Feature/20260623 ai reply speed up 1

package.json

@@ -1,7 +1,7 @@
 {
   "name": "knowforge",
   "private": true,
-  "version": "0.7.1",
+  "version": "0.7.2",
   "license": "Apache-2.0",
   "repository": {
     "type": "git",

src-tauri/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "knowforge"
-version = "0.7.1"
+version = "0.7.2"
 description = "Knowforge desktop app"
 authors = ["caichangqing"]
 edition = "2024"

src-tauri/src/challenge_review.rs

@@ -7,6 +7,8 @@ use std::fs;
 use std::path::{Path, PathBuf};
 use uuid::Uuid;
 
+use std::sync::Arc;
+
 use crate::llm::{create_provider, CompletionOverrides};
 use crate::llm::LlmChatMessage;
 use crate::thought_parser;
@@ -272,6 +274,7 @@ fn normalize_template_kind(raw: Option<&str>) -> String {
 #[tauri::command]
 pub async fn generate_challenge_question(
     workspace: tauri::State<'_, crate::WorkspaceState>,
+    http_client: tauri::State<'_, Arc<reqwest::Client>>,
     args: GenerateChallengeQuestionArgs,
 ) -> Result<GenerateChallengeQuestionResponse, String> {
     let root = crate::lock_workspace_root(&workspace)?;
@@ -282,7 +285,7 @@ pub async fn generate_challenge_question(
     .await
     .map_err(|e| e.to_string())??;
 
-    let provider = match create_provider(&ai, None) {
+    let provider = match create_provider(&ai, None, http_client.inner()) {
         Ok(p) => p,
         Err(_) => {
             return Ok(GenerateChallengeQuestionResponse {
@@ -407,6 +410,7 @@ pub async fn generate_challenge_question(
 #[tauri::command]
 pub async fn evaluate_challenge_answer(
     workspace: tauri::State<'_, crate::WorkspaceState>,
+    http_client: tauri::State<'_, Arc<reqwest::Client>>,
     args: EvaluateChallengeAnswerArgs,
 ) -> Result<EvaluateChallengeAnswerResponse, String> {
     let root = crate::lock_workspace_root(&workspace)?;
@@ -424,7 +428,7 @@ pub async fn evaluate_challenge_answer(
         }
     };
 
-    let provider = match create_provider(&ai, None) {
+    let provider = match create_provider(&ai, None, http_client.inner()) {
         Ok(p) => p,
         Err(_) => {
             return Ok(EvaluateChallengeAnswerResponse {

src-tauri/src/lib.rs

@@ -1540,11 +1540,13 @@ async fn semantic_search(
     args: semantic_index::SemanticSearchArgs,
     app: tauri::AppHandle,
     state: tauri::State<'_, WorkspaceState>,
+    embed_cache: tauri::State<'_, Arc<semantic_index::EmbeddingCache>>,
 ) -> Result<Vec<semantic_index::SemanticSearchHit>, String> {
     let root = lock_workspace_root(&state)?;
     let cache = semantic_index::default_model_cache_dir();
     let bundle = semantic_index::resolve_bundle_model_dir(&app);
-    tauri::async_runtime::spawn_blocking(move || semantic_index::run_semantic_search(&root, &cache, &bundle, args))
+    let embed_cache_arc = Arc::clone(embed_cache.inner());
+    tauri::async_runtime::spawn_blocking(move || semantic_index::run_semantic_search(&root, &cache, &bundle, args, &embed_cache_arc))
         .await
         .map_err(|e| e.to_string())?
 }
@@ -1557,6 +1559,8 @@ async fn suggest_related_notes(
     include_reasons: Option<bool>,
     editor_markdown_override: Option<String>,
     state: tauri::State<'_, WorkspaceState>,
+    http_client: tauri::State<'_, Arc<reqwest::Client>>,
+    embed_cache: tauri::State<'_, Arc<semantic_index::EmbeddingCache>>,
 ) -> Result<Vec<link_recommendation::LinkRecommendation>, String> {
     let canonical_root = lock_workspace_root(&state)?;
     let rel_path = rel_path.trim().to_string();
@@ -1569,6 +1573,7 @@ async fn suggest_related_notes(
     let root = canonical_root.clone();
     let rel = rel_path.clone();
     let md_override = editor_markdown_override.clone();
+    let embed_cache_arc = Arc::clone(embed_cache.inner());
     let mut out = tauri::async_runtime::spawn_blocking(move || -> Result<Vec<link_recommendation::LinkRecommendation>, String> {
         let emb = semantic_index::open_embedding_db(&root)?;
         let thoughts = vault_thoughts_db::open_thoughts_db(&root)?;
@@ -1579,6 +1584,7 @@ async fn suggest_related_notes(
             &thoughts,
             max_results,
             md_override.as_deref(),
+            &embed_cache_arc,
         )
     })
     .await
@@ -1593,7 +1599,7 @@ async fn suggest_related_notes(
         })
         .await
         .map_err(|e| e.to_string())??;
-        link_recommendation::enrich_recommendations_with_reasons(&mut out, &excerpt, &ai).await?;
+        link_recommendation::enrich_recommendations_with_reasons(&mut out, &excerpt, &ai, http_client.inner()).await?;
     }
 
