Module messages 

Unified message architecture for LLM interactions.

This module provides the core UnifiedMessage type that works across all LLM providers. It’s the primary abstraction that makes multi-llm provider-agnostic.

Overview

The unified message system provides:

• Provider-agnostic messages: Same format works with OpenAI, Anthropic, Ollama, and LM Studio
• Caching hints: Native support for Anthropic prompt caching via MessageAttributes
• Priority ordering: Control message ordering with priority-based sorting
• Rich content types: Text, JSON, tool calls, and tool results via MessageContent

Quick Start

use multi_llm::{UnifiedMessage, MessageRole};

// Simple messages using convenience constructors
let user_msg = UnifiedMessage::user("Hello, how are you?");
let system_msg = UnifiedMessage::system("You are a helpful assistant.");
let assistant_msg = UnifiedMessage::assistant("I'm doing well, thank you!");

// Build a conversation
let messages = vec![system_msg, user_msg, assistant_msg];

Caching Support

For Anthropic’s prompt caching (90% cost savings on cache reads):

use multi_llm::UnifiedMessage;

// Mark a system prompt for caching (5-minute TTL)
let cached_system = UnifiedMessage::system("You are a helpful assistant.")
    .with_ephemeral_cache();

// For longer sessions, use extended caching (1-hour TTL)
let long_context = UnifiedMessage::system("Large context here...")
    .with_extended_cache();

Message Categories

Use semantic constructors to get appropriate caching and priority defaults:

use multi_llm::UnifiedMessage;

// System instructions (cacheable, highest priority)
let system = UnifiedMessage::system_instruction(
    "You are a helpful assistant.".to_string(),
    Some("system-v1".to_string())
);

// Context information (cacheable, medium priority)
let context = UnifiedMessage::context(
    "User preferences: dark mode, verbose output".to_string(),
    None
);

// Current user input (not cached, lowest priority)
let current = UnifiedMessage::current_user("What's the weather?".to_string());

Structs

MessageAttributes

Attributes that guide how providers handle a message.

UnifiedLLMRequest

A complete request to an LLM provider.

UnifiedMessage

A provider-agnostic message for LLM interactions.

Enums

CacheType

Cache type for prompt caching (Anthropic-specific feature)

MessageCategory

Semantic category of a message for provider-specific handling.

MessageContent

Content of a message, supporting text, JSON, and tool interactions.

MessageRole

Role of a message in an LLM conversation.