World's first open-source enterprise world model.
Rust agent-memorybitemporalgraphragknowledge-baseknowledge-graphllmontologypgvectorpostgresqlragrustself-hostedsemantic-searchtemporal-knowledge-graphworld-model
Phrase alignment batches signatures by their candidates, casts no second vote after a first that says none, and asks no rules of a bare head word (#999)
WaylandYang7d73ff928083
diff --git a/crates/utopia-server/src/phrase_alignment.rs b/crates/utopia-server/src/phrase_alignment.rsindex 960f843c..b6ed98b3 100644--- a/crates/utopia-server/src/phrase_alignment.rs+++ b/crates/utopia-server/src/phrase_alignment.rs@@ -184,6 +184,64 @@ fn consider<'a>(.collect()}+/// 第二票要不要投。第二票防的是「选第一个」冒充一致,要两票一致的是**绑定**;第一票+/// 说没有一条对得上的签名,第二票怎么答都绑不上:答没有是「无」,答了一条是「拿不定」。+/// 测量库一轮 864 条签名里第一票说没有的 317 条,没有一条最后绑上(bench README,+/// 2026-09-28),这一票省下。第一票没答到的也不投:少一票本来就不下结论。+/// 补问 marks 的照旧投两票:它要两票说同一个值+fn second_vote_is_due(first: &Vote, first_answered: bool, marks_only: bool) -> bool {+ marks_only || (first_answered && first.is_some())+}++/// 类别词规则的宾语是从类别词自己的字里读出来的(「british film」读出英国),只有中心词+/// 的类别词(「state」「ship」)没有可读的字:一轮 159 个类别词里 127 个是单个词,提出+/// 的 8 条规则没有一条读得出宾语(bench README,2026-09-28)。这种不问。+///+/// 只认得出用空格分词的写法:一串 ASCII 字母是一个词。带连字符、数字的(「1968」、+/// 「british-governed」)和不分词的文字(「英国电影」)看不出有没有修饰语,照旧问+fn has_modifier(kind_word: &str) -> bool {+ !kind_word.trim().chars().all(|c| c.is_ascii_alphabetic())+}++/// 把签名分成批,候选相近的放在一起。返回每批里签名的下标。+///+/// 属性定义表在一批里只写一遍,占一次调用输入的将近一半:按到来的次序每十二条切一批时,+/// 同批的签名互不相干(一条讲出生地、一条讲导演、一条讲所属球队),各自十条候选几乎不+/// 重叠,合起来就是本体的一半(测量库上九十六条属性里约五十条)。这里只改排法:每条签名+/// 看到的候选一条不少,判断一字不变。+///+/// 贪心:每批从还没排的第一条起,每次加进让这一批的候选并集长得最少的那一条,并列时取+/// 靠前的——结果只由输入决定,同样的签名每次排成同样的批。没有候选的签名不问模型,+/// 排在最后,不占有候选的批里的位置+fn batch_by_candidates<K: Copy + Eq + std::hash::Hash>(+ sets: &[Vec<K>],+ size: usize,+) -> Vec<Vec<usize>> {+ let size = size.max(1);+ let mut left: Vec<usize> = (0..sets.len()).filter(|i| !sets[*i].is_empty()).collect();+ let empty: Vec<usize> = (0..sets.len()).filter(|i| sets[*i].is_empty()).collect();+ let mut out: Vec<Vec<usize>> = Vec::new();+ while !left.is_empty() {+ let first = left.remove(0);+ let mut batch = vec![first];+ let mut union: HashSet<K> = sets[first].iter().copied().collect();+ while batch.len() < size && !left.is_empty() {+ let (at, _) = left+ .iter()+ .enumerate()+ .map(|(at, i)| (at, sets[*i].iter().filter(|k| !union.contains(k)).count()))+ .min_by_key(|(at, added)| (*added, *at))+ .expect("left is not empty");+ let i = left.remove(at);+ union.extend(sets[i].iter().copied());+ batch.push(i);+ }+ out.push(batch);+ }+ out.extend(empty.chunks(size).map(<[usize]>::to_vec));+ out+}+/// 日志里放得下的一段回复:空白折成一个空格,最多这么多字符。