World's first open-source enterprise world model.
Rust agent-memorybitemporalgraphragknowledge-baseknowledge-graphllmontologypgvectorpostgresqlragrustself-hostedsemantic-searchtemporal-knowledge-graphworld-model
A second question should not re-find what the first one already found (#94)
WaylandYange4f38a553474
diff --git a/crates/utopia-server/src/api/chat.rs b/crates/utopia-server/src/api/chat.rsindex fc6129b2..e6857048 100644--- a/crates/utopia-server/src/api/chat.rs+++ b/crates/utopia-server/src/api/chat.rs@@ -20,6 +20,10 @@ use crate::llm_util;use crate::retrieval;use crate::state::AppState;+/// 回放几个已认下的实体。上限是因为长会话会攒出几十个,全贴回去就把+/// 省下来的上下文又花掉了;按首次出现排序,早认下的通常是这场对话的主角。+const KNOWN_ENTITY_LIMIT: usize = 20;+const MAX_HISTORY: usize = 20;const MAX_ROUNDS: usize = 6;const SEARCH_TOP_K: usize = 6;@@ -262,9 +266,10 @@ pub async fn chat(&query,&json!([]),&json!([]),+ &json!([]),).await?;- let history: Vec<(String, String)> = utopia_store::conversations::recent_context(+ let (history, known_entities) = utopia_store::conversations::recent_context(&state.pool,conversation_id,MAX_HISTORY as i64,@@ -301,6 +306,37 @@ pub async fn chat(for (role, content) in &history {msgs.push(json!({ "role": role, "content": content }));}+ // **前几轮已经认下的实体,连 id 一起交回去。**+ //+ // 少了这一段,模型只看得见上一轮的最终答案文字,不知道自己搜过什么、+ // 拿到过哪些 id,于是从名字重搜一遍。更隐蔽的是同名歧义时两轮可能落到+ // **不同的实体**上,前后两个答案讲的不是同一个节点。+ //+ // 贴在历史之后、当前问题之前——位置就是服从性,跟抽取里 known_block+ // 紧挨正文是同一条理由。+ if !known_entities.is_empty() {+ let lines: Vec<String> = known_entities+ .iter()+ .take(KNOWN_ENTITY_LIMIT)+ .map(|e| {+ format!(+ "{} | {} | {}",+ e["id"].as_str().unwrap_or("?"),+ e["name"].as_str().unwrap_or("?"),+ e["type"].as_str().unwrap_or("?")+ )+ })+ .collect();+ msgs.push(json!({+ "role": "user",+ "content": format!(+ "Entities already identified earlier in this conversation (id | name | type). Call entity_facts with these ids directly; do not look them up by name again:+ {}",+ lines.join("+ ")+ )+ }));+ }// 会话 id 先行下发(新会话由此告知前端)yield Ok(Event::default()@@ -313,6 +349,9 @@ pub async fn chat(// 落库累积:assistant 全文与行动轨迹(历史回放用)let mut answer_acc = String::new();let mut steps_acc: Vec<serde_json::Value> = Vec::new();+ // 这一轮认下的实体,落库供下一轮回放。**只记身份不记正文**——+ // 工具结果里是 chunk 全文,每轮重复堆进上下文几轮就把窗口吃光+ let mut resolved_acc: Vec<serde_json::Value> = Vec::new();let mut rounds = 0usize;loop {@@ -335,6 +374,7 @@ pub async fn chat(&state.pool, conversation_id, "assistant", &answer_acc,&serde_json::Value::Array(steps_acc.clone()),&serde_json::Value::Array(sources.clone()),+ &serde_json::Value::Array(resolved_acc.clone()),).await;yield done_event();}@@ -381,6 +421,7 @@ pub async fn chat(&state.pool, conversation_id, "assistant", &answer_acc,&serde_json::Value::Array(steps_acc.clone()),&serde_json::Value::Array(legacy_sources.clone()),+ &serde_json::Value::Array(resolved_acc.clone()),).await;yield done_event();}@@ -422,6 +463,7 @@ pub async fn chat(&state.pool, conversation_id, "assistant", &answer_acc,&serde_json::Value::Array(steps_acc.clone()),&serde_json::Value::Array(sources.clone()),+ &serde_json::Value::Array(resolved_acc.clone()),).await;yield done_event();}@@ -520,6 +562,11 @@ pub async fn chat(.collect::<Vec<_>>().join("\n")};+ for n in &hits {+ resolved_acc.push(json!