Today, we support prompt templates and schemas for system, user, and assistant. You can mix structured and non-structured inputs by completing them with raw_text.

However, there's a case for using multiple prompt templates and schemas for a given role, e.g. a structured multi-turn or multi-block conversation. We might want multiple user messages with different templates, or even a message with multiple content blocks with different templates.

In the config, today you can specify {system|user|assistant}_{schema|template}.

We can generalize this further with "named input templates". You'd specify:

[functions.my_function]
# ...
input_schemas.abc = "path/to/schema.json"
input_schemas.def = "path/to/schema.json"
output_schema = "path/to/schema.json"  # no change
# Deprecate: user_schema
# Deprecate: assistant_schema
# ...

[functions.my_function.variants.my_function]
# ...
input_templates.abc = "path/to/template.minijinja"
input_templates.def = "path/to/template.minijinja"
# Deprecate: user_template
# Deprecate: assistant_template
# ...
input_wrappers.system = "path/to/template.minijinja"
input_wrappers.assistant = "path/to/template.minijinja"
input_wrappers.user = "path/to/template.minijinja"
# ...

The values input_schemas."tensorzero::*", input_templates."tensorzero::*", and input_wrappers."tensorzero::*" are reserved.

If you define a schema, you must define the corresponding input template, and vice-versa.

At inference time, you'd send messages like:

{
  "role": "user",
  "content": [
    {
      "type": "template",
      "name": "abc",
      "arguments": {
        "x": 1
      }
    }
  ]
}

We'll gradually deprecate arguments in text content blocks.

For system, if you provide an object, it will try to use the input_templates.system template and error if not present.

input_wrappers templates support a single variable, text, which is the text content block. (Any other values currently supported will be gradually deprecated.)