const SNIPPET_CHARS: usize = 240;@@ -538,8 +596,26 @@ async fn align_phrases_locked();// 先把每批的 future 造出来再排队:直接在 map 里返回 async 块会让借用的生命周期// 满足不了 tokio::spawn 要的 Send- let futures: Vec<_> = todo- .chunks(BATCH)+ // 候选相近的签名排进同一批(见 `batch_by_candidates`):属性表一批只写一遍,同批的+ // 签名候选重叠得越多,这张表越短+ let sets: Vec<Vec<Uuid>> = todo+ .iter()+ .map(|s| {+ let fitting = &considered[&s.key()].0;+ if fitting.len() > CANDIDATE_LIMIT {+ Vec::new()+ } else {+ fitting.iter().map(|p| p.id).collect()+ }+ })+ .collect();+ let batches: Vec<Vec<&PhraseSignature>> = batch_by_candidates(&sets, BATCH)+ .into_iter()+ .map(|group| group.into_iter().map(|i| todo[i]).collect())+ .collect();+ let futures: Vec<_> = batches+ .iter()+ .map(|batch| batch.as_slice()).map(|batch| async move {let (mut bound, mut none, mut undecided, mut skipped, mut failed, mut unanswered) =(0usize, 0usize, 0usize, 0usize, 0usize, 0usize);@@ -566,6 +642,7 @@ async fn align_phrases_locked(.iter().enumerate().filter(|(i, _)| !cands[*i].is_empty())+ .filter(|(i, s)| pass == 0 || second_vote_is_due(&votes[*i].0, answered[*i].0, marks_only.contains_key(&s.key()))).map(|(i, s)| {let mut list: Vec<&RelationTypeView> = cands[i].clone();if pass == 1 {@@ -753,7 +830,9 @@ async fn align_phrases_locked(}let (a, b) = &votes[i];let (ans_a, ans_b) = answered[i];- if !ans_a || !ans_b {+ // 第一票说没有的不投第二票(见 `second_vote_is_due`):结论就是没有+ let asked_twice = second_vote_is_due(a, ans_a, false);+ if !ans_a || (asked_twice && !ans_b) {// 有一票没答到:不下结论,下次再问unanswered += 1;continue;@@ -766,7 +845,11 @@ async fn align_phrases_locked(}).unwrap_or(serde_json::Value::Null)};- let record = serde_json::json!({ "first": show(a), "second": show(b) });+ let record = if asked_twice {+ serde_json::json!({ "first": show(a), "second": show(b) })+ } else {+ serde_json::json!({ "first": null, "reason": "first_vote_none" })+ };// 两票选的属性与方向(或都说没有)是否一致let agree = match (a, b) {(Some((ka, da, _)), Some((kb, db, _))) => ka == kb && da == db,@@ -946,11 +1029,11 @@ async fn align_phrases_locked(// 类别词的候选也开短名单:词加例名嵌入后取最近的属性;没有向量时看全部let fresh: Vec<&utopia_store::type_bindings::KindWordSignature> = kind_words.iter()- .filter(|k| !asked_kind.contains(k.kind_word.as_str()))+ .filter(|k| !asked_kind.contains(k.kind_word.as_str()) && has_modifier(&k.kind_word)).collect();let kind_short = shortlist_kind_words(state, settings, kb_id, &fresh).await?;for (k, basis) in kind_words.iter().zip(kind_basis.iter()) {- if asked_kind.contains(k.kind_word.as_str()) {+ if asked_kind.contains(k.kind_word.as_str()) || !has_modifier(&k.kind_word) {continue;}let mut candidates: Vec<&RelationTypeView> = props@@ -1071,6 +1154,54 @@ mod lifecycle_tests;#[cfg(test)]mod tests {+ #[test]+ fn a_kind_word_that_is_only_a_head_word_has_nothing_to_read() {+ for bare in ["state", "ship", " settlement "] {+ assert!(!has_modifier(bare), "{bare}");+ }+ for readable in [+ "british film",+ "1968",+ "british-governed",+ "英国电影",+ "wine area",+ ] {+ assert!