更新 docker-compose.yml

Cargo.toml

@@ -73,8 +73,9 @@ encoding_rs = "0.8"
 etcd-client = { version = "0.14", optional = true }
 serde_with = "3.16.1"
 sysinfo = "0.29"
-# jemalloc allocator
+# jemalloc is only enabled on non-Windows targets. Windows uses the system
-# Note: profiling is only needed for debug/analysis, disabled in release for better performance
+# allocator because tikv-jemalloc-sys does not reliably build for MSVC.
+[target.'cfg(not(windows))'.dependencies]
 tikv-jemallocator = { version = "0.6" }
 tikv-jemalloc-ctl = { version = "0.6", optional = true, features = ["use_std"] }
 

README.md

@@ -87,7 +87,11 @@ docker run -d --name mineru-api --restart unless-stopped --ipc host -p 10001:100
 | `HTKNOW_EMBEDDING_URL` | `http://222.190.139.186:59700/v1/embeddings` | 文本向量服务 |
 | `HTKNOW_IMAGE_EMBEDDING_URL` | `http://192.168.0.46:59802/v1/embeddings/file` | 图片向量服务 |
 | `HTKNOW_RERANK_URL` | `http://222.190.139.186:59600/v1/rerank` | Rerank 服务 |
-
+| `HTKNOW_CUSTOM_IMAGE_URL` | `http://192.168.0.46:59085/split_result` | 自定义图片文本化 解析 |
+| `HTKNOW_CUSTOM_IMAGE_URL` | `http://192.168.0.46:59085/split_result` | 自定义图片文本化 解析 |
+| `HTKNOW_WEKNORA_URL` | `http://172.19.152.51:8081` | weknora 后端服务 |
+| `HTKNOW_WEKNORA_API_KEY` | `sk-6U04CeFoY7JSEBKBp7fQSGYoW3kv7e5qdB14aFnEPt` | weknora api接入服务 |
+| `HTKNOW_WEKNORA_KB_MAP` | `{"*":"881ee2e9-633a-426e-b26a-b9175eec9a64"}` | weknora 知识库id和htknow映射接入服务 |
 ### AI 模型
 | 环境变量 | 默认值 | 说明 |
 | --- | --- | --- |

docker-compose.yml

@@ -1,61 +1,45 @@
-version: "3.8"
-
 services:
   htknow:
-    build:
+    image: gwknow:20260729-weknora-fix
-      context: .
+    container_name: gwhtknow_ceshi
-      dockerfile: Dockerfile
-    container_name: htknow
     ports:
-      - "8080:3000"
+      - "12010:3000"
     volumes:
-      # Mount data directory for persistent storage
       - ./data:/app/data
     environment:
-      - RUST_LOG=info
+      RUST_LOG: "info"
-      # 服务器配置
+      HTKNOW_SERVER_HOST: "0.0.0.0"
-      - HTKNOW_SERVER_HOST=0.0.0.0
+      HTKNOW_SERVER_PORT: "3000"
-      - HTKNOW_SERVER_PORT=3000
-      - HTKNOW_SERVER_UPLOAD_LIMIT_MB=500
-      - HTKNOW_SERVER_PROCESS_INTERVAL_SECS=10
-      - HTKNOW_BUILD_KNOWLEDGE_GRAPH=false
-      - "HTKNOW_LANCEDB_COMPACT_CRON=0 0 3 * * *"
-      # 外部服务地址
-      - HTKNOW_MINERU_URL=http://192.168.0.46:10001/file_parse
-      - HTKNOW_REQUEST_TIMEOUT_SECS=600
-      - HTKNOW_MINERU_MAX_PAGES=0
-      - HTKNOW_OFFICE_CONVERT_URL=http://192.168.0.46:8003/convert
-      - HTKNOW_EMBEDDING_URL=http://192.168.0.46:9700/v1/embeddings
-      - HTKNOW_RERANK_URL=http://192.168.0.46:9600/v1/rerank
-      # AI 模型配置
-      - HTKNOW_EMBEDDING_MODEL=bge-m3
-      - HTKNOW_EMBEDDING_DIM=1024
-      - HTKNOW_EMBEDDING_BATCH_SIZE=8
-      - HTKNOW_RERANK_MODEL=bge-rerank
-      - HTKNOW_RERANK_THRESHOLD=0.1
-      # 数据库配置
-      - DATABASE_URL=sqlite://data/app.sqlite
-      - HTKNOW_DB_MAX_CONNECTIONS=10
-      - HTKNOW_DB_BUSY_TIMEOUT_MS=5000
-      # 存储路径配置
-      - HTKNOW_DATA_DIR=data
-      - HTKNOW_LANCEDB_PATH=data/lancedb_data
-      - HTKNOW_TEMP_PATH=data/temp
-      - HTKNOW_IMAGES_PATH=data/images
-      # 搜索配置
-      - HTKNOW_SEARCH_LIMIT=10
-      - HTKNOW_TANTIVY_INDEX_PATH=data/tantivy_index
-      - HTKNOW_TANTIVY_MEMORY_MB=50
-      # LLM 配置（可选，用于知识图谱构建）
-      # - LLM_API_URL=https://api.openai.com/v1/chat/completions
-      # - LLM_API_KEY=your-api-key
-      # - LLM_MODEL=gpt-3.5-turbo
-    restart: unless-stopped
-    # Uncomment if you need to connect to other services
-    # networks:
-    #   - htknow-network
 
