Is this an "Example Retriever" optimisation or it can also produce prompt from scratch like "LLM as Optimizers" ?

[gabgen]

It is not clear if the teleprompter is just a way to retrieve a good example from your training dataset or it can also write a prompt from scratch , and in this case how .

[okhat]

It's an example generation optimization, for a complete pipeline of steps. This is a lot richer than "LLM as Optimizers", which only optimizes a few words at the start of the prompt.

[okhat]

Learn more in the paper linked from the README (at the top).

[gabgen]

It's an example generation optimization, for a complete pipeline of steps. This is a lot richer than "LLM as Optimizers", which only optimizes a few words at the start of the prompt.

@okhat thanks for the feedback. "LLM as Optimizers" can actually write a prompt from scratch ( the number of words is tunable). This is particularly useful when you have very few examples to provide, so the problem is not to retrieve a good example, but more to provide a good prompt. In this case, would DSPy provide a solution to generate a good prompt, without relying on examples to put in the prompt? I saw you have a "finetune" teleprompter , is this something similar to what I am looking for ?