({+ "id": n.id.to_string(), "name": n.name, "type": n.type_label+ }));+ }(text, json!({ "kind": "entity", "label": name, "detail": format!("{} matches", hits.len()) }))}"entity_facts" => {diff --git a/crates/utopia-store/src/conversations.rs b/crates/utopia-store/src/conversations.rsindex 31dfd933..6168129e 100644--- a/crates/utopia-store/src/conversations.rs+++ b/crates/utopia-store/src/conversations.rs@@ -67,13 +67,19 @@ pub async fn messages(pool: &PgPool, conversation_id: Uuid) -> AppResult<Vec<Con}/// 服务端拼上下文用:最近 n 条的 (role, content),按时间正序。+/// 回放最近几轮:角色、正文,**以及那几轮认下的实体**。+///+/// 只回放角色与正文时,模型看不见自己上一轮搜过什么、拿到过哪些 id,只能从+/// 名字重搜一遍——实测就是这个样子。**但整段工具结果也不该回放**:那里面是+/// chunk 正文,每轮重复堆进上下文,几轮就把窗口吃光。回放身份足矣:有了 id,+/// 下一轮直接调 entity_facts。pub async fn recent_context(pool: &PgPool,conversation_id: Uuid,n: i64,-) -> AppResult<Vec<(String, String)>> {- let mut rows: Vec<(String, String, DateTime<Utc>)> = sqlx::query_as(- "SELECT role, content, created_at FROM conversation_messages+) -> AppResult<(Vec<(String, String)>, Vec<serde_json::Value>)> {+ let mut rows: Vec<(String, String, serde_json::Value, DateTime<Utc>)> = sqlx::query_as(+ "SELECT role, content, resolved, created_at FROM conversation_messagesWHERE conversation_id = $1 ORDER BY created_at DESC LIMIT $2",).bind(conversation_id)@@ -81,7 +87,22 @@ pub async fn recent_context(.fetch_all(pool).await?;rows.reverse();- Ok(rows.into_iter().map(|(r, c, _)| (r, c)).collect())+ // 实体按 id 去重、保持首次出现的顺序:同一个实体在几轮里反复出现是常态,+ // 每轮各列一遍只是把同一件事说三遍+ let mut seen: std::collections::HashSet<String> = std::collections::HashSet::new();+ let mut entities: Vec<serde_json::Value> = Vec::new();+ for (_, _, res, _) in &rows {+ for e in res.as_array().into_iter().flatten() {+ let Some(id) = e["id"].as_str() else { continue };+ if seen.insert(id.to_string()) {+ entities.push(e.clone());+ }+ }+ }+ Ok((+ rows.into_iter().map(|(r, c, _, _)| (r, c)).collect(),+ entities,+ ))}pub async fn append_message(@@ -91,12 +112,15 @@ pub async fn append_message(content: &str,steps: &serde_json::Value,sources: &serde_json::Value,+ // 这一轮认下的实体(id / 名字 / 类型)。下一轮回放,让模型接着走而不是重搜+ resolved: &serde_json::Value,) -> AppResult<Uuid> {let id = Uuid::now_v7();let mut tx = pool.begin().await?;sqlx::query(- "INSERT INTO conversation_messages (id, conversation_id, role, content, steps, sources)- VALUES ($1, $2, $3, $4, $5, $6)",+ "INSERT INTO conversation_messages+ (id, conversation_id, role, content, steps, sources, resolved)+ VALUES ($1, $2, $3, $4, $5, $6, $7)",).bind(id).bind(conversation_id)@@ -104,6 +128,7 @@ pub async fn append_message(.bind(content).bind(steps).bind(sources)+ .bind(resolved).execute(&mut *tx).await?;sqlx::query("UPDATE conversations SET updated_at = now() WHERE id = $1")diff --git a/crates/utopia-store/src/resolution.rs b/crates/utopia-store/src/resolution.rsindex 6af9456d..b84ea697 100644--- a/crates/utopia-store/src/resolution.rs+++ b/crates/utopia-store/src/resolution.rs@@ -334,13 +334,20 @@ const CONTAIN_SCAN_LIMIT: i64 = 16;/// 包含关系绝不能走它:`华瑞集团技术中心` 与 `星云科技技术中心` 都含「技术中心」,/// 同一篇文档里上下文相似度很容易过线,而它们是两个部门。