-# Uncomment if you need a network
+      HTKNOW_MINERU_URL: "http://192.168.0.46:10001/file_parse"
-# networks:
+      HTKNOW_REQUEST_TIMEOUT_SECS: "600"
-#   htknow-network:
+      HTKNOW_MINERU_MAX_PAGES: "0"
-#     driver: bridge
+      HTKNOW_OFFICE_CONVERT_URL: "http://192.168.2.12:8003/convert"
+      HTKNOW_EMBEDDING_URL: "http://192.168.2.12:59700/v1/embeddings"
+      HTKNOW_RERANK_URL: "http://192.168.2.12:59600/v1/rerank"
+      HTKNOW_CUSTOM_PARSE_URL: "http://192.168.2.12:59085/split_result"
+      HTKNOW_IMAGE_PARSE_URL: "http://192.168.2.12:59085/split_image"
+
+      HTKNOW_WEKNORA_URL: "http://192.168.2.12:59098"
+      HTKNOW_WEKNORA_API_KEY: "sk-R7YwleKiYxVlSyp_BPE6Giz5t8rCOQpRrcDBBZ6Bljc"
+      HTKNOW_WEKNORA_KB_MAP: '{"*":"f9241739-537d-4678-8e42-99ce65f45b25"}'
+
+      HTKNOW_EMBEDDING_MODEL: "bge-m3"
+      HTKNOW_EMBEDDING_DIM: "1024"
+      HTKNOW_EMBEDDING_BATCH_SIZE: "8"
+      HTKNOW_RERANK_MODEL: "bge-rerank"
+      HTKNOW_RERANK_THRESHOLD: "0.1"
+
+      DATABASE_URL: "sqlite://data/app.sqlite"
+      HTKNOW_DB_MAX_CONNECTIONS: "10"
+      HTKNOW_DB_BUSY_TIMEOUT_MS: "5000"
+
+      HTKNOW_DATA_DIR: "data"
+      HTKNOW_LANCEDB_PATH: "data/lancedb_data"
+      HTKNOW_TEMP_PATH: "data/temp"
+      HTKNOW_IMAGES_PATH: "data/images"
+      HTKNOW_SEARCH_LIMIT: "10"
+      HTKNOW_TANTIVY_INDEX_PATH: "data/tantivy_index"
+      HTKNOW_TANTIVY_MEMORY_MB: "50"
+
+    restart: unless-stopped

src/api/common.rs

@@ -1,9 +1,11 @@
 use sqlx::{QueryBuilder, Sqlite, SqlitePool};
 
 use crate::{
-    AuthUser, api::{
+    AuthUser,
-        error::{ApiError, ApiResult}, knowledge_base
+    api::{
-    }
+        error::{ApiError, ApiResult},
+        knowledge_base,
+    },
 };
 
 /// 要求当前用户为 admin，否则返回 BadRequest。

src/api/error.rs

@@ -1,7 +1,9 @@
 use std::{fmt, num::ParseIntError, result};
 
 use axum::{
-    Json, http::StatusCode, response::{IntoResponse, Response}
+    Json,
+    http::StatusCode,
+    response::{IntoResponse, Response},
 };
 use serde::Serialize;
 use sqlx;

src/api/file.rs

@@ -456,7 +456,7 @@ pub async fn upload(
 
                 match field_name.as_deref() {
                     Some("file") => {
-                        let filename = field.file_name().unwrap_or("unknown").to_string();
+                        let filename = normalize_upload_filename(field.file_name().unwrap_or("unknown"));
                         debug!("Uploading file: {}", filename);
                         let tempname = uuid::Uuid::new_v4().to_string();
                         let filepath = format!("{}/{}", dir, tempname);
@@ -755,6 +755,54 @@ pub async fn upload(
     Ok(Json(uploaded_files))
 }
 
+fn normalize_upload_filename(raw: &str) -> String {
+    if raw.chars().any(is_cjk) {
+        return raw.to_string();
+    }
+    let Some(bytes) = raw.chars().map(|ch| u8::try_from(ch as u32).ok()).collect::<Option<Vec<_>>>() else {
+        return raw.to_string();
+    };
+    if let Ok(decoded) = std::str::from_utf8(&bytes)
+        && decoded.chars().any(is_cjk)
+    {
+        return decoded.to_string();
+    }
+    let high_bytes = bytes.iter().filter(|byte| **byte >= 0x80).count();
+    if high_bytes >= 4 {
+        let (decoded, _, had_errors) = encoding_rs::GBK.decode(&bytes);
+        if !had_errors && decoded.chars().any(is_cjk) {
+            return decoded.into_owned();
+        }
+    }
+    raw.to_string()
+}
+
+fn is_cjk(ch: char) -> bool {
+    matches!(ch as u32, 0x3400..=0x4DBF | 0x4E00..=0x9FFF | 0xF900..=0xFAFF)
+}
+
+#[cfg(test)]
+mod upload_filename_tests {
+    use super::normalize_upload_filename;
+
+    #[test]
+    fn repairs_utf8_bytes_misread_as_latin1() {
+        assert_eq!(normalize_upload_filename("å\u{8f}\u{91}å\u{8a}¨æ\u{9c}º.pdf"), "发动机.pdf");
+    }
+
+    #[test]
+    fn repairs_gbk_bytes_misread_as_latin1() {
+        assert_eq!(normalize_upload_filename("À×´ï-άÐÞÊÖ²á.pdf"), "雷达-维修手册.pdf");
+    }
+
+    #[test]
+    fn preserves_valid_names() {
+        assert_eq!(normalize_upload_filename("发动机.pdf"), "发动机.pdf");
+        assert_eq!(normalize_upload_filename("resume.pdf"), "resume.pdf");
+        assert_eq!(normalize_upload_filename("résumé.pdf"), "résumé.pdf");
+    }
+}
+
 /// 获取文件详情
 #[utoipa::path(
     get,
@@ -1676,6 +1724,16 @@ async fn execute_batch_delete(
         step_start.elapsed().as_millis()
     );
 
+    // Local database and index deletion has committed. Preserve the captured
+    // KB IDs long enough to propagate idempotent delete events to WeKnora.
+    for file in &files_to_delete {
+        if let Some(kb_id) = file.kb_id {
+            if let Err(err) = crate::weknora_sync::enqueue_file_delete(pool, file.id, kb_id).await {
+                warn!("Failed to enqueue WeKnora delete for file {}: {}", file.id, err);
+            }
+        }
+    }
+
     debug!(
         "file_batch_delete total requested={} accepted={} deleted={} {}ms",
         requested,
@@ -2472,6 +2530,13 @@ pub async fn update_slices(
         warn!("Failed to update full search index for file {}: {}", id, e);
     }
 
