Files
qingjian/apps/linux/server/tests/rescoring.rs
T
yenharvey a41d1cf82a feat: 接入含章·通变与知微模型
接入含章·通变的拼音解码、整句重排与词图兜底生成,扩展 macOS、Windows 和 Linux 的模型加载路径。模型文件按正式标识归档,保留旧用户目录回退,并同步数据打包、评测记录和用户文档。
2026-09-29 12:02:00 +08:00

213 lines
7.6 KiB
Rust

//! 本地整句模型:停顿后重排换首选、动过高亮不换、模型在组句中途接上也补这一轮;
//! 插件定时 Poll 回的帧内容没变时展示身份不变,重排换了顺序才推进。
use std::time::{Duration, Instant};
use qingjian_core::Engine;
use qingjian_core::sentence::SentenceScorer;
use qingjian_dictionary::Dictionary;
use qingjian_linux_server::{Router, RouterConfig};
use qingjian_platform::protocol::{
ClientMessage, Frame, KeyEvent, KeyModifiers, PROTOCOL_VERSION, ServerMessage, SessionId,
};
use serde_json::{Value, json};
const SESSION: SessionId = SessionId(1);
/// 假打分器:偏爱某个文本,其余都给低分(与 Core / Windows Server 的重打分测试同款)。
struct Prefers(&'static str);
impl SentenceScorer for Prefers {
fn score(&self, _context: &str, _keys: &str, texts: &[&str]) -> Vec<f64> {
texts
.iter()
.map(|t| if *t == self.0 { -1.0 } else { -20.0 })
.collect()
}
}
/// k 优路径按末词分状态,几条路径要在末词上不同才都留下来:ni + ta → 你他 / 你她 / 你它。
fn engine() -> Engine {
Engine::new(Dictionary::parse("你\tni\t100\n他\tta\t100\n她\tta\t90\n它\tta\t80\n").unwrap())
}
fn router_with(engine: Engine) -> Router {
let mut router = Router::new(engine, RouterConfig::default());
router.handle(ClientMessage::OpenSession {
session: SESSION,
app: None,
protocol: PROTOCOL_VERSION,
});
router.handle(ClientMessage::Privacy {
session: SESSION,
private: false,
});
router
}
/// 接了假模型的 Router:本地整句模型在 Server 里是异步接法,按键先按词级出候选,停顿后 tick 才换。
fn router_with_scorer(preferred: &'static str) -> Router {
let mut engine = engine();
engine.set_async_sentence_scorer(Some(Box::new(Prefers(preferred))));
router_with(engine)
}
fn press(router: &mut Router, event: KeyEvent) -> Frame {
match router.handle(ClientMessage::Key {
session: SESSION,
event,
}) {
Some(ServerMessage::KeyResult { frame, .. }) => frame,
other => panic!("expected KeyResult, got {other:?}"),
}
}
fn type_letters(router: &mut Router, text: &str) -> Frame {
let mut last = Frame::default();
for c in text.chars() {
last = press(
router,
KeyEvent::new(
c.to_ascii_uppercase() as u32,
Some(c),
KeyModifiers::default(),
),
);
}
last
}
fn first(frame: &Frame) -> Option<&str> {
frame.candidates.items.first().map(|c| c.text.as_str())
}
fn poll(router: &mut Router) -> Frame {
match router.handle(ClientMessage::Poll { session: SESSION }) {
Some(ServerMessage::Update { frame, .. }) => frame,
other => panic!("expected Update, got {other:?}"),
}
}
/// 一直 tick 到首选变成 `text` 或等满 `timeout`;返回最后一帧。
fn tick_until_first(router: &mut Router, text: &str, timeout: Duration) -> Frame {
let started = Instant::now();
loop {
std::thread::sleep(router.next_tick().min(Duration::from_millis(20)));
router.tick();
let frame = poll(router);
if first(&frame) == Some(text) || started.elapsed() > timeout {
return frame;
}
}
}
#[test]
fn local_model_rescoring_reorders_sentence_after_pause() {
let mut router = router_with_scorer("你它");
let frame = type_letters(&mut router, "nita");
