LLM Map

class sycamore.transforms.base_llm.LLMMap(child: ~sycamore.plan_nodes.Node | None, prompt: ~sycamore.llms.prompts.prompts.SycamorePrompt, output_field: str, llm: ~sycamore.llms.llms.LLM, llm_mode: ~sycamore.llms.llms.LLMMode | None = None, iteration_var: str | None = None, validate: ~typing.Callable[[~sycamore.data.document.Document], bool] = <function LLMMap.<lambda>>, max_tries: int = 5, filter: ~typing.Callable[[~sycamore.data.document.Document], bool] = <function LLMMap.<lambda>>, **kwargs)[source]

Bases: MapBatch

The LLMMap transform renders each Document in a docset into a prompt for an LLM, calls the LLM, and attaches the output to the document.

Parameters:

  • child -- Child node in the sycamore execution graph

  • prompt -- The SycamorePrompt to use to render each document. Must implement the render_document method.

  • output_field -- The name of the field in doc.properties in which to store the llm output

  • llm -- The llm to use for inference.

  • llm_mode -- How to call the llm - sync/async/batch. All LLMs do not necessarily implement all options.

  • iteration_var -- Name of the document property to increment with every invalid response. Default is None, which means no re-try.

  • validate -- Function to determine whether an LLM response is valid. Default is 'everything is valid'

  • max_tries -- Hard limit on the number of LLM calls per document. Default is 5

Example

prompt = EntityExtractorZeroShotGuidancePrompt.set(entity="title")

docset.llm_map(
    prompt=prompt,
    output_field="title",
    llm=OpenAI(OpenAIModels.GPT_4O_MINI)
)
class sycamore.transforms.base_llm.LLMMapElements(child: ~sycamore.plan_nodes.Node | None, prompt: ~sycamore.llms.prompts.prompts.SycamorePrompt, output_field: str, llm: ~sycamore.llms.llms.LLM, llm_mode: ~sycamore.llms.llms.LLMMode | None = None, iteration_var: str | None = None, validate: ~typing.Callable[[~sycamore.data.element.Element], bool] = <function LLMMapElements.<lambda>>, max_tries: int = 5, filter: ~typing.Callable[[~sycamore.data.element.Element], bool] = <function LLMMapElements.<lambda>>, **kwargs)[source]

Bases: MapBatch

The LLMMapElements transform renders each Element for each Document in a docset into a prompt for an LLM, calls the LLM, and attaches the output to the element.

Parameters:

  • child -- Child node in the sycamore execution graph

  • prompt -- The SycamorePrompt to use to render each element. Must implement the render_element method.

  • output_field -- The name of the field in elt.properties in which to store the llm output.

  • llm -- The llm to use for inference.

  • llm_mode -- How to call the llm - sync/async/batch. All LLMs do not necessarily implement all options.

  • iteration_var -- Name of the element property to increment with every invalid response. Default is None, which means no re-try.

  • validate -- Function to determine whether an LLM response is valid. Default is 'everything is valid'

  • max_tries -- Hard limit on the number of LLM calls per element. Default is 5

Example

prompt = TextSummarizerGuidancePrompt

docset.llm_map_elements(
    prompt = prompt,
    output_field = "summary",
    llm = OpenAI(OpenAIModels.GPT_4O)