+    // The slice contents and htknow indexes are committed at this point.
+    // Notify WeKnora as well so an edited document updates its external
+    // knowledge record without requiring the source file to be reparsed.
+    if let Err(err) = crate::weknora_sync::enqueue_file_upsert(&pool, id).await {
+        warn!("Failed to enqueue WeKnora sync for file {}: {}", id, err);
+    }
+
     // 8. 返回更新后的切片列表
     get_slices(State(pool), Path(id)).await
 }

src/api/graph.rs

@@ -1,7 +1,9 @@
 use std::collections::HashMap;
 
 use axum::{
-    Extension, extract::{Path, Query, State}, response::Json
+    Extension,
+    extract::{Path, Query, State},
+    response::Json,
 };
 use serde::{Deserialize, Serialize};
 use sqlx::SqlitePool;

src/api/system.rs

@@ -5,9 +5,12 @@ use sqlx::SqlitePool;
 use utoipa::ToSchema;
 
 use crate::{
-    AuthUser, api::{
+    AuthUser,
-        common, error::{ApiError, ApiResult}
+    api::{
-    }, search::{SearchEngine, tantivy_engine::ForceMergeStats}
+        common,
+        error::{ApiError, ApiResult},
+    },
+    search::{SearchEngine, tantivy_engine::ForceMergeStats},
 };
 
 /// 进程内存占用
@@ -350,14 +353,17 @@ fn force_merge_index_stats(index: &str, stats: ForceMergeStats) -> TantivyForceM
     }
 }
 
-#[cfg(feature = "profiling")]
+#[cfg(all(feature = "profiling", not(windows)))]
 async fn heap_profile_impl() -> ApiResult<Response> {
     use std::{
-        ffi::CString, io::Seek, time::{SystemTime, UNIX_EPOCH}
+        ffi::CString,
+        io::Seek,
+        time::{SystemTime, UNIX_EPOCH},
     };
 
     use axum::{
-        body::Body, http::{HeaderValue, header}
+        body::Body,
+        http::{HeaderValue, header},
     };
     use tempfile::NamedTempFile;
     use tikv_jemalloc_ctl::raw;
@@ -407,19 +413,22 @@ async fn heap_profile_impl() -> ApiResult<Response> {
     Ok(response)
 }
 
-#[cfg(not(feature = "profiling"))]
+#[cfg(any(not(feature = "profiling"), windows))]
 async fn heap_profile_impl() -> ApiResult<Response> {
     Err(ApiError::BadRequest("Heap profiling is only available with 'profiling' feature enabled.".to_string()))
 }
 
-#[cfg(feature = "profiling")]
+#[cfg(all(feature = "profiling", not(windows)))]
 async fn heap_profile_pdf_impl() -> ApiResult<Response> {
     use std::{
-        ffi::CString, process::Stdio, time::{SystemTime, UNIX_EPOCH}
+        ffi::CString,
+        process::Stdio,
+        time::{SystemTime, UNIX_EPOCH},
     };
 
     use axum::{
-        body::Body, http::{HeaderValue, header}
+        body::Body,
+        http::{HeaderValue, header},
     };
     use tempfile::NamedTempFile;
     use tikv_jemalloc_ctl::raw;
@@ -512,7 +521,7 @@ async fn heap_profile_pdf_impl() -> ApiResult<Response> {
     Ok(response)
 }
 
-#[cfg(not(feature = "profiling"))]
+#[cfg(any(not(feature = "profiling"), windows))]
 async fn heap_profile_pdf_impl() -> ApiResult<Response> {
     Err(ApiError::BadRequest("Heap profiling is only available with 'profiling' feature enabled.".to_string()))
 }

src/archive.rs

@@ -4,7 +4,8 @@
 //! 7Z 和 RAR 格式暂不支持。
 
 use std::{
-    io::{Read, Write}, path::Path
+    io::{Read, Write},
+    path::Path,
 };
 
 use encoding_rs::GB18030;

src/config.rs

@@ -380,7 +380,7 @@ fn env_or_parse<T: std::str::FromStr>(key: &str, default: T) -> T {
 
 /// 配置加载器 trait
 #[allow(dead_code)]
-pub trait ConfigLoader: Send+Sync {
+pub trait ConfigLoader: Send + Sync {
     /// 加载配置
     fn load(&self) -> anyhow::Result<AppConfig>;
 }
@@ -413,7 +413,8 @@ impl EtcdConfigLoader {
     /// 启动 watch 监听配置变化
     pub async fn watch<F>(&self, _callback: F) -> anyhow::Result<()>
     where
-        F: Fn(AppConfig)+Send+Sync+'static, {
+        F: Fn(AppConfig) + Send + Sync + 'static,
+    {
         // TODO: 实现 etcd watch 逻辑
         // 1. 连接 etcd
         // 2. 监听 prefix 下的 key 变化

src/db.rs

@@ -131,6 +131,7 @@ async fn run_schema_migrations(pool: &SqlitePool) -> anyhow::Result<()> {
         run_image_description_migration(pool).await?;
         run_file_summary_migration(pool).await?;
         run_image_description_jpg_filename_migration(pool).await?;
+        run_weknora_outbox_migration(pool).await?;
         return Ok(());
     }
 
@@ -161,6 +162,46 @@ async fn run_schema_migrations(pool: &SqlitePool) -> anyhow::Result<()> {
     run_image_description_migration(pool).await?;
     run_file_summary_migration(pool).await?;
     run_image_description_jpg_filename_migration(pool).await?;
+    run_weknora_outbox_migration(pool).await?;
+    Ok(())
+}
+
+async fn run_weknora_outbox_migration(pool: &SqlitePool) -> anyhow::Result<()> {
+    const VERSION: i64 = 6;
+    let mut tx = pool.begin().await?;
+    let claimed = sqlx::query("INSERT OR IGNORE INTO schema_migrations(version, name) VALUES (?, ?)")
+        .bind(VERSION)
+        .bind("weknora_sync_outbox")
+        .execute(&mut *tx)
+        .await?
+        .rows_affected();
+    if claimed == 0 {
+        tx.rollback().await?;
+        return Ok(());
+    }
+    sqlx::query(
+        "CREATE TABLE IF NOT EXISTS weknora_sync_outbox (
+            event_id TEXT NOT NULL,
+            event_type TEXT NOT NULL CHECK(event_type IN ('upsert','delete')),
+            file_id INTEGER PRIMARY KEY,
+            htknow_kb_id INTEGER NOT NULL,
+            attempts INTEGER NOT NULL DEFAULT 0,
+            next_attempt_at INTEGER NOT NULL DEFAULT (strftime('%s','now')),
+            last_error TEXT NOT NULL DEFAULT '',
+            created_at INTEGER NOT NULL DEFAULT (strftime('%s','now')),
+            updated_at INTEGER NOT NULL DEFAULT (strftime('%s','now'))
+        )",
+    )
+    .execute(&mut *tx)
+    .await?;
+    sqlx::query(
+        "CREATE INDEX IF NOT EXISTS idx_weknora_sync_outbox_due
+         ON weknora_sync_outbox(next_attempt_at, created_at)",
+    )
+    .execute(&mut *tx)
+    .await?;
+    tx.commit().await?;
+    info!("Applied schema migration {}: weknora_sync_outbox", VERSION);
     Ok(())
 }
 

