In test.json, the prompt can be specified in three ways, by providing either the prompt, prompt_file or multi_run_prompt.
However, only one of these may be specified at the same time.
If prompt is specified, it's value should be a list of JSON objects, so called "messages".
Each message must have at least the field content.
Optionally, each message may also have the role field with either the values user, system or assistant. If no role is specified, it defaults to user.
This mirrors the Chat ML format of the OpenAI Chat API.
We permit further fields in the context of Multi-Variable Prompts.
After running the LLM, the assistants response will be appended as a new message with the assistant role.
If prompt_file is specified, it is a path (absolute or relative, though we recommend relative) to a .prompt file containing the prompt. Below we how this file can be formatted.
Lastly, we discuss how multi_run_prompt can be specified in Multi-Run Prompts.
Prompt File Format
A prompt file can contain either plain text, a single message as a single JSON object or a sequence of messages with one JSON object per line.
If just plain text is provided, it read as a single message (with the user role).
Each JSON object must have the field content and optionally role as discussed above.
Multi-Run Prompts
By default, for each instance the LLM is only invoked one.
However, via the field multi_run_prompt it is possible to specify multiple prompts that are all run and the results are jointly fed to the checker as one instance.
The field multi_run_prompt can be a list of JSON objects, which each have the fields name (optional, currently ignored), repetitions (optional, defaulting to 1) and either prompt or prompt_file as outlined above. Each prompt specified this way is executed repetitions many times.
For an example see the corresponding dummy LVE.
Multi-Variable Prompts
By default, the LLM is just invoked to complete the Chat prompt with a single trailing assistant message. However, it is also possible to specify a longer back-and-forth.
To do this, messages with the role assistant, but content set to null can be specified. Whenever such a message is encountered, all the messages prior to it are send to the LLM as a prompt, the completion is saved as the content of that message and the remainder of the prompt is executed in the same fashion.
If such messages are present, they need to also have the field variable set. In the checker the results of the individual LLM completions are available through the respective variable names.
Unless the last message in the prompt is an assistant message with null content one will be appended.
assistant messages with non-null content are still handled in the default way.