(has_modifier(readable), "{readable}");+ }+ }++ #[test]+ fn signatures_with_the_same_candidates_share_a_batch() {+ // 六条签名、三种候选集,交错着来;每批两条+ let sets: Vec<Vec<u32>> = vec![+ vec![1, 2],+ vec![7, 8],+ vec![],+ vec![2, 1],+ vec![8, 9],+ vec![1, 2, 3],+ ];+ let batches = batch_by_candidates(&sets, 2);+ // 每条签名恰好出现一次+ let mut all: Vec<usize> = batches.iter().flatten().copied().collect();+ all.sort_unstable();+ assert_eq!(all, vec![0, 1, 2, 3, 4, 5]);+ // 候选一样的排在一起,没有候选的在最后+ assert_eq!(batches, vec![vec![0, 3], vec![1, 4], vec![5], vec![2]]);+ // 同样的输入排成同样的批+ assert_eq!(batch_by_candidates(&sets, 2), batches);+ // 一批的候选并集比按次序切的小+ let union = |b: &Vec<usize>| {+ b.iter()+ .flat_map(|i| sets[*i].iter())+ .collect::<HashSet<_>>()+ .len()+ };+ let in_order: usize = [vec![0, 1], vec![2, 3], vec![4, 5]].iter().map(union).sum();+ let grouped: usize = batches.iter().map(union).sum();+ assert!(grouped < in_order, "{grouped} < {in_order}");+ }+#[test]fn a_shortlist_keeps_label_matches_and_fills_by_distance_within_the_fitting_set() {let ids: Vec<Uuid> = (0..6).map(|_| Uuid::now_v7()).collect();diff --git a/crates/utopia-server/src/phrase_alignment_tests.rs b/crates/utopia-server/src/phrase_alignment_tests.rsindex a63f0a04..cfa15275 100644--- a/crates/utopia-server/src/phrase_alignment_tests.rs+++ b/crates/utopia-server/src/phrase_alignment_tests.rs@@ -322,7 +322,7 @@ fn bound() -> Vec<Value> {]}fn none() -> Vec<Value> {- vec![vote(None, None), vote(None, None), nothing_implied()]+ vec![vote(None, None)]}/// 类别词对齐还排着时短语对齐不动手:两端的类还没定,判了也是重判。它给自己排一份半分钟@@ -591,8 +591,8 @@ async fn an_edit_during_the_model_request_leaves_the_decision_stale() -> anyhow:f.run().await?;assert_eq!(f.requests().len(),- 4,- "two votes, then two votes again (a none is not asked for rules): the basis differs, not the clock"+ 3,+ "one vote that says none, then two votes (a none is not asked for rules): the basis differs, not the clock");assert_eq!(f.binding().await?.status, "bound");anyhow::Ok(())@@ -657,6 +657,29 @@ async fn a_person_decision_made_during_the_request_is_not_overwritten() -> anyho/// 类别词那边 bench 里「12 篇文档一个词都没绑上」的回复形状搬到短语上:DeepSeek-V3.2/// 把整段答成按 id 作键的对象,第二票把 `b` 写成对象、值写成 `{"key","direction"}`。+/// 第一票说没有,结论就是没有:第二票不投,也不问规则+#[tokio::test]+async fn a_first_vote_that_says_none_is_not_followed_by_a_second() -> anyhow::Result<()> {+ let Some(f) = Fx::new().await? else {+ return Ok(());+ };+ let run = async {+ f.script(vec![+ vote(None, None),+ vote(Some("based_in"), Some("forward")),+ ]);+ f.run().await?;+ assert_eq!(f.requests().len(), 1, "one vote");+ assert_eq!(f.binding().await?.status, "none");+ assert_eq!(f.reason().await?.as_deref(), Some("first_vote_none"));+ assert_eq!