     Ok(out)
@@ -1655,6 +1661,8 @@ async fn add_manual_topic_semantic(
 pub fn run() {
     tauri::Builder::default()
         .manage(WorkspaceState::default())
+        .manage(Arc::new(llm::build_shared_http_client()))
+        .manage(Arc::new(semantic_index::EmbeddingCache::new()))
         .manage(Arc::new(llm::LlmSessionState::default()))
         .manage(Arc::new(llm::approval::ToolApprovalState::new()))
         .manage(Arc::new(tools::ToolRegistry::new()))

src-tauri/src/link_recommendation.rs

@@ -1,6 +1,8 @@
 //! 文档级语义相似度 → 双向链接推荐候选（迭代 6.3 步骤 15）。
 //! 无语义索引时不提供推荐（无关键词兜底）。
 
+use std::sync::Arc;
+
 use crate::llm::create_provider;
 use crate::llm::LlmChatMessage;
 use crate::note_privacy;
@@ -435,6 +437,7 @@ pub fn suggest_related_notes(
     _thoughts_db: &Connection,
     max_results: usize,
     editor_markdown_override: Option<&str>,
+    embed_cache: &semantic_index::EmbeddingCache,
 ) -> Result<Vec<LinkRecommendation>, String> {
     if max_results == 0 {
         return Ok(Vec::new());
@@ -456,7 +459,7 @@ pub fn suggest_related_notes(
         return Err("semantic_index_not_ready: kf-private notes are excluded from link recommendations".to_string());
     }
 
-    let all_chunks = semantic_index::load_all_doc_embeddings(embedding_db)?;
+    let all_chunks = embed_cache.get_docs(embedding_db);
     if all_chunks.is_empty() {
         return Err(
             "semantic_index_not_ready: no document chunks in embedding index; rebuild embeddings first"
@@ -623,8 +626,9 @@ async fn link_reason_completion_body(
     ai: &ResolvedAiConfig,
     messages: &[LlmChatMessage],
     timeout_ms: u64,
+    http_client: &Arc<reqwest::Client>,
 ) -> Option<String> {
-    let provider = create_provider(ai, None).ok()?;
+    let provider = create_provider(ai, None, http_client).ok()?;
     let overrides = crate::llm::CompletionOverrides {
         timeout_ms: Some(timeout_ms),
         ..Default::default()
@@ -639,6 +643,7 @@ pub async fn enrich_recommendations_with_reasons(
     candidates: &mut [LinkRecommendation],
     current_doc_excerpt: &str,
     config: &ResolvedAiConfig,
+    http_client: &Arc<reqwest::Client>,
 ) -> Result<(), String> {
     if candidates.is_empty() {
         return Ok(());
@@ -680,7 +685,7 @@ pub async fn enrich_recommendations_with_reasons(
         },
     ];
 
-    let raw = match link_reason_completion_body(config, &msgs, timeout_ms).await {
+    let raw = match link_reason_completion_body(config, &msgs, timeout_ms, http_client).await {
         Some(s) => {
             let t = s.trim();
             if t.is_empty() {
@@ -813,7 +818,8 @@ mod tests {
         semantic_index::upsert_doc_chunk(&conn, "far.md#0", "far.md", 0, "t", &far, "m").unwrap();
 
