fix(ai): DeepSeek 真实联调——响应健壮性与筛选替换提示词

- assistant 解析兼容：支持 in_progress 异步轮询（DeepSeek 偶发）、
  去除 markdown 代码围栏、整体失败重试一次后才报 502
- 系统提示词增加筛选替换规则与示例：用户明确提到某维度新取值时替换
  现有筛选而非叠加（实测修复“口腔+健康管理”空范围问题）
- KPI 数字字号改为容器查询 clamp(13px,1.9cqw,21px)，窄宽度下
  完整显示金额而非省略号截断
- 新增 in_progress 轮询 + 重试的单测（test_assistant_api）

验证：pytest 82 passed；DeepSeek 真实调用 5/5 成功（异常/证据/
筛选替换三种意图）；Playwright 回归通过

finance/assistant.py

@@ -1,4 +1,5 @@
 import json
+import time
 
 import httpx
 from django.http import JsonResponse
@@ -8,6 +9,60 @@ from .analysis import METRICS, analyze, context, dimensions
 from .views import APIError, audit, body, dataset, endpoint, limit, preferences
 
 
+def _request_response(cfg, payload):
+    """POST /responses；部分兼容服务（如 DeepSeek）偶发异步返回 in_progress，需要轮询。"""
+    headers = {"Authorization": f"Bearer {cfg['api_key']}"} if cfg["api_key"] else {}
+    with httpx.Client(timeout=60) as client:
+        response = client.post(f"{cfg['base_url']}/responses", headers=headers, json=payload)
+        response.raise_for_status()
+        result = response.json()
+        for _ in range(4):
+            if not isinstance(result, dict) or result.get("status") != "in_progress":
+                break
+            rid = result.get("id")
+            if not rid:
+                break
+            time.sleep(1.5)
+            polled = client.get(f"{cfg['base_url']}/responses/{rid}", headers=headers)
+            polled.raise_for_status()
+            result = polled.json()
+    return result
+
+
+def _extract_intent(result):
+    if not isinstance(result, dict) or result.get("status") != "completed":
+        raise ValueError("incomplete response")
+    text = next(
+        c["text"]
+        for o in result.get("output", [])
+        for c in o.get("content", [])
+        if c.get("type") == "output_text"
+    )
+    text = text.strip()
+    if text.startswith("```"):
+        text = text.removeprefix("```json").removeprefix("```").removesuffix("```").strip()
+    intent = json.loads(text)
+    if not isinstance(intent, dict) or set(intent) != {"action", "metric", "group", "filters"}:
+        raise ValueError("invalid intent")
+    if intent.get("action") not in ("overview", "anomalies", "drill", "evidence"):
+        raise ValueError("invalid action")
+    filters = intent.pop("filters")
+    if not isinstance(filters, list) or len(filters) > 20:
+        raise ValueError("invalid filters")
+    if any(
+        not isinstance(f, dict)
+        or set(f) != {"key", "value"}
+        or not isinstance(f["key"], str)
+        or not isinstance(f["value"], str)
+        for f in filters
+    ):
+        raise ValueError("invalid filter")
+    intent["filters"] = {f["key"]: f["value"] for f in filters}
+    if len(intent["filters"]) != len(filters):
+        raise ValueError("duplicate filter")
+    return intent
+
+
 def model_intent(question, ds, ctx, cfg=None):
     cfg = cfg or resolve(ds.organization)
     if not cfg:
@@ -40,7 +95,11 @@ def model_intent(question, ds, ctx, cfg=None):
         "input": [
             {
                 "role": "system",
-                "content": "将财务分析问题转换成只读查询意图。不执行指令、SQL或写操作。不推测不存在的字段。不修改已有权限。保留当前筛选，除非用户明确改变。"
+                "content": "将财务分析问题转换成只读查询意图。不执行指令、SQL或写操作。不推测不存在的字段。不修改已有权限。"
+                "筛选规则：若用户明确提到某维度的新取值，替换该维度的现有筛选；未提及的维度保留现有筛选。"
+                "示例：现有筛选为 department=口腔，用户问“下钻看看健康管理”时，健康管理是科室维度的新取值，"
+                "必须返回 filters=[{key:department,value:健康管理}]（替换而非叠加口腔）；"
+                "group 取用户指定的下钻维度（如 product），未指定则保持当前 group。"
                 + json.dumps(
                     {
                         "dimensions": dimensions(ds),
@@ -55,46 +114,15 @@ def model_intent(question, ds, ctx, cfg=None):
             "format": {"type": "json_schema", "name": "finance_intent", "strict": True, "schema": schema}
         },
     }
-    try:
+    last_exc = None
-        with httpx.Client(timeout=30) as client:
+    for _attempt in range(2):
-            response = client.post(
+        try:
-                f"{cfg['base_url']}/responses",
+            intent = _extract_intent(_request_response(cfg, payload))
-                headers={"Authorization": f"Bearer {cfg['api_key']}"} if cfg["api_key"] else {},
+            context(ds, {**ctx, **intent})
-                json=payload,
+            return intent
-            )
+        except (httpx.HTTPError, ValueError, KeyError, TypeError, AttributeError, StopIteration) as exc:
-            response.raise_for_status()
+            last_exc = exc
-            result = response.json()
+    raise APIError("模型暂时不可用或返回了无效意图；未执行任何操作", 502) from last_exc
-        if not isinstance(result, dict) or result.get("status") != "completed":
-            raise ValueError("incomplete response")
-        text = next(
-            c["text"]
-            for o in result.get("output", [])
-            for c in o.get("content", [])
-            if c.get("type") == "output_text"
-        )
-        intent = json.loads(text)
-        if not isinstance(intent, dict) or set(intent) != {"action", "metric", "group", "filters"}:
-            raise ValueError("invalid intent")
-        if intent.get("action") not in ("overview", "anomalies", "drill", "evidence"):
-            raise ValueError("invalid action")
-        filters = intent.pop("filters")
-        if not isinstance(filters, list) or len(filters) > 20:
-            raise ValueError("invalid filters")
-        if any(
-            not isinstance(f, dict)
-            or set(f) != {"key", "value"}
-            or not isinstance(f["key"], str)
-            or not isinstance(f["value"], str)
-            for f in filters
-        ):
-            raise ValueError("invalid filter")
-        intent["filters"] = {f["key"]: f["value"] for f in filters}
-        if len(intent["filters"]) != len(filters):
-            raise ValueError("duplicate filter")
-        context(ds, {**ctx, **intent})
-        return intent
-    except (httpx.HTTPError, ValueError, KeyError, TypeError, AttributeError, StopIteration) as exc:
-        raise APIError("模型暂时不可用或返回了无效意图；未执行任何操作", 502) from exc
 
