Crate rune_chain_core 

Core traits and types for the rune-chain LLM orchestration framework.

rune-chain-core is the foundation layer of the rune-chain-* crate family — a modern, async-first Rust port of the LangChain concept. It defines the shared vocabulary (types, traits, error variants) that every other rune-chain-* crate depends on, without pulling in any LLM provider SDK itself.

Features

• Chain — the central trait; implement it to create any unit of LLM work
• Llm — provider-agnostic interface supporting text generation, streaming, and native function/tool calling
• Tool — async, fallible tool implementations for agent function calling
• Memory — pluggable conversation history (in-process, DB, vector store, …)
• Message / Role — typed conversation turns (System, Human, Ai, Tool)
• MessageContent / ContentPart — multi-modal content (text + images)
• GenerateResult / TokenUsage — structured output with token accounting
• ToolCall / FunctionDefinition — native function-calling protocol
• StreamData — incremental token chunks for streaming responses
• PromptArgs + prompt_args! — ergonomic key-value input for templates
• ChainError / LlmError / ToolError — structured error types

Quick Start

use rune_chain_core::{Chain, ChainError, GenerateResult, PromptArgs, prompt_args};
use async_trait::async_trait;

struct EchoChain;

#[async_trait]
impl Chain for EchoChain {
    async fn call(&self, input: PromptArgs) -> Result<GenerateResult, ChainError> {
        let text = input
            .get("input")
            .and_then(|v| v.as_str())
            .unwrap_or("")
            .to_string();
        Ok(GenerateResult::from_text(text))
    }
}

let chain = EchoChain;
let result = chain.invoke(prompt_args! { "input" => "hello" }).await.unwrap();
assert_eq!(result, "hello");

Macros

prompt_args

Build a PromptArgs map from key-value pairs.

Structs

FunctionDefinition

JSON-Schema-based description of a tool the LLM may call.

GenerateResult

The result of a single LLM generation call.

Message

A single turn in a conversation, carrying a Role, rich MessageContent, and optional tool-call metadata for function-calling workflows.

StreamData

A single chunk delivered by a streaming LLM response.

TokenUsage

Token consumption reported by a single LLM generation call.

ToolCall

A single tool/function call requested by the model in a generation response.

Enums

ChainError

Errors produced by a Chain during execution.

ContentPart

A single part within a multi-part message content.

ImageData

Raw image payload that can be embedded in a message.

LlmError

Errors produced by an Llm during generation.

MessageContent

The body of a Message: either plain text or a list of rich content parts.

Role

The originator of a Message in a conversation turn.

ToolChoice

How the model should choose whether to call a tool.

ToolError

Errors produced by a Tool during execution.

Traits

Chain

A composable unit of LLM work that maps PromptArgs to a GenerateResult.

Llm

An async interface to a large language model.

Memory

Persistent conversation history attached to a Chain.

Tool

An LLM-callable tool that an agent can invoke during its reasoning loop.

Type Aliases

PromptArgs

Key-value map of variables substituted into prompt templates.