         let tconn = Connection::open_in_memory().unwrap();
-        let rec = suggest_related_notes(root, "cur.md", &conn, &tconn, 5, None).unwrap();
+        let ec = semantic_index::EmbeddingCache::new();
+        let rec = suggest_related_notes(root, "cur.md", &conn, &tconn, 5, None, &ec).unwrap();
         assert_eq!(rec.len(), 2);
         assert_eq!(rec[0].target_rel_path, "near.md");
         assert_eq!(rec[1].target_rel_path, "mid.md");
@@ -837,7 +843,8 @@ mod tests {
         semantic_index::upsert_doc_chunk(&conn, "linked#0", "linked", 0, "t", &near, "m").unwrap();
 
         let tconn = Connection::open_in_memory().unwrap();
-        let rec = suggest_related_notes(root, "cur.md", &conn, &tconn, 5, None).expect("chunks align");
+        let ec = semantic_index::EmbeddingCache::new();
+        let rec = suggest_related_notes(root, "cur.md", &conn, &tconn, 5, None, &ec).expect("chunks align");
         assert_eq!(rec.len(), 1);
         assert_eq!(rec[0].target_rel_path, "near.md");
     }
@@ -862,13 +869,15 @@ mod tests {
         semantic_index::upsert_doc_chunk(&conn, "far.md#0", "far.md", 0, "t", &far, "m").unwrap();
 
         let tconn = Connection::open_in_memory().unwrap();
+        let ec = semantic_index::EmbeddingCache::new();
         let rec = suggest_related_notes(
             root,
             "cur.md",
             &conn,
             &tconn,
             5,
             Some("x [[linked]]\n"),
+            &ec,
         )
         .unwrap();
         assert_eq!(rec.len(), 1);
@@ -931,7 +940,8 @@ mod tests {
             existing_link: false,
             reason: None,
         }];
-        super::enrich_recommendations_with_reasons(&mut c, "excerpt", &ai)
+        let http_client = Arc::new(reqwest::Client::new());
+        super::enrich_recommendations_with_reasons(&mut c, "excerpt", &ai, &http_client)
             .await
             .unwrap();
         assert!(c[0].reason.is_none());

src-tauri/src/llm/agent_loop.rs

@@ -11,11 +11,11 @@ use std::time::Duration;
 
 use futures_util::future::join_all;
 use serde_json::{json, Value};
-use tauri::{AppHandle, Emitter};
+use tauri::{AppHandle, Emitter, Manager};
 use tokio_util::sync::CancellationToken;
 
 use super::approval::ToolApprovalState;
-use super::context_guard::ContextGuard;
+use super::context_guard::{ContextGuard, PrecomputedSummary};
 use super::memory;
 use super::provider::{LlmProvider, NormalizedToolCall};
 use super::{LlmChatMessage, LlmToolCall, LlmToolCallFunction};
@@ -137,6 +137,7 @@ pub async fn run_agent_stream(
         ContextGuard::with_provider(config.max_context_tokens, provider.clone())
     };
     let mut loop_detector = LoopDetector::new();
+    let mut pending_summary: Option<tokio::task::JoinHandle<Option<PrecomputedSummary>>> = None;
 
     loop {
         if cancel.is_cancelled() {
@@ -154,6 +155,11 @@ pub async fn run_agent_stream(
             }
         }
 
+        if let Some(handle) = pending_summary.take() {
+            if let Ok(Some(cached)) = handle.await {
+                context_guard.apply_cached_summary(&mut messages, &cached);
+            }
+        }
         context_guard.trim_with_summary(&mut messages).await;
 
         // 1. 流式请求（携带 tools 字段；本轮文字会通过 emit_chunk/emit_done 推给前端）
@@ -384,7 +390,14 @@ pub async fn run_agent_stream(
             store_extraction_msgs(&memory_manager, &messages).await;
             return String::new();
         }
-        // 继续 loop：下一轮流式请求会带上完整的 messages 历史
+
+        if context_guard.budget_pressure(&messages) > 0.7 {
+            let msgs_snapshot = messages.clone();
+            let guard_clone = context_guard.clone();
+            pending_summary = Some(tokio::spawn(async move {
+                guard_clone.pre_summarize(&msgs_snapshot).await
+            }));
+        }
     }
 }
 
@@ -484,6 +497,9 @@ pub(crate) async fn execute_tool(
     );
     ctx.call_id = Some(tc.id.clone());
     ctx.provider = provider;
+    if let Some(ec) = app.try_state::<Arc<crate::semantic_index::EmbeddingCache>>() {
+        ctx.embed_cache = Some(Arc::clone(&*ec));
+    }
 
     let manifest = tool.manifest().clone();
     let start = std::time::Instant::now();