dspy
dspy copied to clipboard
stanfordnlp
add system prompt design to GPT model
It seems missing a key feature: "system prompt" "system prompt" impacts gpt-model significantly at least let as can put in config or somewhere else?
Good point. We need to think harder about chat model support. We may put the instructions in the system prompt, or something else.
Out of curiosity, what would you like to put in the system prompt?
Out of curiosity, what would you like to put in the system prompt?
Want to use the same pipeline across projects, with varying background context. The pipelines capture the cognitive processes well, those don't vary across projects. But the background context changes from project to project. (Maybe there is a diff way to do this other than system prompt? I'm new to DSPy.)
Good point. We need to think harder about chat model support. We may put the instructions in the system prompt, or something else.
Out of curiosity, what would you like to put in the system prompt?
My best scenarios are almost:
Packing what I want GPT to do into a Prompt Class structure, similar to object-oriented programming, but with a very programmatic way of writing prompts.
And putting the data and context in the User Prompt.
system prompt = method, user prompt = user input data
GPT API will be significantly work that way.It is much stronger than putting everything in the user prompt.
It may not just be the specific LM, but also the native language. I work at S&N (Germany)
+1. Have experimented with guiding behavior in system prompt as well as in user prompt. For certain (admittedly raunchy) behaviors, putting them in the system prompt works, whereas shoving it in the user prompt results in "I'm sorry, I cannot assist.." etc.
DSPy is unusable for building complex gpt-based chatting products at the moment, imo. Will explore other models or other tools in the meantime.
The ideal behavior for me would be that for models which have a chat completion API, DSPy at the very least builds a chat history of messages with roles, as the api expects, instead of putting everything in a single user message. Perhaps having a special "chat history" input parameter, along with either letting us specify the system prompts, or intelligently building the system prompts.
