World's first open-source enterprise world model.
Rust agent-memorybitemporalgraphragknowledge-baseknowledge-graphllmontologypgvectorpostgresqlragrustself-hostedsemantic-searchtemporal-knowledge-graphworld-model
The tab strip starts where the rows do (#771)
WaylandYang03417131d151
diff --git a/crates/utopia-ingest/src/lib.rs b/crates/utopia-ingest/src/lib.rsindex 94a76e44..b340695c 100644--- a/crates/utopia-ingest/src/lib.rs+++ b/crates/utopia-ingest/src/lib.rs@@ -173,10 +173,12 @@ mod table;pub mod transcript;pub use chunker::{chunk_segments, chunk_text, chunk_with_budget, ChunkPiece, BUDGET_TOKENS};+// 抽取那一侧要按值找回它那一列的表头(#729):写出这个表格式的模块,也负责读回它/// Decode fetched text with the same encoding detection as file ingestion.pub use parsers::plain_text as decode_text;pub use provenance::{Origin, Provenance, Segment};pub use reading::Reading;+pub use table::column_header;/// 解析产物:纯文本 + 可选结构信息。#[derive(Debug)]diff --git a/crates/utopia-ingest/src/table.rs b/crates/utopia-ingest/src/table.rsindex dc03b3ab..e56a5c9d 100644--- a/crates/utopia-ingest/src/table.rs+++ b/crates/utopia-ingest/src/table.rs@@ -603,10 +603,160 @@ fn collapse_spans(cols: Vec<usize>, headers: &[Row], data: &[(String, Row)]) ->out}+/// 一行切成格:每格的文字,以及它在这一行里的字节偏移。行首行尾的空段不算格。+/// **带着偏移**,因为表头里两列同名是常事(两个「Amount」),光按文字回头找会指错格+fn cells(row: &str) -> Vec<(&str, usize)> {+ let lead = row.len() - row.trim_start().len();+ let trimmed = row.trim();+ let (inner, open) = match trimmed.strip_prefix('|') {+ Some(rest) => (rest, 1),+ None => (trimmed, 0),+ };+ let inner = inner.strip_suffix('|').unwrap_or(inner);+ let mut at = lead + open;+ let mut out = Vec::new();+ for piece in inner.split('|') {+ let pad = piece.len() - piece.trim_start().len();+ out.push((piece.trim(), at + pad));+ at += piece.len() + 1;+ }+ out+}++/// 比较一格与一个值:空白、千分位逗号、货币符号不算数,别的一个字都不能差。+/// **不做包含**:一格写着 `$1,234`,值是 `234` 的那一条不该认领它+fn same_number(cell: &str, value: &str) -> bool {+ let strip = |s: &str| {+ s.chars()+ .filter(|c| !c.is_whitespace() && !is_currency(*c) && *c != ',')+ .collect::<String>()+ };+ let (a, b) = (strip(cell), strip(value));+ !a.is_empty() && a == b+}++/// 这个值落在这一行的哪一列,以及那一列的表头是什么(#729:财务表格单元的期间列)。+///+/// 一张财务表的列头写着期间(「Three Months Ended July 26, 2026」),值写在行里。**人是+/// 按列读表的,抽取却是按行读的**:一行五个数进了图,五条陈述除了值一模一样,没有任何+/// 东西说得出哪个数属于哪个期间。这里把那一列的表头找回来,交给时间解析去判它是不是一个+/// 期间——代码不认「Months Ended」这类字眼,只认位置。+///+/// 它宁可什么都不给,也不给错的:+///+/// * 这一行在这一块里出现不止一次——说不清是哪张表的哪一行;+/// * 这一行里没有恰好一格等于这个值——两格相等就分不出是哪一列;+/// * 这一行上面不是 `| --- |` 分隔线,或分隔线上面不是表头行;+/// * 表头那一格是空的,或一个数字都没有(「Amount」不是期间,而任何期间都会写出年份、+/// 日期或季度的数字)。+///+/// 返回表头那一格的文字和它在这一块里的字节偏移——一条提及要说得出自己在原文的哪个位置。