Templates

template_format on the factory selects how the messages are rendered; its default, langfuse, matches the playground. A prompt may carry its own format in the config key template_format, which applies when allowed_config_keys names it or is "any" (see Models and config keys). The key never reaches the model.

{"provider": "openai", "model": "gpt-5.5", "template_format": "jinja2"}

Format	Syntax	Rendered by
langfuse (default)	{{ name }}	The TemplateParser of the Langfuse SDK
f-string	{name}	str.format, as in LangChain prompt templates
mustache	{{ name }}, {{{ name }}}, {{#flag}}...{{/flag}}	The LangChain mustache renderer
jinja2	{{ name \| upper }}, {% if %}	Jinja2 in a sandbox, as in LangChain; needs pip install "langfuse-hotswap[jinja2]"

The format applies to every message of the prompt. In every format the variables are required: the input is validated by the LangChain prompt template, so a missing variable raises a KeyError that names it.

The langfuse format

Messages are rendered by the TemplateParser of the Langfuse SDK, so the result matches the playground. A contract test pins its rules:

• A variable is whatever stands between {{ and the next }}, stripped. Values are inserted with str() without escaping and None becomes an empty string.
• Single braces and HTML are left as they are. There are no sections: {{#flag}} is a variable named #flag, and {{a.b}} is a variable named a.b.
• Unlike prompt.compile(), a missing variable is an error, not left as is.

The LangChain formats

f-string, mustache and jinja2 are rendered by the functions behind ChatPromptTemplate with the same template_format, with their rules: f-string treats single braces as fields, mustache escapes HTML in {{ }}, renders sections and treats a.b as a nested lookup, jinja2 evaluates expressions and filters. Use them for prompts that need conditions or loops, and keep in mind that the Langfuse playground renders {{ name }} only.

Messages

• A text prompt becomes one human message. Roles LangChain does not know, such as developer, are kept as ChatMessage with that role.
• Placeholders are filled from the input variable of the same name, which must be a list of messages. Variables inside those messages are not rendered.
• Variables named stop and self are rejected by LangChain.