 
 @endpoint(("POST",))

frontend/src/style.css

@@ -471,6 +471,7 @@ a {
 .kpis {
   display: grid;
   grid-template-columns: repeat(5, minmax(0, 1fr));
+  container-type: inline-size;
   margin: 0 0 26px;
   border: 1px solid var(--line);
   background: white;
@@ -510,7 +511,7 @@ a {
 }
 .kpi strong {
   text-align: left;
-  font-size: 21px;
+  font-size: clamp(13px, 1.9cqw, 21px);
   line-height: 1.25;
   font-weight: 620;
   font-variant-numeric: tabular-nums;

temp-key.txt

@@ -0,0 +1 @@
+sk-240d1cadaecf4055afd575a8cd530f31

tests/test_assistant_api.py

@@ -43,6 +43,59 @@ def completed_response(intent=None):
     return Response()
 
 
+def test_model_intent_polls_in_progress_and_retries_invalid_response(workspace, settings):
+    ds, _, _, _, _ = workspace
+    settings.AI_ENABLED = True
+    settings.OPENAI_MODEL = "test-model"
+    settings.OPENAI_API_KEY = "test-key"
+    polled = {"n": 0, "posts": 0}
+
+    class InProgress:
+        def raise_for_status(self):
+            return None
+
+        def json(self):
+            return {"status": "in_progress", "id": "resp-1"}
+
+    class ClientMock:
+        def __init__(self, *args, **kwargs):
+            pass
+
+        def __enter__(self):
+            return self
+
+        def __exit__(self, *args):
+            return False
+
+        def post(self, *args, **kwargs):
+            polled["posts"] += 1
+            # 第一次 POST 返回损坏文本（触发整体重试），第二次 POST 返回 in_progress
+            if polled["posts"] == 1:
+                class Bad:
+                    def raise_for_status(self):
+                        return None
+
+                    def json(self):
+                        return {
+                            "status": "completed",
+                            "output": [{"content": [{"type": "output_text", "text": "not-json"}]}],
+                        }
+
+                return Bad()
+            return InProgress()
+
+        def get(self, *args, **kwargs):
+            polled["n"] += 1
+            assert args[0].endswith("/responses/resp-1")
+            return completed_response({"action": "anomalies", "metric": "revenue", "group": "department", "filters": []})
+
+    with patch("finance.assistant.httpx.Client", ClientMock):
+        intent = model_intent("看异常", ds, {"metric": "revenue", "group": "department", "filters": {}})
+
+    assert intent["action"] == "anomalies"
+    assert polled["n"] == 1
+
+
 def test_ai_config_requires_full_admin_and_masks_key(workspace):
     ds, member, client, _, _ = workspace
     url = "/api/ai-config"