From 40dc2e07fe9a1050d0f81b67c15fdab3680234fb Mon Sep 17 00:00:00 2001
From: aaron <>
Date: Mon, 8 Jun 2026 09:06:34 +0800
Subject: [PATCH] 1

---
 app/services/llm_insights.py      | 80 ++++++++++++++++++++++---------
 app/web/routes_recommendations.py |  3 +-
 static/llm_insights.html          |  1 +
 tests/test_llm_insights.py        | 42 ++++++++++++++++
 4 files changed, 103 insertions(+), 23 deletions(-)

diff --git a/app/services/llm_insights.py b/app/services/llm_insights.py
index 968f852..f14ee06 100644
--- a/app/services/llm_insights.py
+++ b/app/services/llm_insights.py
@@ -114,6 +114,33 @@ def _parse_json_object_text(content: str) -> dict:
     return parsed
 
 
+def _chat_completion_request(base_url, api_key, model, messages, max_tokens, timeout, *, response_format=True):
+    body = {
+        "model": model,
+        "messages": messages,
+        "temperature": 0.2,
+        "max_tokens": max_tokens,
+    }
+    if response_format:
+        body["response_format"] = {"type": "json_object"}
+    return requests.post(
+        f"{base_url}/chat/completions",
+        headers={"Authorization": f"Bearer {api_key}", "Content-Type": "application/json"},
+        json=body,
+        timeout=timeout,
+    )
+
+
+def _parse_llm_response(resp, model):
+    if resp.status_code >= 400:
+        return {"status": "failed", "error": f"http_{resp.status_code}", "raw": resp.text[:1000], "model": model}
+    data = resp.json()
+    message = (((data.get("choices") or [{}])[0]).get("message") or {})
+    content = _message_content_text(message)
+    parsed = _parse_json_object_text(content)
+    return {"status": "success", "content": parsed, "model": model}
+
+
 def _call_llm_json(prompt_version, payload):
     params = get_llm_params()
     api_key = os.getenv(str(params.get("api_key_env") or "OPENAI_API_KEY"), "").strip()
@@ -132,29 +159,38 @@ def _call_llm_json(prompt_version, payload):
         "input": payload,
         "output_schema_hint": "JSON object with concise Chinese fields only",
     })
+    messages = [
+        {"role": "system", "content": system_prompt},
+        {"role": "user", "content": user_prompt},
+    ]
     try:
-        resp = requests.post(
-            f"{base_url}/chat/completions",
-            headers={"Authorization": f"Bearer {api_key}", "Content-Type": "application/json"},
-            json={
-                "model": model,
-                "messages": [
-                    {"role": "system", "content": system_prompt},
-                    {"role": "user", "content": user_prompt},
-                ],
-                "temperature": 0.2,
-                "max_tokens": max_tokens,
-                "response_format": {"type": "json_object"},
-            },
-            timeout=timeout,
-        )
-        if resp.status_code >= 400:
-            return {"status": "failed", "error": f"http_{resp.status_code}", "raw": resp.text[:1000], "model": model}
-        data = resp.json()
-        message = (((data.get("choices") or [{}])[0]).get("message") or {})
-        content = _message_content_text(message)
-        parsed = _parse_json_object_text(content)
-        return {"status": "success", "content": parsed, "model": model}
+        resp = _chat_completion_request(base_url, api_key, model, messages, max_tokens, timeout, response_format=True)
+        first = _parse_llm_response(resp, model)
+        if first.get("status") == "success":
+            return first
+        return first
+    except ValueError as exc:
+        # Some OpenAI-compatible providers occasionally return an empty message
+        # when strict JSON response_format is enabled. Retry once with prompt-only
+        # JSON enforcement so the job does not fail silently on provider quirks.
+        retryable = str(exc) in ("llm_empty_content", "llm_empty_json_object")
+        if not retryable:
+            return {"status": "failed", "error": str(exc)[:1000], "model": model}
+        try:
+            retry_messages = [
+                {
+                    "role": "system",
+                    "content": system_prompt + " Return one valid JSON object. No markdown, no prose.",
+                },
+                {"role": "user", "content": user_prompt},
+            ]
+            resp = _chat_completion_request(base_url, api_key, model, retry_messages, max_tokens, timeout, response_format=False)
+            result = _parse_llm_response(resp, model)
+            if result.get("status") == "success":
+                result["retry"] = "without_response_format"
+            return result
+        except Exception as retry_exc:
+            return {"status": "failed", "error": f"{exc}; retry_failed:{str(retry_exc)[:800]}", "model": model}
     except json.JSONDecodeError as exc:
         return {"status": "failed", "error": f"invalid_json:{exc}", "model": model}
     except Exception as exc:
diff --git a/app/web/routes_recommendations.py b/app/web/routes_recommendations.py
index f778d9c..e567996 100644
--- a/app/web/routes_recommendations.py
+++ b/app/web/routes_recommendations.py
@@ -34,6 +34,7 @@ def _friendly_llm_item(item):
     type_label = {
         "recommendation": "推荐解释",
         "sentiment": "舆情解读",
+        "sentiment_batch": "批量舆情分析",
         "review": "复盘 memo",
     }.get(target_type, target_type or "未知任务")
     status_label = {
@@ -41,7 +42,7 @@ def _friendly_llm_item(item):
         "failed": "失败",
         "skipped": "跳过",
     }.get(status, status or "未知")
-    subject = payload.get("symbol") or payload.get("related_symbol") or payload.get("title") or payload.get("run_date") or item.get("target_id")
+    subject = payload.get("symbol") or payload.get("related_symbol") or payload.get("title") or payload.get("run_date") or payload.get("target_id") or item.get("target_id")
     summary = content.get("summary") or content.get("memo") or content.get("why_now_or_not") or content.get("raw") or item.get("error") or ""
     return {
         "id": item.get("id"),
diff --git a/static/llm_insights.html b/static/llm_insights.html
index 1742340..7e08388 100644
--- a/static/llm_insights.html
+++ b/static/llm_insights.html
@@ -17,6 +17,7 @@
         <option value="">全部类型</option>
         <option value="recommendation">推荐解释</option>
         <option value="sentiment">舆情解读</option>
+        <option value="sentiment_batch">批量舆情分析</option>
         <option value="review">复盘 memo</option>
       </select>
       <select class="select" id="statusSel" onchange="reloadFirst()">
diff --git a/tests/test_llm_insights.py b/tests/test_llm_insights.py
index fa4b69e..5750c76 100644
--- a/tests/test_llm_insights.py
+++ b/tests/test_llm_insights.py
@@ -152,6 +152,48 @@ def test_llm_empty_content_is_not_marked_success(monkeypatch):
     assert "llm_empty_content" in result["error"]
 
