--- production/wechat_rpa/admin_api.py
+++ release/wechat_rpa/admin_api.py
@@ -42,6 +42,7 @@
 import admin_backend as backend
 import app_version
 import archive_api
+import knowledge_api
 import zyt_auth
 
 # 账号下拉框里「全部」的取值。用 -1 而不是 0：0 早就有含义了（旧版未归属数据），
@@ -1277,12 +1278,19 @@
     ) -> dict:
         # 这一份的结构由 `admin_backend.desktop_config_payload` 统一给出。
         # 两处各写一份的话迟早漂移，表现成"某台客户端少了个字段"，极难查。
-        return backend.desktop_config_payload(
+        payload = backend.desktop_config_payload(
             database,
             scheme=request.headers.get("x-forwarded-proto", request.url.scheme),
             host=request.headers.get("host", ""),
             desktop_account=account,
         )
+        knowledge = app.state.knowledge_store
+        tenant = str(account["tenant_id"])
+        counts = await asyncio.to_thread(knowledge.stats, tenant)
+        settings = await asyncio.to_thread(knowledge.settings, tenant)
+        payload["knowledge"] = {"published": counts.get("published", 0),
+                                "enabled": bool(settings["enabled"]), "updated_at": settings["updated_at"]}
+        return payload
 
     # 桌面端回流一次编排调用。补上这条之前，客户端改指 8766 会丢掉全部调用留痕
     # ——配置照常同步、回复照常发，唯独调用记录一条不进库，而那恰恰是出事后用来
@@ -1325,6 +1333,10 @@
         resolve_admin_tenant=resolve_admin_tenant,
         zyt_api_url=lambda: desktop_login_api_url()[0],
         zyt_patient_searcher=zyt_patient_searcher,
+    )
+
+    knowledge_api.register_knowledge_routes(
+        app, database, require, client_ip, resolve_admin_tenant, current_desktop,
     )
 
     # ── 前端静态资源 ─────────────────────────────────────────────────────

--- production/wechat_rpa/model_gateway.py
+++ release/wechat_rpa/model_gateway.py
@@ -44,6 +44,8 @@
 
 import model_protocol
 import admin_backend
+from knowledge_store import KnowledgeStore
+from knowledge_retriever import KnowledgeRetriever, inject_references
 
 # ── 超时分层 ─────────────────────────────────────────────────────────────────
 # 整体必须小于桌面端的超时，否则桌面端先超时重发，网关还在跑，双倍扣费。
@@ -477,6 +479,9 @@
     app = FastAPI(title="模型统一网关", version="1.0")
     database = admin_backend.Database(db_path)
     database.migrate()
+    knowledge_store = KnowledgeStore(database)
+    knowledge_store.initialize()
+    knowledge_retriever = KnowledgeRetriever(knowledge_store)
     catalog = Catalog(db_path)
     idempotency: dict[str, tuple[float, dict]] = {}
     log_queue: asyncio.Queue = asyncio.Queue(maxsize=2000)
@@ -522,6 +527,9 @@
             record = await log_queue.get()
             try:
                 await asyncio.to_thread(_write_log, db_path, record)
+                if record.get("knowledge_retrieval") is not None:
+                    await asyncio.to_thread(knowledge_retriever.record, record["tenant_id"],
+                                            record["task_id"], record["knowledge_retrieval"])
             except Exception:
                 pass
             finally:
@@ -584,7 +592,12 @@
         )
         if idempotency_key and idempotency_key in idempotency:
             # 桌面端重发（超时后重试）不该重复扣费
-            return JSONResponse(idempotency[idempotency_key][1])
+            cached = idempotency[idempotency_key][1]
+            cached_hits = (cached.get("knowledge") or {}).get("hits", [])
+            if not cached_hits or await asyncio.to_thread(
+                    knowledge_retriever.references_current, str(account["tenant_id"]), cached_hits):
+                return JSONResponse(cached)
+            idempotency.pop(idempotency_key, None)
 
         body = await request.json()
         answers, judge, mode, version = catalog.plan()
@@ -606,6 +619,32 @@
         tools = body.get("tools") or None
 