+pub fn column_header<'a>(chunk: &'a str, row: &str, value: &str) -> Option<(&'a str, usize)> {+ let row = row.trim();+ if !row.starts_with('|') {+ return None;+ }+ let at = chunk.find(row)?;+ if chunk[at + row.len()..].contains(row) {+ return None;+ }+ let mut hit = cells(row)+ .into_iter()+ .enumerate()+ .filter(|(_, (c, _))| same_number(c, value));+ let (col, _) = hit.next()?;+ if hit.next().is_some() {+ return None;+ }+ // 紧挨着这一行往上:一条分隔线,再上去一行是表头。位置用 rfind 从原文里量,+ // 不靠换行符占几个字节+ let before = &chunk[..at];+ let mut lines = before.lines().rev();+ let sep = lines.next()?;+ if !sep.trim_start().starts_with("| ---") {+ return None;+ }+ let head = lines.next()?;+ if !head.trim_start().starts_with('|') {+ return None;+ }+ let sep_at = before.rfind(sep)?;+ let head_at = before[..sep_at].rfind(head)?;+ let (cell, in_line) = *cells(head).get(col)?;+ if cell.is_empty() || !cell.chars().any(|c| c.is_ascii_digit()) {+ return None;+ }+ let from = head_at + in_line;+ Some((&chunk[from..from + cell.len()], from))+}+#[cfg(test)]mod tests {use super::*;+ /// #729:一行五个数,每个数的期间在它那一列的表头上+ const COSTS: &str = "(A) Acquisition-related and other costs are comprised of amortization.\n\n\+| | Three Months Ended July 26, 2026 | Three Months Ended April 26, 2026 | Six Months Ended July 26, 2026 |\n\+| --- | --- | --- | --- |\n\+| Cost of revenue | $46 | $47 | $93 |\n\+| Sales, general and administrative | $6 | $5 | $11 |\n";++ #[test]+ fn a_cell_takes_the_period_of_its_own_column() {+ let row = "| Cost of revenue | $46 | $47 | $93 |";+ let (head, at) = column_header(COSTS, row, "$46").expect("46 是第一列的值");+ assert_eq!(head, "Three Months Ended July 26, 2026");+ assert_eq!(&COSTS[at..at + head.len()], head, "偏移指回原文那一格");+ assert_eq!(+ column_header(COSTS, row, "$47").unwrap().0,+ "Three Months Ended April 26, 2026",+ "同一行的下一个数换了一列,期间也跟着换"+ );+ assert_eq!(+ column_header(COSTS, row, "93").unwrap().0,+ "Six Months Ended July 26, 2026",+ "货币符号与千分位不算数"+ );+ }++ #[test]+ fn a_header_that_names_no_time_gives_nothing() {+ let t = "| Item | Amount |\n| --- | --- |\n| Cost of revenue | $46 |\n";+ assert_eq!(column_header(t, "| Cost of revenue | $46 |", "$46"), None);+ }++ #[test]+ fn an_ambiguous_cell_gives_nothing() {+ // 同一行两格相同:说不出是哪一列+ let t = "| | Q1 2026 | Q2 2026 |\n| --- | --- | --- |\n| Cost | $46 | $46 |\n";+ assert_eq!(column_header(t, "| Cost | $46 | $46 |", "$46"), None);+ // 同一块里两张表有一模一样的行:说不出是哪张表+ let twice = format!("{COSTS}\n{COSTS}");+ assert_eq!(+ column_header(&twice, "| Cost of revenue | $46 | $47 | $93 |", "$46"),+ None+ );+ }++ #[test]+ fn a_row_with_no_table_above_it_gives_nothing() {+ // 没有分隔线就没有表头+ let t = "| Cost of revenue | $46 |\n";+ assert_eq!(column_header(t, "| Cost of revenue | $46 |", "$46"), None);+ // 引文不是一行表格+ assert_eq!