// 按键时只按词级模型:他 的词频高,首选是「你他」
assert_eq!(first(&frame), Some("你他"));
// 在等防抖,主循环该在 80 ms 内醒来
assert!(router.next_tick() <= Duration::from_millis(80));
let frame = tick_until_first(&mut router, "你它", Duration::from_secs(3));
assert_eq!(
first(&frame),
Some("你它"),
"停顿后模型偏爱的整句应换到首位,实际:{:?}",
frame.candidates.items
);
// 换完不再等;空闲节拍回到落盘学习的一秒
assert_eq!(router.next_tick(), Duration::from_secs(1));
}
#[test]
fn local_model_does_not_touch_a_navigated_page() {
let mut router = router_with_scorer("你它");
type_letters(&mut router, "nita");
// 用户动过高亮(下方向键):模型的结果只留在缓存里,不换正在看的这页
let frame = press(
&mut router,
KeyEvent::new(0x28, None, KeyModifiers::default()),
);
assert_eq!(first(&frame), Some("你他"));
// 防抖 80 ms + 假模型立即回分,300 ms 足够等到结果;首选仍是原来的
let frame = tick_until_first(&mut router, "你它", Duration::from_millis(300));
assert_eq!(first(&frame), Some("你他"));
}
#[test]
fn model_attached_during_composition_rescores_the_current_round() {
// 模型还在后台加载时用户已经在组句:这一轮的查询从没见过打分器
let mut router = router_with(engine());
let frame = type_letters(&mut router, "nita");
assert_eq!(first(&frame), Some("你他"));
assert_eq!(router.next_tick(), Duration::from_secs(1));
// 加载完成接上:不用再敲键,这一轮也补一次重排
router.attach_sentence_scorer(Box::new(Prefers("你它")));
assert!(router.next_tick() <= Duration::from_millis(80));
let frame = tick_until_first(&mut router, "你它", Duration::from_secs(3));
assert_eq!(first(&frame), Some("你它"));
}
/// 插件走 JSON 通道:握手报了展示身份,之后每帧都带 `identity`。
fn poll_json(router: &mut Router) -> Value {
router
.handle_linux(json!({"Poll": {"session": 1}}))
.expect("poll answers")
}
#[test]
fn poll_keeps_the_display_identity_until_the_frame_changes() {
let mut router = router_with(engine());
router.handle_linux(json!({"DisplayReporting": {"session": 1, "identity": {"generation": 2, "context": "first", "revision": 0}}}));
let mut shown = Value::Null;
for c in "nita".chars() {
let key = ClientMessage::Key {
session: SESSION,
event: KeyEvent::new(
c.to_ascii_uppercase() as u32,
Some(c),
KeyModifiers::default(),
),
};
shown = router
.handle_linux(serde_json::to_value(key).unwrap())
.unwrap()["KeyResult"]["identity"]
.take();
}
// 组句期间插件每 80 ms 问一次:内容没变就沿用身份,插件不重画、不重复回报曝光
for _ in 0..3 {
let update = poll_json(&mut router);
assert_eq!(update["Update"]["identity"], shown);
assert_eq!(
update["Update"]["frame"]["candidates"]["items"][0]["text"],
"你他"
);
}
// 模型接上、重排换了顺序:这一帧才是新的展示,版本号推进
router.attach_sentence_scorer(Box::new(Prefers("你它")));
let started = Instant::now();
let changed = loop {
std::thread::sleep(Duration::from_millis(20));
router.tick();
let update = poll_json(&mut router);
if update["Update"]["frame"]["candidates"]["items"][0]["text"] == "你它"
|| started.elapsed() > Duration::from_secs(3)
{
break update;
}
};
assert_eq!(
changed["Update"]["frame"]["candidates"]["items"][0]["text"],
"你它"
);
assert!(changed["Update"]["identity"]["revision"].as_u64() > shown["revision"].as_u64());
// 换完之后再问:又是沿用
assert_eq!(
poll_json(&mut router)["Update"]["identity"],
changed["Update"]["identity"]
);
}