Prompt templates

Placeholders

prompt: "Translate to English: {sentence}"

Each {name} is replaced with that row's value from an input or context column. Column names can contain spaces: {Source sentence} works. An empty cell becomes an empty string.

The template must use at least one placeholder, or every row would get the same prompt. A placeholder that is not an input or context column is an error, reported before any request is sent. The expected column is never available, so the answer cannot leak into the prompt.

Literal braces

Write {{ and }} for literal braces, for example when asking for JSON:

prompt: |-
  Extract the entities from the text as JSON, like {{"person": [...], "place": [...]}}.
  Text: {text}

System prompts

Give the prompt as an object to add a system message. Placeholders work in both parts.

prompt:
  system: You are a careful annotator for sentiment analysis research.
  user: |-
    Classify the sentiment of this {language} text as "positive" or "negative".
    Answer with one word.

    Text: {text}

Multi-line prompts in YAML

Use |- for a block that keeps line breaks and drops the final newline. Indent every line of the block by the same amount.

Tips for scoreable answers

• Ask for the answer format you will score, such as "Answer with one word" for classification or "Answer with the number only" for arithmetic.
• For classification, give the accuracy or f1 metric a labels list, so a chatty answer like "The sentiment is positive." still counts as positive.
• Check a few outputs in results.xlsx before trusting a score. Formatting problems show up in the Outputs sheet straight away.

From Python

from paperlab_space import PromptTemplate

template = PromptTemplate("Translate to English: {sentence}", system="Be literal.")
print(template.variables)  # ['sentence']
print(template.render({"sentence": "สวัสดี"}))  # chat messages

A plain string works wherever a template is expected, such as RunConfig(prompt="...").