PromptBuilder

[ZanSara]

See the LLM proposal: #5540

---

The goal of PromptBuilder is to transform prompt templates, which are strings with variables, specifically Jinja templates, and fill up those variables with values that come from other components in the pipeline or from the pipeline inputs. The output of this component is one (or more) prompts, where prompts means strings that the LLM can directly use.

PromptBuilder is not tokenizer aware: the prompt will be checked for length by the LLM component before inference. If the need arise, we may later extend the component.

Draft I/O for PromptBuilder:

@component
class PromptBuilder:

    def __init__(self, template: Union[str, Path]):
        self.template = # Download the template
        template_variables = # extracts the variables from the template text
		component.set_input_parameters(**{var: Any for var in template_variables})

  	@component.output_types(prompts=List[str])
    def run(self, **kwargs):
        # Render the template using the variables
        return {"prompts": prompts}

Due to the dynamic nature of prompt templates, the PromptBuilder.run() method takes **kwargs, which contains all the variables that will be filled in the template. However, for this component to work with Canals, we need to know in advance which values this dict will contain: therefore, we need the users to specify in the __init__ of the component the template to use.

Such template names cannot be changed at runtime.

[silvanocerza]

Closed by #5713