found that a structured format (like lists or short paragraphs) improves clarity and avoids over-explanation.
To validate answers, I prefer cross-referencing multiple sources. For example, if I request documentation for a database, I look for the original source outside the model. I also ask the model to explain its reasoning (e.g., "How did you arrive at this conclusion?") and verify its logic. As for length, remember that most users need a direct response that gets to the point quickly, unless the context requires more detail.
This looks like it could be helpful, thanks. In terms of experience, what kind of systems have you worked with—do you develop them yourself or work with existing ones?