src/frontend.rs

@@ -1,5 +1,9 @@
 use axum::{
-    Router, body::Body, http::{StatusCode, Uri, header}, response::{IntoResponse, Response}, routing::get
+    Router,
+    body::Body,
+    http::{StatusCode, Uri, header},
+    response::{IntoResponse, Response},
+    routing::get,
 };
 use rust_embed::Embed;
 

src/image_description.rs

@@ -23,7 +23,8 @@ pub struct ImageDescription {
 
 /// 保存或更新单条图片描述
 pub async fn save(
-    pool: &SqlitePool, file_id: i64, image_filename: &str, description: &str, raw_response: &str, source: &str, jpg_filename: Option<&str>,
+    pool: &SqlitePool, file_id: i64, image_filename: &str, description: &str, raw_response: &str, source: &str,
+    jpg_filename: Option<&str>,
 ) -> Result<()> {
     sqlx::query(
         "INSERT INTO image_descriptions (file_id, image_filename, description, raw_response, source, jpg_filename)
@@ -68,7 +69,9 @@ pub async fn delete(pool: &SqlitePool, file_id: i64, image_filename: &str) -> Re
         .bind(image_filename)
         .execute(pool)
         .await
-        .with_context(|| format!("failed to delete image_description for file_id={} filename={}", file_id, image_filename))?;
+        .with_context(|| {
+            format!("failed to delete image_description for file_id={} filename={}", file_id, image_filename)
+        })?;
     Ok(())
 }
 
@@ -94,13 +97,16 @@ struct FileImageDescriptionRow {
 }
 
 /// 批量查询多个文件的图片描述，返回 (file_id, image_filename) -> description 映射。
-pub async fn list_by_file_ids(pool: &SqlitePool, file_ids: &[i64]) -> Result<HashMap<(i64, String), (String, Option<String>)>> {
+pub async fn list_by_file_ids(
+    pool: &SqlitePool, file_ids: &[i64],
+) -> Result<HashMap<(i64, String), (String, Option<String>)>> {
     if file_ids.is_empty() {
         return Ok(HashMap::new());
     }
 
-    let mut query_builder: QueryBuilder<'_, Sqlite> =
+    let mut query_builder: QueryBuilder<'_, Sqlite> = QueryBuilder::new(
-        QueryBuilder::new("SELECT file_id, image_filename, description, jpg_filename FROM image_descriptions WHERE file_id IN (");
+        "SELECT file_id, image_filename, description, jpg_filename FROM image_descriptions WHERE file_id IN (",
+    );
     let mut separated = query_builder.separated(", ");
     for file_id in file_ids {
         separated.push_bind(file_id);
@@ -123,7 +129,16 @@ pub async fn copy_for_file(pool: &SqlitePool, src_file_id: i64, dst_file_id: i64
     .with_context(|| format!("failed to copy image_descriptions from file_id={}", src_file_id))?;
 
     for row in rows {
-        save(pool, dst_file_id, &row.image_filename, &row.description, &row.raw_response, &row.source, row.jpg_filename.as_deref()).await?;
+        save(
+            pool,
+            dst_file_id,
+            &row.image_filename,
+            &row.description,
+            &row.raw_response,
+            &row.source,
+            row.jpg_filename.as_deref(),
+        )
+        .await?;
     }
     Ok(())
 }

src/lib.rs

@@ -1,5 +1,8 @@
 use axum::{
-    body::Body, http::{Request, StatusCode}, middleware::Next, response::{IntoResponse, Response}
+    body::Body,
+    http::{Request, StatusCode},
+    middleware::Next,
+    response::{IntoResponse, Response},
 };
 use base64::{Engine, engine::general_purpose::STANDARD};
 
@@ -19,6 +22,7 @@ pub mod pdf_highlight;
 pub mod processor;
 pub mod search;
 pub mod slice_content;
+pub mod weknora_sync;
 
 /// User authentication info extracted from request headers
 #[derive(Clone, Debug)]

src/log4rs.rs

@@ -1,7 +1,9 @@
 use std::{env, str::FromStr};
 
 use log4rs::{
-    append::console::ConsoleAppender, config::{Appender, Config, Logger, Root}, encode::pattern::PatternEncoder
+    append::console::ConsoleAppender,
+    config::{Appender, Config, Logger, Root},
+    encode::pattern::PatternEncoder,
 };
 
 fn parse_rust_log(rust_log: &str) -> (Option<log::LevelFilter>, Vec<(String, log::LevelFilter)>) {

src/main.rs

@@ -1,8 +1,8 @@
-#[cfg(debug_assertions)]
+#[cfg(all(not(windows), debug_assertions))]
 #[global_allocator]
 static GLOBAL: tikv_jemallocator::Jemalloc = tikv_jemallocator::Jemalloc;
 
