I like Edit 2. I often load the context and reference material as the Agent. i often get weird results if a lot of context is loaded.
Does anyone know the limitations of the System prompt or if it is good practice to have multiple system prompts?
I agree this has been a good philosophical discussion.
However, I think there were some very good points with agreement. One being the System Prompt. Rather that saying “You are a happy assistant that give concise helpful answers.”, say “You are a happy assistant that give concise helpful answers that works in an auto shop in Dallas, TX that performs mechanical services but no collision services …”. Maybe we should send the whole discourse to GPT and have it summarize the findings.
Actually, it might b an idea for OpenAI. Once a thread goes beyond 10 replies, provide a short synopsis at the top.
I’ve learned a bit, since my first replies, and this would be my guess at how RAG tools solutions (like in LangChain) would format a question from a user:
Answer the following question using the context provided:
<question>
What is the capitol of Texas?
</question>
<context>
....all text found from RAG matching (or top-N chunks, etc)
</conext>
I think XML is best, for delimiting subparts in a query, and probably the System Prompt should be more general and not specific to any particular question in the conversation. So the above would go in a User Message, not a System Message. Hope this helps.