(+ column_header(COSTS, "Acquisition-related and other costs", "$46"),+ None+ );+ // 一格写着 $1,234,值是 234 的不该认领它+ let t2 = "| | Q1 2026 |\n| --- | --- |\n| Cost | $1,234 |\n";+ assert_eq!(column_header(t2, "| Cost | $1,234 |", "234"), None);+ assert_eq!(+ column_header(t2, "| Cost | $1,234 |", "1234").unwrap().0,+ "Q1 2026"+ );+ }+fn render(html: &str) -> String {let (_, tables) = lift_tables(html);tables.join("\n=====\n")diff --git a/crates/utopia-server/src/extraction_open.rs b/crates/utopia-server/src/extraction_open.rsindex b6d2da2d..1bb4e658 100644--- a/crates/utopia-server/src/extraction_open.rs+++ b/crates/utopia-server/src/extraction_open.rs@@ -59,6 +59,20 @@ fn locate(hay: &str, needle: &str) -> Option<(i32, i32)> {/// 时间词在块里的字符起点。**必须在这条陈述自己的那句引文里**:模型会把一个时间词挂到/// 好几条陈述上(FDA 语料实测「week 4」挂到了「不应由过敏患者服用」上,六条带时间的陈述错了三条),/// 整块里搜得到不等于这句说了它。引文里有、但引文本身没在块里定位到的,起点退回整块里的第一处+/// 这条陈述的值有没有真的写在它的引文里(#729)。+///+/// 千分位逗号、空白和货币符号两边都不算数;别的按原样比。**故意放得松**:这是一条丢弃+/// 规则,放过一条可疑的,好过丢掉一条对的+fn shows_value(quote: &str, value: &str) -> bool {+ let strip = |s: &str| {+ s.chars()+ .filter(|c| !c.is_whitespace() && *c != ',' && !matches!(c, '$' | '¥' | '€' | '£'))+ .collect::<String>()+ };+ let v = strip(value);+ v.is_empty() || strip(quote).contains(&v)+}+fn locate_time(chunk: &str, quote: Option<(&str, Option<(i32, i32)>)>, words: &str) -> Option<i32> {let (q, span) = quote?;if let Some((inner, _)) = locate(q, words) {@@ -376,6 +390,9 @@ pub(crate) async fn run_open(.as_deref().map(str::trim).filter(|v| v.chars().any(char::is_alphanumeric));+ // 模型自己写的那个值,在下面被「宾语没声明」顶替之前记下来:引文要核的是它,+ // 不是一个实体的名字(#729)+ let stated_value = value;let object = match s.object.as_deref().map(str::trim).filter(|o| !o.is_empty()) {Some(name) => match place(pool,@@ -478,6 +495,23 @@ pub(crate) async fn run_open().await;}+ // **值必须出现在它自己的引文里**(#729)。一行五列被压成五条只有值不同的+ // 陈述、引文却是表上面那句导语时,五条里至多一条对,而谁也分不出是哪一条。+ // 只查带数字的值:实体名当值写的那一类(object_undeclared)不在此列+ if let (Some(v), Some((q, _))) = (stated_value, quote) {+ if v.chars().any(|c| c.is_ascii_digit()) && !shows_value(q, v) {+ drop_signal(+ state,+ kb_id,+ document_id,+ reason::VALUE_NOT_IN_QUOTE,+ "a statement's value is not in its own quoted sentence",+ Some(v),+ )+ .await;+ continue;+ }+ }// 开放陈述没有模型自报的置信度:它说的是「文档这么说了」。看图描述出来的块// 照旧压上限(0040 决定 4)let confidence = origin_ceiling(&chunk.origin, 1.0);@@ -526,6 +560,20 @@ pub(crate) async fn run_open(}}}+ // **表格单元的期间在它那一列的表头上**(#729)。模型自己说了时间就不动它——+ // 它看得见整块原文,说得出的比一根列头多。这一条不是模型报的,所以不走+ // `time_not_in_quote`:它的出处是位置(这个值在这一行的第几格),+ // 而那一格的表头上写着期间。