宁分勿合。///-/// **已知不覆盖**:中间插词(`启明 X7 加速卡` 与 `启明 X7 推理加速卡` 谁也不含谁)。-/// 那要三元组相似度,而 `CREATE EXTENSION pg_trgm` 需要超级权限——本仓库刚改成-/// 受限角色连库(迁移 0031),装扩展会在部署上失败。留给需要时再说。+/// **别名一并参与**:合并会把名字搬进 `aliases`,只查 `canonical_name` 的话,+/// 每成功合并一次就少一条召回的桥——`Holmes` 并入 `Sherlock Holmes` 之后,+/// 后来的 `Mr. Holmes` 就再也搭不上了(它跟 `Sherlock Holmes` 谁也不含谁)。+/// 合并越成功、漏得越多,是个会自我加剧的洞。///-/// **性能**:反向那半(新名字包含旧名字)用不上任何索引,所以这里靠-/// `kb_id + type_id` 收窄行集并设上限,且只在**新建实体时**跑一次,不是每条 mention。-/// 大库上如果不够,正解是建一张「后缀键」表走等值查,而不是加模糊索引。+/// **已知不覆盖**:两个名字既不互相包含、又没有共同别名做桥的情形+///(`启明 X7 加速卡` 与 `启明 X7 推理加速卡`)。那要三元组相似度,而+/// `CREATE EXTENSION pg_trgm` 需要超级权限——本仓库是受限角色连库(迁移 0031),+/// 装扩展会在部署上失败。留给需要时再说。+///+/// **性能**:反向那半(新名字包含旧名字)用不上任何索引,别名那半同样,+/// 所以这里靠 `kb_id` 收窄行集并设上限,且只在**新建实体时**跑一次,+/// 不是每条 mention。大库上如果不够,正解是建一张「后缀键」表走等值查,+/// 而不是加模糊索引。async fn containment_reviews(pool: &PgPool,kb_id: Uuid,@@ -368,9 +375,25 @@ async fn containment_reviews(WHERE e.kb_id = $1 AND e.merged_into IS NULLAND e.id <> $2AND lower(e.canonical_name) <> $3- AND char_length(e.canonical_name) >= $4- AND (lower(e.canonical_name) LIKE '%' || $3 || '%'- OR $3 LIKE '%' || lower(e.canonical_name) || '%')+ AND (+ (char_length(e.canonical_name) >= $4+ AND (lower(e.canonical_name) LIKE '%' || $3 || '%'+ OR $3 LIKE '%' || lower(e.canonical_name) || '%'))+ -- **别名也要参与召回,否则每合并一次就少一条桥。**+ --+ -- 合并把名字搬进 aliases:Holmes 并入 Sherlock Holmes 之后,+ -- Holmes 那一行 merged_into 非空、被上面第一个条件滤掉了。可十分钟后+ -- 出现的 Mr. Holmes 跟 Sherlock Holmes 谁也不含谁——本来正是靠+ -- Holmes 才桥得上。实测就是这么漏的:同类型那个 bug 修好、Holmes+ -- 正确合并之后,Mr. Holmes 反而永远进不了队列。+ --+ -- 合并越成功,桥拆得越多。这个洞会自我加剧。+ OR EXISTS (+ SELECT 1 FROM unnest(e.aliases) AS alias+ WHERE char_length(alias) >= $4+ AND (lower(alias) LIKE '%' || $3 || '%'+ OR $3 LIKE '%' || lower(alias) || '%'))+ )ORDER BY char_length(e.canonical_name)LIMIT $5",)diff --git a/migrations/0043_conversation_resolved_entities.sql b/migrations/0043_conversation_resolved_entities.sqlnew file mode 100644index 00000000..813cf6dd--- /dev/null+++ b/migrations/0043_conversation_resolved_entities.sql@@ -0,0 +1,16 @@+-- 这一轮认下了哪些实体:id、名字、类型。下一轮回放它。+--+-- 会话历史此前只回放 role + content,工具调用与结果一条都不进。于是第二轮+-- 开始时模型只看得见上一轮的最终答案文字,不知道自己搜过什么、拿到过哪些 id,+-- 只能从名字重搜一遍。实测就是这个样子:问"华生和夏洛克的关系",再问一次+-- 换个名字,两轮各检索一遍。+--+-- **不回放整段工具结果**:那里面是 chunk 正文,每轮重复堆进上下文,几轮就把+-- 窗口吃光。要回放的是身份——有了 id,下一轮直接调 entity_facts,不必再从+-- 名字查起。顺带治掉一个更隐蔽的毛病:同名歧义时两轮可能查到**不同的实体**,+-- 于是前后两个答案讲的不是同一个节点。+--+-- 与抽取里 known_block 同一个套路(前面分块已认下的实体注入后续分块),+-- 理由也一样:让模型接着上一次的结论走,而不是每次从零开始猜。+ALTER TABLE conversation_messages+ ADD COLUMN resolved jsonb NOT NULL DEFAULT '[]'::jsonb;