-#[cfg(all(debug_assertions, feature = "profiling"))]
+#[cfg(all(not(windows), debug_assertions, feature = "profiling"))]
 #[unsafe(export_name = "_rjem_malloc_conf")]
 // lg_prof_sample:0 表示每次分配都采样(最详细,但性能开销大)
 // lg_prof_sample:10 表示每 1KB 采样一次(推荐)
@@ -13,7 +13,7 @@ use std::{net::SocketAddr, sync::Arc};
 
 use axum::{Router, extract::DefaultBodyLimit, middleware, response::Html, routing::get};
 use chrono::Local;
-use htknow::{api, auth, config, db, frontend, log4rs, processor, search};
+use htknow::{api, auth, config, db, frontend, log4rs, processor, search, weknora_sync};
 use tokio::net::TcpListener;
 use tokio_cron_scheduler::{Job, JobScheduler};
 use utoipa_swagger_ui::SwaggerUi;
@@ -27,6 +27,7 @@ async fn main() -> anyhow::Result<()> {
     log::info!("Configuration loaded: server={}:{}", cfg.server.host, cfg.server.port);
 
     let pool = db::init().await?;
+    weknora_sync::start_worker(pool.clone());
     log::info!("Initializing search engine...");
     let search_init_started_at = std::time::Instant::now();
     let search_engine = search::SearchEngine::init().await.with_pool(pool.clone());

src/pdf_highlight.rs

@@ -5,7 +5,9 @@
 use anyhow::Result;
 use log::info;
 use lopdf::{
-    Document, Object, ObjectId, Stream, content::{Content, Operation}, dictionary
+    Document, Object, ObjectId, Stream,
+    content::{Content, Operation},
+    dictionary,
 };
 use serde::{Deserialize, Serialize};
 

src/processor.rs

@@ -179,8 +179,16 @@ async fn load_or_parse_image_descriptions(
         }
         match image_parse::parse_image_file(path, name, None).await {
             Ok(resp) => {
-                if let Err(err) =
+                if let Err(err) = image_description::save(
-                    image_description::save(pool, file_id, name, &resp.description, &resp.raw_response, source, resp.jpg_filename.as_deref()).await
+                    pool,
+                    file_id,
+                    name,
+                    &resp.description,
+                    &resp.raw_response,
+                    source,
+                    resp.jpg_filename.as_deref(),
+                )
+                .await
                 {
                     warn!("Failed to save image description for {}: {}", name, err);
                 }
@@ -2046,12 +2054,7 @@ impl FileProcessor {
         if output.status.success() {
             return Ok(());
         }
-        anyhow::bail!(
+        anyhow::bail!("{} exited with {}: {}", program, output.status, String::from_utf8_lossy(&output.stderr).trim());
-            "{} exited with {}: {}",
-            program,
-            output.status,
-            String::from_utf8_lossy(&output.stderr).trim()
-        );
     }
 
     async fn process_uploaded_image_file(
@@ -2165,7 +2168,10 @@ impl FileProcessor {
             if is_image {
                 let removed = retain_one_image_block_for_uploaded_image(&mut content_list);
                 if removed > 0 {
-                    warn!("Keeping one MinerU image block for uploaded image {}, dropped {} duplicate blocks", file.filename, removed);
+                    warn!(
+                        "Keeping one MinerU image block for uploaded image {}, dropped {} duplicate blocks",
+                        file.filename, removed
+                    );
                 }
             }
         } else {
@@ -2179,7 +2185,10 @@ impl FileProcessor {
             if is_image {
                 let removed = retain_one_image_block_for_uploaded_image(&mut content_list);
                 if removed > 0 {
-                    warn!("Keeping one MinerU image block for uploaded image {}, dropped {} duplicate blocks", file.filename, removed);
+                    warn!(
+                        "Keeping one MinerU image block for uploaded image {}, dropped {} duplicate blocks",
+                        file.filename, removed
+                    );
                 }
             }
             if !self.ensure_file_exists(file.id, "before writing pdf contents").await? {
@@ -2295,11 +2304,7 @@ impl FileProcessor {
                     text_level: None,
                     text_format: None,
                     img_path: Some(file.filename.clone()),
-                    image_caption: if caption.is_empty() {
+                    image_caption: if caption.is_empty() { None } else { Some(vec![caption]) },
-                        None
-                    } else {
-                        Some(vec![caption])
-                    },
                     table_body: None,
                     table_caption: None,
                 });
@@ -2947,6 +2952,10 @@ impl FileProcessor {
         })
         .await?;
 
+        if let Err(err) = crate::weknora_sync::enqueue_file_upsert(&self.pool, file.id).await {
+            warn!("Failed to enqueue WeKnora sync for file {}: {}", file.id, err);
+        }
+
         Ok(())
     }
 
@@ -3566,6 +3575,9 @@ impl FileProcessor {
             updated_file.content = Some(indexed_content);
             updated_file.summary = summary.clone();
             self.maybe_build_knowledge_graph(&updated_file).await;
+            if let Err(err) = crate::weknora_sync::enqueue_file_upsert(&self.pool, target.id).await {
+                warn!("Failed to enqueue WeKnora sync for file {}: {}", target.id, err);
+            }
             return Ok(());
         }
 
@@ -3640,6 +3652,9 @@ impl FileProcessor {
         updated_file.content = Some(full_content.clone());
         updated_file.summary = source.summary.clone();
         self.maybe_build_knowledge_graph(&updated_file).await;
+        if let Err(err) = crate::weknora_sync::enqueue_file_upsert(&self.pool, target.id).await {
+            warn!("Failed to enqueue WeKnora sync for file {}: {}", target.id, err);
+        }
 
         Ok(())
     }

src/search/mod.rs

@@ -29,8 +29,48 @@ pub use tantivy_engine::{FullSearchResultItem, SearchResultItem};
 const FULL_SNIPPET_MAX_CHARS: usize = 400;
 const MAX_QUERY_TERMS_FOR_SYNONYM_LOOKUP: usize = 100;
 const DEFAULT_REBUILD_BATCH_SIZE: i64 = 100;
+/// Reciprocal Rank Fusion constant. 60 is the commonly used default.
+const RRF_K: f32 = 60.0;
 static RERANK_HTTP_CLIENT: Lazy<Client> = Lazy::new(Client::new);
 
+/// Fuse ranked recall lists without comparing incompatible raw score scales.
+fn reciprocal_rank_fusion(result_lists: impl IntoIterator<Item = Vec<SearchResultItem>>) -> Vec<SearchResultItem> {
+    let mut fused: HashMap<i64, (SearchResultItem, f32, usize)> = HashMap::new();
+
+    for results in result_lists {
+        for (index, mut result) in results.into_iter().filter(|result| !result.content.trim().is_empty()).enumerate() {
+            let rank = index + 1;
+            let contribution = 1.0 / (RRF_K + rank as f32);
+            fused
+                .entry(result.id)
+                .and_modify(|(_, score, best_rank)| {
+                    *score += contribution;
+                    *best_rank = (*best_rank).min(rank);
+                })
+                .or_insert_with(|| {
+                    result.score = contribution;
+                    (result, contribution, rank)
+                });
+        }
+    }
+
+    let mut results: Vec<(SearchResultItem, usize)> = fused
+        .into_values()
+        .map(|(mut result, score, best_rank)| {
+            result.score = score;
+            (result, best_rank)
+        })
+        .collect();
+    results.sort_by(|(a, a_rank), (b, b_rank)| {
+        b.score
+            .partial_cmp(&a.score)
+            .unwrap_or(std::cmp::Ordering::Equal)
+            .then_with(|| a_rank.cmp(b_rank))
+            .then_with(|| a.id.cmp(&b.id))
+    });
+    results.into_iter().map(|(result, _)| result).collect()
+}
+
 /// /v1/rerank 格式的请求体
 #[derive(Debug, Serialize)]
 struct RerankRequest {
@@ -1207,37 +1247,8 @@ impl SearchEngine {
             }
         };
 
