agent/advisor_agent/data_query.py

@@ -13,6 +13,7 @@ from config.settings import settings
 from nl2sql.contracts import DataQueryRequest, DataQueryResult
 from nl2sql.embedding import EmbeddingError
 from nl2sql.retrieval import retrieve_metadata
+from nl2sql.cache import cache_get, cache_set, build_question_cache_key
 from nl2sql.runtime_config import runtime_config
 from nl2sql.schema import load_authoritative_schema
 from repositories.customer_relation import CustomerRelationRepo
@@ -154,6 +155,7 @@ async def execute_advisor_data_query(
         pass
 
     if scope == "customer" and _is_current_holdings_query(question):
+        redis = redis or database.redis.client()
         return await _query_current_holdings(
             db,
             customer_id=customer_id,
@@ -166,6 +168,21 @@ async def execute_advisor_data_query(
 
     milvus = milvus or database.milvus.client()
     redis = redis or database.redis.client()
+    cache_key = build_question_cache_key(
+        question,
+        permission=permission,
+        data_scope={"customer_ids": customer_ids},
+        page=page,
+        page_size=page_size,
+        sort_by=sort_by,
+        sort_order=sort_order,
+    )
+    cached = await cache_get(redis, cache_key)
+    if cached is not None:
+        cached["trace_id"] = trace_id
+        cached["customer_id"] = customer_id
+        cached["warnings"] = [*cached.get("warnings", []), "cache_hit"]
+        return cached
     llm_client = llm_client or default_llm
     request = DataQueryRequest(
         question=question,
@@ -222,4 +239,5 @@ async def execute_advisor_data_query(
         payload["answer"] = f"{existing_answer.rstrip('。')}。客户姓名：{enriched['name_summary']}。"
     payload["sql"] = None
     payload["customer_id"] = customer_id
+    await cache_set(redis, cache_key, payload, ttl=runtime_config.cache_ttl)
     return payload

agent/advisor_agent/intent/classifier.py

@@ -108,6 +108,17 @@ async def classify_advisor_intent(
             "rule",
             "识别为系统提示文本而非业务查询",
         )
+    fast_intent = recognize_advisor_intent(query)
+    if (
+        fast_intent == AGENT_INTENT_DATA_QUERY
+        and not any(term in query for term in _RECOMMENDATION_TERMS)
+    ):
+        return IntentClassification(
+            fast_intent,
+            0.95,
+            "rule_fast",
+            "明显查询类问题",
+        )
     if llm_client is not None:
         try:
             raw = await asyncio.wait_for(

api/routers/advisor_agent.py

@@ -331,6 +331,7 @@ async def chat_stream(
                 session_id=chat_request.session_id,
                 conversation_context=conversation_context,
                 llm_client=getattr(_advisor_runtime(request), "llm_client", None),
+                redis=redis_db.client(),
             )
         except QueryServiceError as exc:
             payload = agent_failure(ERR_CODE_LLM_ERROR, _data_query_error_message(exc), trace_id=trace_id)
@@ -565,6 +566,7 @@ async def advisor_data_query(
             page_size=body.page_size,
             sort_by=body.sort_by,
             sort_order=body.sort_order,
+            redis=redis_db.client(),
         )
     except QueryServiceError as exc:
         return agent_failure(

api/routers/nl2sql.py

@@ -248,6 +248,7 @@ async def query_data(
             db,
             database=settings.mysql.database,
             candidate_tables=table_names,
+            redis=redis,
         )
 
     request_contract = DataQueryRequest(

frontend/lib/assistant-api.ts

@@ -72,10 +72,7 @@ export const assistantApi = {
   sessionHistory: (sessionId: string) =>
     apiFetch<AgentMessage[]>(`/advisor-agent/session/${encodeURIComponent(sessionId)}/history`),
 
-  /**
-   * POST /chat/stream —— 响应是 text/event-stream，但后端整包输出，
-   * 因此读完整条响应后一次性返回聚合结果（非流式渲染）。
-   */
+  /** POST /chat/stream —— 读取 SSE，并支持调用方实时消费事件。 */
   chatOnce(
     body: {
       query: string;

main.py

@@ -90,6 +90,7 @@ async def lifespan(app: FastAPI):
         await app.state.advisor_event_consumer.stop()
     if app.state.advisor_scheduler is not None:
         app.state.advisor_scheduler.shutdown()
+    await llm_client.aclose()
     await database.dispose()
 
 

nl2sql/cache.py

@@ -49,6 +49,32 @@ def build_cache_key(
     return f"nl2sql:{digest}"
 