         started = time.monotonic()
+        knowledge_result = None
+        knowledge_trace = {"enabled": False, "hits": [], "mode": "disabled", "elapsed_ms": 0}
+        tenant = str(account["tenant_id"])
+        if str(body.get("purpose") or "chat") == "chat":
+            settings = await asyncio.to_thread(knowledge_store.settings, tenant)
+            if settings["enabled"]:
+                from knowledge_processing import latest_query
+                query = str(body.get("customer_text") or "")
+                if not query:
+                    for turn in reversed(messages):
+                        if turn.get("role") == "user" and isinstance(turn.get("content"), str):
+                            query = turn["content"]
+                            break
+                query = latest_query(query) or query
+                try:
+                    knowledge_result = await asyncio.wait_for(
+                        asyncio.to_thread(knowledge_retriever.search, tenant, query), timeout=4.0)
+                    knowledge_trace = {"enabled": True, "mode": knowledge_result["mode"],
+                        "elapsed_ms": knowledge_result["elapsed_ms"], "hits": [
+                            {k: h[k] for k in ("id", "revision", "score")} for h in knowledge_result["hits"]]}
+                    messages = inject_references(messages, knowledge_result["hits"])
+                except Exception:
+                    knowledge_trace = {"enabled": True, "hits": [], "mode": "unavailable", "elapsed_ms": 4000}
+                    from knowledge_processing import redact
+                    knowledge_result = {**knowledge_trace, "query": redact(query)[:1000]}
+                    messages = inject_references(messages, [])
 
         if tools:
             # 工具轮：只问主出口，不并发也不评审。
@@ -634,12 +673,24 @@
                 "judge": dict(_EMPTY_VERDICT),
                 "judge_mode": mode,
                 "roles_version": version,
+                "knowledge": knowledge_trace,
                 "total_ms": int((time.monotonic() - started) * 1000),
             }
             if single.get("error") and not payload["reply"] and not payload["tool_calls"]:
                 return JSONResponse(payload, status_code=502)
-            if x_idempotency_key:
-                idempotency[x_idempotency_key] = (now, payload)
+            if idempotency_key:
+                idempotency[idempotency_key] = (now, payload)
+            try:
+                log_queue.put_nowait({
+                    "desktop_account_id": int(account["id"]), "tenant_id": tenant,
+                    "device_id": x_device_id, "task_id": str(body.get("task_id") or ""),
+                    "roles_version": version, "judge_mode": mode, "chosen": payload["chosen"],
+                    "judge": payload["judge"], "candidates": [single], "total_ms": payload["total_ms"],
+                    "customer_text": str(body.get("customer_text") or ""), "reply_text": payload["reply"],
+                    "purpose": str(body.get("purpose") or "chat"), "knowledge_retrieval": knowledge_result,
+                })
+            except asyncio.QueueFull:
+                pass
             return JSONResponse(payload)
 
         results = await asyncio.gather(
@@ -683,6 +734,7 @@
 
         payload = {
             "reply": chosen["text"],
+            "knowledge": knowledge_trace,
             "tool_calls": [],
             "chosen": chosen["provider"],
             "candidates": results,
@@ -708,6 +760,7 @@
                 "customer_text": str(body.get("customer_text") or ""),
                 "reply_text": payload["reply"],
                 "purpose": str(body.get("purpose") or "chat"),
+                "knowledge_retrieval": knowledge_result,
             })
         except asyncio.QueueFull:
             pass       # 观测数据丢一条，不能因此拖慢回复

--- production/wechat_rpa/rbac.py
+++ release/wechat_rpa/rbac.py
@@ -27,6 +27,11 @@
     ("model:write",   "增删改模型与编排", "模型"),
     ("agent:read",    "查看智能体",       "模型"),
     ("agent:write",   "增删改智能体与协作方案", "模型"),
+    ("knowledge:read", "查看知识中心", "知识库"),
+    ("knowledge:write", "加工和编辑知识草稿", "知识库"),
+    ("knowledge:review", "审核知识", "知识库"),
+    ("knowledge:publish", "发布和停用知识", "知识库"),
+    ("knowledge:export", "导出已发布知识", "知识库"),
     ("review:read",   "查看待审队列",     "审核"),
     ("review:write",  "通过/驳回待审回复", "审核"),
     ("stats:read",    "查看调用统计",     "运营"),