How to Optimize Existing Prompt

[NickL77]

Hi, I'm looking to optimize an existing system prompt we already have. How can I initialize the program with my existing prompt and have GEPA start from there? I tried something like:

state = program.signature.dump_state()
state["instructions"] = STARTING_INSTRUCTIONS
program.signature.load_state(state)

but GEPA seemed to just reinitialize the instructions.

A couple other features I'm interested in (and haven't found in tutorials or documentation):

1. Freezing certain parts of the prompt
2. Certain frozen parts can be re-ordered (ie I have a set of in-context learning examples that I don't want to be modified but GEPA can change the order of them

[sc23333]

I also have the same problem..

[LakshyAAAgrawal]

Context: https://discord.com/channels/1161519468141355160/1415751892906152127/1415751892906152127
For (2) you can use the .demos field of DSPy module predictors, and set it to the few-shot examples you have. GEPA will not touch/modify it during evaluation.
For starting from a custom instruction, simply doing: pred.signature = pred.signature.with_instructions("") where pred is a specific predictor in your DSPy module should work
For (1), can you give an example? We have recently integrated a pipeline to enable users to provide a custom instruction proposer.

The feature is fully integrated and ready to use, but the documentation is WIP.

You can have a look at https://github.com/stanfordnlp/dspy/pull/8775/files to find how to use the custom instruction proposer, which should enable you to control exactly how GEPA will update your prompts.

You will need to install DSPy from main: pip install git+https://github.com/stanfordnlp/dspy.git
Would be glad to receive any feedback on this!
Happy to help answer any questions regarding any of these!

[LakshyAAAgrawal]

Hi @NickL77 and @sc23333 , can you let us know if it ended up working for you? I am closing the issue for now, but please feel free to reopen it in case your query wasn't addressed, or you faced any other issues!