Module normalizer 

Tool-call normalizer: converts raw model text output (in-band markers) into structured normalizer::ToolInvocations for kernel dispatch. ToolCallNormalizer — converts raw LLM text output into structured ToolInvocations.

Models served via OpenAI-compatible APIs (e.g. mlx_lm.server) sometimes emit tool-intent as in-band text markers rather than the structured tool_calls JSON field. Normalizers detect and decode those markers so the kernel can dispatch them identically to first-class tool calls.

Built-in implementations

Type	Format
LfmNormalizer	LiquidAI LFM <|tool_call_start|>[fn(k=v)]<|tool_call_end|>
StructuredToolCallNormalizer	OpenAI Responses function_call output and Chat Completions tool_calls

Example

use rig_compose::normalizer::{LfmNormalizer, ToolCallNormalizer};

let raw = "<|tool_call_start|>[get_weather(city='Berlin')]<|tool_call_end|>";
let normalizer = LfmNormalizer;
let calls = normalizer.normalize(raw).expect("parse ok");
assert_eq!(calls[0].name, "get_weather");

Structs

LfmNormalizer

Normalizer for LiquidAI LFM models (e.g. LFM2.5-1.2B-Thinking) served through mlx_lm.server or similar OpenAI-compatible shims that emit tool intent as in-band text rather than the structured tool_calls field.

StructuredToolCallNormalizer

Normalizer for structured tool-call JSON returned by common provider APIs.

ToolInvocation

A structured tool invocation extracted from raw model output.

ToolInvocationResult

Result of dispatching one normalized ToolInvocation.

Enums

ToolDispatchAction

Decision returned by a ToolDispatchHook before a tool invocation runs.

Traits

ToolCallNormalizer

Normalizes raw LLM text output into structured ToolInvocations.

ToolDispatchHook

Hook for policy, accounting, and tracing around normalized tool dispatch.

Functions

dispatch_tool_invocations

Dispatch normalized tool invocations sequentially through a ToolRegistry.

dispatch_tool_invocations_with_hooks

Dispatch normalized tool invocations with policy/accounting hooks.