它是不是一个期间,由时间解析去判(0045:+ // 模型读、代码算),这里一个字眼都不认+ if !time_words.iter().any(|(_, _, role)| *role == "when") {+ if let (Some(v), Some((q, _))) = (stated_value, quote) {+ if let Some((head, at)) = utopia_ingest::column_header(&chunk.text, q, v) {+ if let Ok(at) = i32::try_from(at) {+ time_words.push((head, at, "when"));+ }+ }+ }+ }if await_nod {// 记忆日志:一切原样进待确认表(0015),人点头时 `pending::confirm` 才把它// 落成开放陈述——同样的短语、限定、时间词、引文偏移@@ -769,6 +817,31 @@ pub(crate) async fn run_open(mod tests {use super::*;+ /// #729:引文里没有这个数,这条陈述就没有依据。丢弃规则要宁松勿紧+ #[test]+ fn a_value_must_be_in_the_sentence_that_is_quoted() {+ let row = "| Cost of revenue | $46 | $47 | $93 |";+ assert!(shows_value(row, "$46"));+ assert!(shows_value(row, "46"), "货币符号不算数");+ assert!(+ shows_value("| Accounts payable | 1,915 |", "1915"),+ "千分位不算数"+ );+ assert!(shows_value("revenue was $ 1,234", "$1,234"), "空白不算数");+ assert!(+ !shows_value(+ "(A) Acquisition-related costs are comprised of amortization.",+ "$46"+ ),+ "导语里没有这个数——正是压平表格那一行的错法"+ );+ assert!(shows_value("anything", ""), "没有值就没什么可核的");+ assert!(+ shows_value("| a | 4,600 |", "46"),+ "放得松:像是命中就放过,丢错一条比留下一条可疑的更糟"+ );+ }+#[test]fn offsets_are_in_characters_not_bytes() {let text = "北京的冬天很冷。去年冬天下了三场雪。";diff --git a/crates/utopia-store/src/extraction_drops.rs b/crates/utopia-store/src/extraction_drops.rsindex e25d92b4..b68e613a 100644--- a/crates/utopia-store/src/extraction_drops.rs+++ b/crates/utopia-store/src/extraction_drops.rs@@ -27,6 +27,12 @@ pub mod reason {pub const MALFORMED_ITEM: &str = "malformed_item";/// 模型输出被截断(撞上 max_tokens)→ 已完整的那些留下,尾巴丢掉pub const TRUNCATED_REPLY: &str = "truncated_reply";+ /// 这条陈述的值不在它自己的引文里(#729):证据没写着这个数。+ ///+ /// 表格是这个错的产地——模型把一行五列压成五条只有值不同的陈述,引文却指向表上面+ /// 那句导语。五条里至多一条对,而图上没有任何东西分得出是哪一条。**引文是这条陈述+ /// 的全部依据**:值不在里面,它就不是文档说的+ pub const VALUE_NOT_IN_QUOTE: &str = "value_not_in_quote";/// 模型报的别名(或它的引文)不在这一块原文里(0041 决定 2):不记这个名字。/// 名字是召回的桥,一座凭空的桥会把两个不相干的实体接到一起pub const NAME_NOT_IN_TEXT: &str = "name_not_in_text";diff --git a/web/src/pages/Graph.tsx b/web/src/pages/Graph.tsxindex 73e16a46..b6d2fef9 100644--- a/web/src/pages/Graph.tsx+++ b/web/src/pages/Graph.tsx@@ -2894,13 +2894,14 @@ function EntityPanel({{/* 视图切换:Relations(一张表,过去的折在组尾)| History(记录轴)| Derived */}- {/* **左边比头部多一档**(24 而不是 16)。两个盒子本来都从 px-4 起,可- 分段控件自己还有一圈内距(p-1 加按钮的 px-2),于是「Relations」四个字- 落在 29,而标题「OpenAI」落在 35——标题前面是色点加 gap,正好差 6px,- 看着就是这一排比标题往左漏出去一截。加一档之后字落在 37,压回标题上。- 这一条只给这里:本体页那条是 `fill` 的整条,与下面正文同宽,- 它的左缘该跟正文对齐,不跟标题对齐 */}- <div className="pl-6 pr-4 pt-3">+ {/* **左缘与下面那一列折叠箭头同起**(16)。它们上下紧挨着,中间没有别的东西,+ 于是这一段读起来是一条竖线;从前对的是标题里的字(pl-6,让「Relations」+ 四个字落在 37、压住标题),可标题在另一个盒子里、还隔着一条分隔线,那条+ 对齐看不见,看得见的是它比下面整整齐齐的一列往右缩了一档。+ 16 = 滚动区的 px-2 加行自己的 px-2,也就是每一行底色开始的地方;+ 12 的箭头在 16 的图标格里居中,所以字形还在里面 2px——那 2px 是这个+ 产品里每一个图标都有的,不是这一处没对齐 */}+ <div className="pl-4 pr-4 pt-3"><Segmentedsize="sm"value={view}