-        // 合并结果：使用 HashMap 按 id 去重，保留最高分数，同时去除内容为空的结果
+        // 两个召回分支的原始分数量纲不同，按排名进行 RRF 融合。
-        let mut merged_map: HashMap<i64, SearchResultItem> = HashMap::new();
+        let merged_results = reciprocal_rank_fusion([tantivy_results, lancedb_results]);
-
-        for result in tantivy_results {
-            // 跳过空内容（包括仅有空白的情况）
-            if result.content.trim().is_empty() {
-                continue;
-            }
-            merged_map.insert(result.id, result);
-        }
-
-        for result in lancedb_results {
-            // 跳过空内容（包括仅有空白的情况）
-            if result.content.trim().is_empty() {
-                continue;
-            }
-
-            merged_map
-                .entry(result.id)
-                .and_modify(|e| {
-                    // 如果已存在，取两者中分数较高的
-                    if result.score > e.score {
-                        *e = result.clone();
-                    }
-                })
-                .or_insert(result);
-        }
-
-        // 转换为 Vec 并按分数降序排序
-        let mut merged_results: Vec<SearchResultItem> = merged_map.into_values().collect();
-        merged_results.sort_by(|a, b| b.score.partial_cmp(&a.score).unwrap_or(std::cmp::Ordering::Equal));
 
         info!("Merged results count: {}", merged_results.len());
 
@@ -1366,29 +1377,7 @@ impl SearchEngine {
             }
         };
 
-        let mut merged_map: HashMap<i64, SearchResultItem> = HashMap::new();
+        let merged_results = reciprocal_rank_fusion([tantivy_results, lancedb_results]);
-        for result in tantivy_results {
-            if result.content.trim().is_empty() {
-                continue;
-            }
-            merged_map.insert(result.id, result);
-        }
-        for result in lancedb_results {
-            if result.content.trim().is_empty() {
-                continue;
-            }
-            merged_map
-                .entry(result.id)
-                .and_modify(|e| {
-                    if result.score > e.score {
-                        *e = result.clone();
-                    }
-                })
-                .or_insert(result);
-        }
-
-        let mut merged_results: Vec<SearchResultItem> = merged_map.into_values().collect();
-        merged_results.sort_by(|a, b| b.score.partial_cmp(&a.score).unwrap_or(std::cmp::Ordering::Equal));
         info!("Image-by-text merged results count: {}", merged_results.len());
 
         if merged_results.is_empty() {
@@ -1967,3 +1956,33 @@ fn reload_reader(reader: &IndexReader, label: &str) -> tantivy::Result<()> {
     debug!("Tantivy {} reader reload {}ms", label, start.elapsed().as_millis());
     Ok(())
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn result(id: i64, content: &str, score: f32) -> SearchResultItem {
+        SearchResultItem { id, file_id: id, kb_id: Some(1), content: content.to_string(), score }
+    }
+
+    #[test]
+    fn rrf_rewards_results_recalled_by_both_branches() {
+        let text = vec![result(1, "text-first", 100.0), result(2, "shared", 1.0)];
+        let vector = vec![result(2, "shared", 0.1), result(3, "vector-second", 0.09)];
+
+        let fused = reciprocal_rank_fusion([text, vector]);
+
+        assert_eq!(fused.iter().map(|item| item.id).collect::<Vec<_>>(), vec![2, 1, 3]);
+        assert!((fused[0].score - (1.0 / 62.0 + 1.0 / 61.0)).abs() < f32::EPSILON);
+    }
+
+    #[test]
+    fn rrf_uses_rank_instead_of_raw_score_and_drops_empty_content() {
+        let text = vec![result(10, "first", 0.01), result(11, "second", 10_000.0), result(12, " ", 20_000.0)];
+
+        let fused = reciprocal_rank_fusion([text, Vec::new()]);
+
+        assert_eq!(fused.iter().map(|item| item.id).collect::<Vec<_>>(), vec![10, 11]);
+        assert!(fused[0].score > fused[1].score);
+    }
+}

src/search/tantivy_engine.rs

@@ -1,14 +1,25 @@
 use std::{
-    collections::{HashMap, HashSet}, path::Path, sync::{
+    collections::{HashMap, HashSet},
-        Arc, mpsc::{self, Sender}
+    path::Path,
-    }, thread, time::{Duration, Instant, SystemTime, UNIX_EPOCH}
+    sync::{
+        Arc,
+        mpsc::{self, Sender},
+    },
+    thread,
+    time::{Duration, Instant, SystemTime, UNIX_EPOCH},
 };
 
 use aho_corasick::{AhoCorasick, AhoCorasickBuilder, MatchKind};
 use log::{debug, info, warn};
 use serde::Serialize;
 use tantivy::{
-    Index, IndexReader, Result, TantivyDocument, TantivyError, Term, collector::TopDocs, directory::error::LockError, doc, merge_policy::LogMergePolicy, query::{BooleanQuery, BoostQuery, Occur, Query, TermQuery}, schema::{FAST, Field, INDEXED, IndexRecordOption, STORED, Schema, TextFieldIndexing, TextOptions, Value as _}
+    Index, IndexReader, Result, TantivyDocument, TantivyError, Term,
+    collector::TopDocs,
+    directory::error::LockError,
+    doc,
+    merge_policy::LogMergePolicy,
+    query::{BooleanQuery, BoostQuery, Occur, Query, TermQuery},
+    schema::{FAST, Field, INDEXED, IndexRecordOption, STORED, Schema, TextFieldIndexing, TextOptions, Value as _},
 };
 
