LLM Calls Utility Actors

Invoking an LLM

The LLMInvoke actor is the primary method through which requests are sent to an LLM. It sends structured messages and receives responses.

Message Roles

When sending a prompt to an LLM, the messages are structured with distinct roles to indicate the purpose of each message:

Role	Purpose	Example
System	High-level instructions, tone, and behavioral guidelines	"You are an AI assistant for Elite Bank. Always answer in a friendly, professional tone."
User	The actual question or input submitted by the end user	"What is the interest rate on a 30-year fixed mortgage?"
Assistant	Previous model outputs for conversation continuity	[Previous response in multi-turn conversation]. The assistance

LLMInvoke Actor Parameters

Parameter	Type	Description
`interface`	string	Interface identifier (e.g., `llm://default`)
`messages`	object	Array of role-based messages
`stream`	boolean	Determines whether the response is streamed during generation

Output Parameters

Output	Description
`stream`	The response content
`total_tokens`	Total tokens used
`input_tokens`	Tokens in the prompt
`output_tokens`	Tokens in the response
`cache_read_tokens`	Tokens read from cache
`cache_write_tokens`	Tokens written to cache

Prompt Construction Actors

Fabric provides utility actors that simplify and streamline the process of building prompts:

LLMConst

This actor constructs LLM prompt messages for any role and is used for defining the static portions of a prompt.

The content can include parameter references in the form of \${value}.

If the value key is not found, the parameter reference will remain in the text for later phases.

LLMAppend

This actor merges multiple message arrays into a single array before passing them to the LLMInvoke actor.

Use Case: Combine system instructions, few-shot examples, and user query into one prompt.

Example Flow:

LLMConst (System) ────┐
                      │
LLMConst (Examples) ──┼──▶ LLMAppend ──▶ LLMInvoke
                      │
LLMConst (User) ──────┘