Struct LfmNormalizer 

pub struct 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.

Recognised format:

<|tool_call_start|>[get_weather(city='Berlin')]<|tool_call_end|>

Multiple calls per block ([fn1(a=1), fn2(b=2)]) and multiple blocks per message are both handled.

Example

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

let raw = "<|tool_call_start|>[add(x=3, y=4)]<|tool_call_end|>";
let calls = LfmNormalizer.normalize(raw).unwrap();
assert_eq!(calls[0].name, "add");
assert_eq!(calls[0].args, json!({"x": 3, "y": 4}));

Trait Implementations

impl Clone for LfmNormalizer

fn clone(&self) -> LfmNormalizer

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for LfmNormalizer

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for LfmNormalizer

fn default() -> LfmNormalizer

Returns the “default value” for a type.

impl ToolCallNormalizer for LfmNormalizer

fn is_applicable(&self, raw: &str) -> bool

Quick scan: does raw contain markers this normalizer handles?

fn normalize(&self, raw: &str) -> Result<Vec<ToolInvocation>, KernelError>

Parse raw text into zero or more tool invocations.