(f.typed().await?, 0);+ anyhow::Ok(())+ }+ .await;+ f.cleanup().await?;+ run+}+/// 两票都得读出来,签名才绑得上#[tokio::test]async fn an_id_keyed_reply_still_binds_the_phrase() -> anyhow::Result<()> {@@ -702,7 +725,11 @@ async fn an_unreadable_reply_is_asked_again_a_bounded_number_of_times() -> anyhojson!({"answer": "based_in", "direction": "forward"}),]);f.run().await?;- assert_eq!(f.requests().len(), 2, "two votes, neither readable");+ assert_eq!(+ f.requests().len(),+ 1,+ "the first vote is unreadable, the second is not cast"+ );assert!(phrase_bindings::bindings(&f.pool, f.kb).await?.is_empty(),"an unreadable reply writes no decision"@@ -722,7 +749,7 @@ async fn an_unreadable_reply_is_asked_again_a_bounded_number_of_times() -> anyhof.clear_jobs().await?;// 已经是最后一次自己排的:不再排align_phrases_reasking(&f.state, f.kb, MAX_REASK).await?;- assert_eq!(f.requests().len(), 4, "the last allowed round still asks");+ assert_eq!(f.requests().len(), 2, "the last allowed round still asks");let after: i64 = sqlx::query_scalar("SELECT count(*) FROM jobs WHERE kind='align_phrases' AND payload->>'kb_id'=$1",)diff --git a/scripts/bench/README.md b/scripts/bench/README.mdindex ae6e0e8a..0f5aba1a 100644--- a/scripts/bench/README.md+++ b/scripts/bench/README.md@@ -830,3 +830,37 @@ bound / none / undecided。0044 的门槛:裁判精度不低于完整原型的新签名、每条约 1.4 千 token,这就是它的地板。0044 的 token 门槛在 Re-DocRED 上要么换口径(每条新签名的成本),要么换同领域语料再量。++### 第六次真跑:对齐少问一些(2026-09-28,dev @ b036f45,claude-haiku-4-5 经 apexin.net,bge-m3,20 篇)++20 篇按调用分(`typed.mjs` 的 token 表):短语对齐的两票占 38%,抽取 20%,提规则 14%,类别词对齐 9%。+短语对齐一次调用的输入约 4600 token,其中属性表将近一半——按到来次序切批时,同批签名的候选几乎不重叠。+三处改动:++1. **按候选分批**:候选相近的签名排进同一批。只改排法,每条签名看到的候选不变。+2. **第一票说没有的不投第二票**:第二票防的是「选第一个」冒充一致;第一票说没有,第二票怎么答都绑不上。+3. **只有中心词的类别词不问规则**:类别词规则的宾语从类别词自己的字里读(「british film」读出英国)。+ 基线一轮问了 159 个类别词,127 个是单个词,提出的 8 条规则(「state」→ 国家)没有一条读得出宾语。++两轮完整的 20 篇之间抽取自己就不一样(签名 229 对 210,只有 137 条相同),整轮的数字分不出对齐的+变化。所以**在同一个库的同一批陈述上**把绑定清掉重判,210 条签名,每种构建各判一遍或两遍:++| | 基线 | 基线 再一遍 | 分批 | 分批 再一遍 | 三处都改 |+|---|---|---|---|---|---|+| 短语对齐调用 | 46 | 47 | 36 | 40 | 32 |+| 输入 token | 190572 | 199086 | 128338 | 143900 | 103412 |+| 输出 token | 46537 | 45422 | 35809 | 36644 | 29703 |+| 合计 | 237109 | 244508 | 164147 | 180544 | 133115 |+| 签名 bound / none / undecided | 53 / 110 / 47 | 66 / 108 / 36 | 69 / 108 / 33 | 67 / 109 / 34 | 63 / 119 / 28 |+| 类型化事实 | 82 | 87 | 99 | 92 | 84 |++读数:+- 分批省短语对齐的三成(24.1 万 → 17.2 万),加上条件第二票省四成五(→ 13.3 万)。+- **模型自己的波动比改动大**:同一个构建、同一批签名、温度 0,两遍判定完全相同的只有 166 / 210(基线)+ 和 163 / 210(分批)。基线对分批是 146 到 159。分批让判定多变了约十条,方向两边都有(绑上的数和+ 类型化行数没有变少);20 篇整轮的裁判精度 79.3% 对 78.8%。+- 条件第二票的代价:它记成「无」的 98 条,在另外四遍里 81 到 86 条本来也是「无」,8 到 12 条是+ 「拿不定」(第二票答了一条属性),4 到 5 条绑上了(那一遍第一票答的不一样,是波动)。「拿不定」的那+ 十条上下不再进人的队列,`votes.reason = "first_vote_none"` 查得到。+- 20 篇整轮(只有分批):合计 82.1 万 → 62.6 万,每篇 4.1 万 → 3.1 万;其中约一半来自分批,其余是+ 这一轮抽出的签名少、重问少。