 use super::chinese_tokenizer;
@@ -451,7 +462,7 @@ pub async fn create_rebuild_writer(index: &Index, label: &str) -> tantivy::Resul
 }
 
 pub fn add_documents(
-    index_writer: &mut tantivy::IndexWriter, schema: &Schema, docs: impl IntoIterator<Item=Document>,
+    index_writer: &mut tantivy::IndexWriter, schema: &Schema, docs: impl IntoIterator<Item = Document>,
 ) -> tantivy::Result<usize> {
     let mut count = 0_usize;
     for doc in docs {

src/weknora_sync.rs

@@ -0,0 +1,265 @@
+use std::{collections::HashMap, time::Duration};
+
+use anyhow::{Context, Result, anyhow};
+use log::{info, warn};
+use once_cell::sync::Lazy;
+use reqwest::Client;
+use serde::Serialize;
+use sha2::{Digest, Sha256};
+use sqlx::{FromRow, SqlitePool};
+
+use crate::api::{FILE_COLS_NO_CONTENT, File};
+
+static CLIENT: Lazy<Client> =
+    Lazy::new(|| Client::builder().timeout(Duration::from_secs(120)).build().expect("build WeKnora sync HTTP client"));
+
+#[derive(Debug, Serialize)]
+struct ExternalChunk {
+    external_slice_id: String,
+    index: usize,
+    content: String,
+}
+
+#[derive(Debug, Serialize)]
+struct ExternalImport {
+    source: &'static str,
+    external_file_id: String,
+    source_version: String,
+    title: String,
+    filename: String,
+    file_type: String,
+    file_size: i64,
+    file_hash: String,
+    chunks: Vec<ExternalChunk>,
+}
+
+#[derive(Debug, FromRow)]
+struct SyncEvent {
+    event_id: String,
+    event_type: String,
+    file_id: i64,
+    htknow_kb_id: i64,
+    attempts: i64,
+}
+
+pub async fn enqueue_file_upsert(pool: &SqlitePool, file_id: i64) -> Result<()> {
+    if !sync_enabled() {
+        return Ok(());
+    }
+    let kb_id: i64 = sqlx::query_scalar("SELECT kb_id FROM files WHERE id = ?")
+        .bind(file_id)
+        .fetch_optional(pool)
+        .await?
+        .flatten()
+        .ok_or_else(|| anyhow!("file {} has no knowledge base", file_id))?;
+    enqueue(pool, "upsert", file_id, kb_id).await
+}
+
+pub async fn enqueue_file_delete(pool: &SqlitePool, file_id: i64, htknow_kb_id: i64) -> Result<()> {
+    if !sync_enabled() {
+        return Ok(());
+    }
+    enqueue(pool, "delete", file_id, htknow_kb_id).await
+}
+
+async fn enqueue(pool: &SqlitePool, event_type: &str, file_id: i64, htknow_kb_id: i64) -> Result<()> {
+    let event_id = uuid::Uuid::new_v4().to_string();
+    sqlx::query(
+        "INSERT INTO weknora_sync_outbox
+         (event_id, event_type, file_id, htknow_kb_id, attempts, next_attempt_at, last_error, created_at, updated_at)
+         VALUES (?, ?, ?, ?, 0, strftime('%s','now'), '', strftime('%s','now'), strftime('%s','now'))
+         ON CONFLICT(file_id) DO UPDATE SET
+           event_id = excluded.event_id,
+           event_type = excluded.event_type,
+           htknow_kb_id = excluded.htknow_kb_id,
+           attempts = 0,
+           next_attempt_at = excluded.next_attempt_at,
+           last_error = '',
+           updated_at = excluded.updated_at",
+    )
+    .bind(event_id)
+    .bind(event_type)
+    .bind(file_id)
+    .bind(htknow_kb_id)
+    .execute(pool)
+    .await?;
+    Ok(())
+}
+
+pub fn start_worker(pool: SqlitePool) {
+    if !sync_enabled() {
+        info!("WeKnora synchronization is disabled");
+        return;
+    }
+    tokio::spawn(async move {
+        loop {
+            if let Err(err) = process_due_events(&pool).await {
+                warn!("WeKnora outbox worker failed: {}", err);
+            }
+            tokio::time::sleep(Duration::from_secs(2)).await;
+        }
+    });
+}
+
+async fn process_due_events(pool: &SqlitePool) -> Result<()> {
+    let events: Vec<SyncEvent> = sqlx::query_as(
+        "SELECT event_id, event_type, file_id, htknow_kb_id, attempts
+         FROM weknora_sync_outbox
+         WHERE next_attempt_at <= strftime('%s','now')
+         ORDER BY created_at, file_id
+         LIMIT 10",
+    )
+    .fetch_all(pool)
+    .await?;
+    for event in events {
+        let result = match event.event_type.as_str() {
+            "upsert" => sync_file(pool, event.file_id).await,
+            "delete" => delete_file(event.file_id, event.htknow_kb_id).await,
+            other => Err(anyhow!("unsupported outbox event type {}", other)),
+        };
+        match result {
+            Ok(()) => {
+                sqlx::query("DELETE FROM weknora_sync_outbox WHERE file_id = ? AND event_id = ?")
+                    .bind(event.file_id)
+                    .bind(&event.event_id)
+                    .execute(pool)
+                    .await?;
+                info!("WeKnora {} accepted for file {}", event.event_type, event.file_id);
+            }
+            Err(err) => {
+                let attempts = event.attempts + 1;
+                let delay = (2_i64.saturating_pow(attempts.min(10) as u32)).min(3600);
+                sqlx::query(
+                    "UPDATE weknora_sync_outbox
+                     SET attempts = ?, next_attempt_at = strftime('%s','now') + ?, last_error = ?, updated_at = strftime('%s','now')
+                     WHERE file_id = ? AND event_id = ?",
+                )
+                .bind(attempts)
+                .bind(delay)
+                .bind(err.to_string())
+                .bind(event.file_id)
+                .bind(&event.event_id)
+                .execute(pool)
+                .await?;
+                warn!(
+                    "WeKnora {} attempt {} failed for file {}; retry in {}s: {}",
+                    event.event_type, attempts, event.file_id, delay, err