 
+def build_question_cache_key(
+    question: str,
+    *,
+    permission: dict[str, Any],
+    data_scope: dict[str, Any] | None = None,
+    page: int = 1,
+    page_size: int = 100,
+    sort_by: str | None = None,
+    sort_order: str = "asc",
+) -> str:
+    """生成自然语言查询结果缓存键，覆盖问题、权限和数据范围。"""
+    payload = {
+        "question": " ".join((question or "").split()),
+        "permission": _permission_payload(permission),
+        "data_scope": data_scope or {},
+        "page": page,
+        "page_size": page_size,
+        "sort_by": sort_by or "",
+        "sort_order": sort_order,
+    }
+    digest = hashlib.sha256(
+        json.dumps(payload, ensure_ascii=False, sort_keys=True, separators=(",", ":")).encode()
+    ).hexdigest()
+    return f"nl2sql:question:{digest}"
+
+
 async def cache_get(redis, key: str) -> dict[str, Any] | None:
     """读取缓存，Redis 异常或内容损坏时返回空结果。"""
     try:

nl2sql/query_rewriter.py

@@ -16,6 +16,11 @@ _SYSTEM_PROMPT = """你是基金平台 NL2SQL 的问题改写器。
 不能扩大权限范围，不能臆造查询结果；上下文不足时保留原问题。
 """
 
+_REFERENCE_TERMS = (
+    "他", "她", "它", "他们", "她们", "它们", "这个客户", "该客户", "那个客户",
+    "这个基金", "该基金", "那只基金", "这只基金", "这个结果", "上一轮", "刚才", "前面",
+)
+
 
 def _rewrite_explicit_customer_identity(question: str) -> str | None:
     normalized = re.sub(r"\s+", "", question)
@@ -68,7 +73,7 @@ async def rewrite_query(
     contextual_identity = _rewrite_customer_name_follow_up(original, context)
     if contextual_identity:
         return contextual_identity
-    if not context or llm_client is None:
+    if not context or llm_client is None or not any(term in original for term in _REFERENCE_TERMS):
         return original
 
     prompt = (

nl2sql/result.py

@@ -67,6 +67,7 @@ async def render_answer(
                 {"role": "user", "content": prompt},
             ],
             temperature=0,
+            max_tokens=256,
         )
     except Exception:  # noqa: BLE001 摘要失败回退固定文本
         return fallback

nl2sql/retrieval.py

@@ -1,6 +1,7 @@
 """NL2SQL 元数据召回与授权过滤。"""
 from __future__ import annotations
 
+import asyncio
 from collections.abc import Awaitable, Callable, Iterable
 from typing import Any
 
@@ -70,15 +71,21 @@ async def retrieve_metadata(
         "is_valid",
         "is_deprecated",
     ]
-    results: list[dict[str, Any]] = []
-    for chunk_type in ("table_meta", "field_meta"):
-        hits = await milvus_client.search(
+    async def search_chunk(chunk_type: str):
+        return await milvus_client.search(
             collection_name=NL2SQL_COLLECTION,
             data=[vector],
             limit=top_k,
             filter=f'chunk_type == "{chunk_type}" and is_valid == true',
             output_fields=output_fields,
         )
+
+    table_hits, field_hits = await asyncio.gather(
+        search_chunk("table_meta"),
+        search_chunk("field_meta"),
+    )
+    results: list[dict[str, Any]] = []
+    for hits in (table_hits, field_hits):
         for hit in _flatten_hits(hits):
             entity = hit.get("entity") or hit
             if not entity.get("is_valid", True) or entity.get("is_deprecated", False):

nl2sql/schema.py

@@ -1,12 +1,23 @@
 """从 MySQL information_schema 加载并校验 NL2SQL 权威 Schema。"""
 from __future__ import annotations
 
+import hashlib
+import json
+import logging
 from typing import Any
 
 from sqlalchemy import bindparam, text
 
 from nl2sql.metadata import normalize_column_row, normalize_table_row
 
+logger = logging.getLogger("nl2sql.schema")
+SCHEMA_CACHE_TTL = 60
+
+
+def _schema_cache_key(database: str, candidates: list[str]) -> str:
+    payload = json.dumps([database, candidates], ensure_ascii=False, separators=(",", ":"))
+    return "nl2sql:schema:" + hashlib.sha256(payload.encode()).hexdigest()
+
 
 TABLES_SQL = text(
     """
@@ -34,13 +45,25 @@ async def load_authoritative_schema(
     *,
     database: str,
     candidate_tables: set[str] | list[str],
+    redis=None,
+    cache_ttl: int = SCHEMA_CACHE_TTL,
 ) -> dict[str, list[dict[str, Any]]]:
     """从权威元数据源加载候选表，并剔除不存在的表和孤立字段。"""
     candidates = {str(name).strip() for name in candidate_tables if str(name).strip()}
     if not candidates:
         return {"tables": [], "columns": []}
 
-    params = {"database": database, "candidate_tables": sorted(candidates)}
+    candidate_list = sorted(candidates)
+    cache_key = _schema_cache_key(database, candidate_list)
+    if redis is not None:
+        try:
+            cached = await redis.get(cache_key)
+            if cached:
+                return json.loads(cached)
+        except Exception:  # noqa: BLE001 缓存异常回退权威查询
+            logger.warning("Schema 缓存读取失败", exc_info=True)
+
+    params = {"database": database, "candidate_tables": candidate_list}
     table_result = await session.execute(TABLES_SQL, params)
     column_result = await session.execute(COLUMNS_SQL, params)
 
