feat: 知识批量沉淀改为后台任务,避免 nginx 504

- /ingest/threads/batch 新增 background 参数(默认 true),立即返回 {queued:true}
- 抽取循环抽出 _run_batch_ingest,后台路径显式传入 created_by
- 前端知识库页与输入框、线程 hooks/types 配套调整
- config.yaml 上下文压缩测试值

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

frontend-web/src/core/knowledge/types.ts

@@ -147,6 +147,13 @@ export interface BatchIngestPayload {
   include_sources?: boolean;
   skip_existing?: boolean;
   mode?: "summary" | "full";
+  /**
+   * When true (default), the backend runs the per-thread extraction loop in a
+   * background task and returns immediately (`queued:true`) instead of blocking
+   * the request until every thread is processed — which otherwise outlives the
+   * nginx gateway timeout and surfaces as a generic "request failed".
+   */
+  background?: boolean;
   /** Optional model override for knowledge extraction (null = configured default). */
   model_name?: string | null;
   /** Optional extraction-prompt template id (null = scope default). */
@@ -157,6 +164,8 @@ export interface BatchIngestPayload {
 
 export interface BatchIngestResult {
   ok: boolean;
+  /** True when the backend queued the batch as a background task (see payload). */
+  queued?: boolean;
   processed: number;
   created: number;
   skipped: number;

frontend-web/src/core/settings/local.ts

@@ -7,6 +7,7 @@ export const DEFAULT_LOCAL_SETTINGS: LocalSettings = {
   display: {
     streamSpeedMeterEnabled: true,
     ragModeEnabled: false,
+    contextCompressionEnabled: false,
     citationCleanupWords: [],
     fontScale: 1,
   },
@@ -43,6 +44,12 @@ export interface LocalSettings {
      * messages render tool results with the legacy generic search UI only.
      */
     ragModeEnabled: boolean;
+    /**
+     * Per-user switch for context compression (上下文压缩). When enabled, long
+     * conversations are summarized transiently before the model call to save
+     * tokens. Default off; toggled from the chat input box (below 参考文献).
+     */
+    contextCompressionEnabled: boolean;
     /**
      * Admin-maintained words or short phrases removed from reference snippets
      * and dialogs, one entry per line in the settings UI.

frontend-web/src/core/threads/hooks.ts

@@ -664,6 +664,9 @@ export function useThreadStream({
           lockThreadRagMode(threadId);
         }
 
+        const contextCompressionEnabled =
+          localSettings.display.contextCompressionEnabled ?? false;
+
         const mergedContext = {
           ...context,
           ...extraContext,
@@ -715,6 +718,7 @@ export function useThreadStream({
               memory_injection_enabled:
                 mergedContext.memory_injection_enabled ?? true,
               rag_mode_enabled: ragModeEnabled,
+              summarization_enabled: contextCompressionEnabled,
               thread_id: threadId,
             },
           },

frontend-web/src/core/threads/types.ts

@@ -20,6 +20,8 @@ export interface AgentThreadContext extends Record<string, unknown> {
   writing_artifact_path?: string;
   memory_injection_enabled?: boolean;
   rag_mode_enabled?: boolean;
+  /** Per-request switch for context compression (上下文压缩). Default off. */
+  summarization_enabled?: boolean;
   /** Per-conversation switch for knowledge-base RAG injection (phase 2). */
   knowledge_rag_enabled?: boolean;
   agent_id?: string;