+                );
+            }
+        }
+    }
+    Ok(())
+}
+
+fn sync_enabled() -> bool {
+    std::env::var("HTKNOW_WEKNORA_URL").ok().is_some_and(|v| !v.trim().is_empty())
+}
+
+fn mapped_weknora_kb_id(htknow_kb_id: i64) -> Result<String> {
+    let kb_map_raw = std::env::var("HTKNOW_WEKNORA_KB_MAP").context("HTKNOW_WEKNORA_KB_MAP is required")?;
+    let kb_map: HashMap<String, String> =
+        serde_json::from_str(&kb_map_raw).context("parse HTKNOW_WEKNORA_KB_MAP JSON")?;
+    kb_map
+        .get(&htknow_kb_id.to_string())
+        .or_else(|| kb_map.get("*"))
+        .cloned()
+        .ok_or_else(|| anyhow!("no WeKnora KB mapping for htknow KB {} and no '*' default mapping", htknow_kb_id))
+}
+
+async fn sync_file(pool: &SqlitePool, file_id: i64) -> Result<()> {
+    let file: File = sqlx::query_as(&format!("SELECT {} FROM files WHERE id = ?", FILE_COLS_NO_CONTENT))
+        .bind(file_id)
+        .fetch_one(pool)
+        .await
+        .context("load parsed file")?;
+    if file.status != 1 {
+        return Err(anyhow!("file status is {}, expected completed", file.status));
+    }
+    let htknow_kb_id = file.kb_id.ok_or_else(|| anyhow!("file has no knowledge base"))?;
+    let weknora_kb_id = mapped_weknora_kb_id(htknow_kb_id)?;
+
+    let source_id = crate::api::effective_parse_file_id(pool, file_id).await?;
+    let ids: Vec<i64> = sqlx::query_scalar("SELECT id FROM slices WHERE file_id = ? ORDER BY ordinal, id")
+        .bind(source_id)
+        .fetch_all(pool)
+        .await?;
+    let contents = crate::slice_content::read_all(source_id).await?;
+    let chunks = ids
+        .into_iter()
+        .enumerate()
+        .filter_map(|(index, id)| {
+            let content = contents.get(&id)?.trim();
+            if content.is_empty() {
+                return None;
+            }
+            Some(ExternalChunk { external_slice_id: id.to_string(), index, content: content.to_string() })
+        })
+        .collect::<Vec<_>>();
+    if chunks.is_empty() {
+        return Err(anyhow!("parsed file contains no text chunks"));
+    }
+
+    let mut version_hasher = Sha256::new();
+    version_hasher.update(file.filename.as_bytes());
+    version_hasher.update(file.hash.as_bytes());
+    for chunk in &chunks {
+        version_hasher.update(chunk.external_slice_id.as_bytes());
+        version_hasher.update([0]);
+        version_hasher.update(chunk.content.as_bytes());
+        version_hasher.update([0xff]);
+    }
+    let source_version = hex::encode(version_hasher.finalize());
+    let file_type =
+        std::path::Path::new(&file.filename).extension().and_then(|v| v.to_str()).unwrap_or("markdown").to_lowercase();
+    let payload = ExternalImport {
+        source: "htknow",
+        external_file_id: file.id.to_string(),
+        source_version,
+        title: file.filename.clone(),
+        filename: file.filename,
+        file_type,
+        file_size: file.size,
+        file_hash: file.hash,
+        chunks,
+    };
+    let base = std::env::var("HTKNOW_WEKNORA_URL")?;
+    let url =
+        format!("{}/api/v1/knowledge-bases/{}/knowledge/external-chunks", base.trim_end_matches('/'), weknora_kb_id);
+    let mut request = CLIENT.post(url).json(&payload);
+    if let Ok(key) = std::env::var("HTKNOW_WEKNORA_API_KEY")
+        && !key.trim().is_empty()
+    {
+        request = request.header("X-API-Key", key);
+    }
+    let response = request.send().await.context("call WeKnora external chunk import")?;
+    let status = response.status();
+    if !status.is_success() {
+        let body = response.text().await.unwrap_or_default();
+        return Err(anyhow!("WeKnora returned {}: {}", status, body));
+    }
+    Ok(())
+}
+
+async fn delete_file(file_id: i64, htknow_kb_id: i64) -> Result<()> {
+    let weknora_kb_id = mapped_weknora_kb_id(htknow_kb_id)?;
+    let base = std::env::var("HTKNOW_WEKNORA_URL")?;
+    let url = format!(
+        "{}/api/v1/knowledge-bases/{}/knowledge/external-chunks/htknow/{}",
+        base.trim_end_matches('/'),
+        weknora_kb_id,
+        file_id
+    );
+    let mut request = CLIENT.delete(url);
+    if let Ok(key) = std::env::var("HTKNOW_WEKNORA_API_KEY")
+        && !key.trim().is_empty()
+    {
+        request = request.header("X-API-Key", key);
+    }
+    let response = request.send().await.context("call WeKnora external knowledge delete")?;
+    let status = response.status();
+    if !status.is_success() {
+        let body = response.text().await.unwrap_or_default();
+        return Err(anyhow!("WeKnora returned {}: {}", status, body));
+    }
+    Ok(())
+}

tests/api_concurrency.rs

@@ -1,6 +1,7 @@
 mod common;
 use std::{
-    fs, time::{Duration, Instant}
+    fs,
+    time::{Duration, Instant},
 };
 
 use axum::http::StatusCode;

tests/common/mod.rs

@@ -1,14 +1,19 @@
 #![allow(dead_code, unused_imports)]
 
 use std::{
-    fs, path::PathBuf, sync::{
+    fs,
-        OnceLock, atomic::{AtomicUsize, Ordering}
+    path::PathBuf,
-    }, time::{SystemTime, UNIX_EPOCH}
+    sync::{
+        OnceLock,
+        atomic::{AtomicUsize, Ordering},
+    },
+    time::{SystemTime, UNIX_EPOCH},
 };
 
 use axum::{Router, extract::DefaultBodyLimit, middleware};
 pub use axum::{
-    body::Body, http::{Request, Response, StatusCode, header}
+    body::Body,
+    http::{Request, Response, StatusCode, header},
 };
 use htknow::{api, auth, db, search::SearchEngine, slice_content};
 pub use http_body_util::BodyExt;