@@ -56,4 +79,10 @@ async def load_authoritative_schema(
         normalized = normalize_column_row(dict(row))
         if normalized["table_name"] in valid_table_names and normalized["field_name"]:
             columns.append(normalized)
-    return {"tables": tables, "columns": columns}
+    schema = {"tables": tables, "columns": columns}
+    if redis is not None:
+        try:
+            await redis.set(cache_key, json.dumps(schema, ensure_ascii=False), ex=cache_ttl)
+        except Exception:  # noqa: BLE001 缓存异常不阻断查询
+            logger.warning("Schema 缓存写入失败", exc_info=True)
+    return schema

nl2sql/sql_agent.py

@@ -74,7 +74,7 @@ async def generate_sql(
         },
     ]
     try:
-        output = await llm_client.chat(messages, temperature=0)
+        output = await llm_client.chat(messages, temperature=0, max_tokens=256)
     except Exception as exc:  # noqa: BLE001 统一收敛模型异常
         raise SqlGenerationError("SQL 模型调用失败") from exc
     sql = _clean_model_output(output)

service/client_agent/runtime.py

@@ -304,7 +304,7 @@ def _build_data_query(*, db_session_factory, milvus_client, llm_client, config_g
 
         async def schema_loader(table_names: set[str], _permission: dict):
             return await load_authoritative_schema(
-                db, database=schema_database, candidate_tables=table_names
+                db, database=schema_database, candidate_tables=table_names, redis=redis
             )
 
         request = DataQueryRequest(

tool/llm.py

@@ -103,6 +103,21 @@ class LLMClient:
     def __init__(self, cfg: LLMCfg | None = None):
         self.cfg = cfg or settings.llm
         self.backends = resolve_backends(self.cfg)
+        self._clients: dict[str, httpx.AsyncClient] = {}
+
+    def _client_for(self, backend: Backend) -> httpx.AsyncClient:
+        clients = getattr(self, "_clients", None)
+        if clients is None:
+            clients = {}
+            self._clients = clients
+        if backend.name not in clients:
+            clients[backend.name] = backend.client(self.cfg.timeout)
+        return clients[backend.name]
+
+    async def aclose(self) -> None:
+        for client in getattr(self, "_clients", {}).values():
+            await client.aclose()
+        getattr(self, "_clients", {}).clear()
 
     # ---- 首选后端（健康检查/日志/embed 使用） -----------------------------
     @property
@@ -186,9 +201,9 @@ class LLMClient:
                 "stream": False,
             }
             try:
-                async with backend.client(self.cfg.timeout) as client:
-                    r = await client.post(url, headers=backend.headers, json=payload)
-                    r.raise_for_status()
+                client = self._client_for(backend)
+                r = await client.post(url, headers=backend.headers, json=payload)
+                r.raise_for_status()
                 choice = r.json()["choices"][0]
                 message = choice["message"]
                 content = message.get("content")
@@ -250,20 +265,16 @@ class LLMClient:
         response = None
         max_retries = max(1, int(getattr(self.cfg, "max_retries", 1)))
         retry_backoff_sec = float(getattr(self.cfg, "retry_backoff_sec", 0))
-        async with backend.client(self.cfg.timeout) as client:
-            for attempt in range(max_retries):
-                try:
-                    response = await client.post(
-                        url,
-                        headers=backend.headers,
-                        json=payload,
-                    )
-                    response.raise_for_status()
-                    break
-                except Exception:
-                    if attempt == max_retries - 1:
-                        raise
-                    await asyncio.sleep(retry_backoff_sec * (2**attempt))
+        client = self._client_for(backend)
+        for attempt in range(max_retries):
+            try:
+                response = await client.post(url, headers=backend.headers, json=payload)
+                response.raise_for_status()
+                break
+            except Exception:
+                if attempt == max_retries - 1:
+                    raise
+                await asyncio.sleep(retry_backoff_sec * (2**attempt))
         if response is None:  # pragma: no cover - defensive guard
             raise LLMFailError("Embedding 请求未返回响应")
         data = response.json()
@@ -286,10 +297,10 @@ class LLMClient:
             url = backend.base_url.removesuffix("/v1") + "/api/tags"
         else:
             url = f"{backend.base_url}/models"
-        async with backend.client(min(self.cfg.timeout, 10)) as client:
-            r = await client.get(url, headers=backend.headers)
-            if r.status_code >= 500:
-                r.raise_for_status()
+        client = self._client_for(backend)
+        r = await client.get(url, headers=backend.headers)
+        if r.status_code >= 500:
+            r.raise_for_status()
 
 
 # 全局单例：Agent 统一引入