 
+def test_llm_empty_content_retries_without_response_format(monkeypatch):
+    monkeypatch.setattr(llm_insights, "get_llm_params", lambda: {
+        "enabled": True,
+        "base_url": "https://llm.example/v1",
+        "api_key_env": "TEST_LLM_KEY",
+        "model": "deepseek-v4-pro",
+        "timeout": 5,
+        "max_tokens": 100,
+        "modules": {},
+    })
+    monkeypatch.setenv("TEST_LLM_KEY", "test-key")
+    calls = []
+
+    class EmptyResp:
+        status_code = 200
+        text = "{}"
+
+        def json(self):
+            return {"choices": [{"message": {"content": ""}}]}
+
+    class JsonResp:
+        status_code = 200
+        text = "{}"
+
+        def json(self):
+            return {"choices": [{"message": {"content": "{\"summary\":\"retry ok\"}"}}]}
+
+    def fake_post(*args, **kwargs):
+        calls.append(kwargs.get("json") or {})
+        return EmptyResp() if len(calls) == 1 else JsonResp()
+
+    monkeypatch.setattr(llm_insights.requests, "post", fake_post)
+
+    result = llm_insights._call_llm_json("test_prompt", {"symbol": "AAA/USDT"})
+
+    assert result["status"] == "success"
+    assert result["content"]["summary"] == "retry ok"
+    assert result["retry"] == "without_response_format"
+    assert calls[0]["response_format"] == {"type": "json_object"}
+    assert "response_format" not in calls[1]
+
+
 def test_llm_json_code_fence_is_parsed(monkeypatch):
     monkeypatch.setattr(llm_insights, "get_llm_params", lambda: {
         "enabled": True,