Merge pull request #240 from yushenglin01/fix/ai-analysis-focus

修复 AI 分析关注重点不突出及静默失效问题

backend/app/services/ai_provider.py

@@ -72,8 +72,8 @@ _ANSI_RE = re.compile(r"\x1b\[[0-9;?]*[ -/]*[@-~]")
 
 
 # ----------------------------------------------------------------
-# 用户 focus 输入净化 — 防止通过"特别关注"绕过红线诱导 AI 给出买卖建议
-# 命中任一敏感词时,整个 focus 被丢弃(返回空串),由各 analyzer 据此跳过注入。
+# 用户 focus 输入规范化。交易建议类表达不会被静默丢弃,而是由统一提示词
+# 转换成客观价位、风险和情景分析,避免历史报告显示了 focus、模型却没有收到。
 # ----------------------------------------------------------------
 _FOCUS_BLOCKLIST = re.compile(
     r"买入|卖出|加仓|减仓|轻仓|重仓|半仓|全仓|仓位|止损|止盈|"
@@ -87,19 +87,36 @@ _FOCUS_BLOCKLIST = re.compile(
 
 
 def sanitize_focus(focus: str) -> str:
-    """净化用户输入的 focus 文本。
-
-    命中交易指令/投资建议类敏感词时返回空串,阻止其注入 AI 提示词。
-    这是对系统提示词红线的兜底:即便用户试图通过 focus 绕过,也不会生效。
-    """
+    """规范化 focus 中的首尾空白与连续换行。"""
     if not focus:
         return ""
-    text = focus.strip()
+    text = re.sub(r"\s+", " ", focus).strip()
+    return text
+
+
+def build_focus_instruction(focus: str, *, report_name: str = "分析报告") -> str:
+    """构建所有报告共用的关注重点指令。
+
+    有关注点时要求模型在固定报告结构之前先直接回应。若原问题涉及交易
+    建议,保留问题语义但要求转换成中立的数据分析,不再无提示地整段丢弃。
+    """
+    text = sanitize_focus(focus)
     if not text:
         return ""
+
+    lines = [
+        "## 用户关注重点(必须优先回应)",
+        f"用户关注: {text}",
+        f"请在完整{report_name}最前面先输出 `### 0. 🔎 关注重点回应`,"
+        "用 2-4 条带具体数据的结论直接回应;随后继续完成既定报告结构,"
+        "并在相关章节加深分析。不要只复述问题。",
+    ]
     if _FOCUS_BLOCKLIST.search(text):
-        return ""
-    return text
+        lines.append(
+            "该关注点含有买卖、仓位、目标价或预测类表达。不得给出相应操作结论;"
+            "请将其转换为客观的技术/财务状态、关键价位、风险因素和条件情景后回应。"
+        )
+    return "\n".join(lines)
 
 
 def current_ai_provider() -> str:
@@ -309,11 +326,13 @@ async def stream_ai_text(
     temperature: float | None = 0.5,
     max_tokens: int | None = 4000,
     timeout: float = 180.0,
+    prefer_final_answer: bool = False,
 ) -> AsyncIterator[str]:
     """Yield text deltas from the configured provider.
 
     Codex CLI only exposes the final assistant message for this use case, so it
-    yields one complete chunk after the command exits.
+    yields one complete chunk after the command exits. ``prefer_final_answer``
+    lets compatible providers prioritize visible content over hidden reasoning.
 
     max_tokens=None 表示不限制输出(同 generate_ai_text 的说明)。
     """
@@ -328,6 +347,7 @@ async def stream_ai_text(
         temperature=temperature,
         max_tokens=max_tokens,
         timeout=timeout,
+        prefer_final_answer=prefer_final_answer,
     ):
         yield chunk
 
@@ -374,6 +394,7 @@ async def _stream_openai(
     temperature: float | None,
     max_tokens: int | None,
     timeout: float,
+    prefer_final_answer: bool,
 ) -> AsyncIterator[str]:
     ai_key = secrets_store.get_ai_key()
     if not ai_key:
@@ -381,15 +402,16 @@ async def _stream_openai(
 
     client = _openai_client(ai_key, timeout)
     model = current_ai_model()
+    base_url = secrets_store.get_ai_config("ai_base_url", settings.ai_base_url)
     req_messages = list(messages)
 
-    async def _iter(stream):
-        async for chunk in stream:
-            delta = chunk.choices[0].delta if chunk.choices else None
-            if delta and delta.content:
-                yield delta.content
-
-    kwargs = _openai_kwargs(temperature=temperature, max_tokens=max_tokens)
+    kwargs = _openai_kwargs(
+        temperature=temperature,
+        max_tokens=max_tokens,
+        model=model,
+        base_url=base_url,
+        prefer_final_answer=prefer_final_answer,
+    )
     while True:
         try:
             stream = await client.chat.completions.create(
@@ -410,7 +432,7 @@ async def _stream_openai(
             raise
 
     try:
-        async for piece in _iter(stream):
+        async for piece in _iter_openai_text(stream):
             yield piece
     except Exception as exc:
         if _is_openai_transport_error(exc):
@@ -418,6 +440,53 @@ async def _stream_openai(
         raise
 
 
+_LENGTH_FINISH_REASONS = {"length", "max_tokens", "max_output_tokens"}
+
+
+async def _iter_openai_text(stream) -> AsyncIterator[str]:
+    """Normalize an OpenAI-compatible stream into complete text deltas.
+
+    Reasoning models may spend the entire completion budget on
+    ``reasoning_content`` and finish with HTTP 200 but no user-visible text.
+    Treat that response, and any length-truncated partial response, as a
+    terminal generation error instead of silently reporting success.
+    """
+    content_seen = False
+    reasoning_seen = False
+    finish_reason = ""
+
+    async for chunk in stream:
+        choices = getattr(chunk, "choices", None) or []
+        if not choices:
+            continue
+        choice = choices[0]
+        reason = getattr(choice, "finish_reason", None)
+        if reason:
+            finish_reason = str(reason)
+
+        delta = getattr(choice, "delta", None)
+        if delta is None:
+            continue
+        if getattr(delta, "reasoning_content", None):
+            reasoning_seen = True
+        content = getattr(delta, "content", None)
+        if content:
+            content_seen = True
+            yield content
+
+    if finish_reason in _LENGTH_FINISH_REASONS:
+        if reasoning_seen and not content_seen:
+            raise RuntimeError(
+                "AI 推理达到输出长度上限, 未生成正文; 请提高输出 Token 上限或改用非推理模型"
+            )
+        raise RuntimeError("AI 输出达到长度上限, 内容不完整; 请提高输出 Token 上限后重试")
+
+    if not content_seen:
+        if reasoning_seen:
+            raise RuntimeError("AI 仅返回推理内容, 未生成正文; 请检查模型配置或改用非推理模型")
+        raise RuntimeError("AI 服务未返回正文内容; 请检查模型配置或稍后重试")
+
+
 def _openai_client(api_key: str, timeout: float):
     from openai import AsyncOpenAI
 
@@ -479,8 +548,18 @@ def _openai_retry_kwargs(exc: Exception, kwargs: dict) -> dict | None:
     return None
 
 
-def _openai_kwargs(*, temperature: float | None, max_tokens: int | None) -> dict:
-    """Build OpenAI create() kwargs; optional parameters are omitted when empty.
+_DEEPSEEK_V4_MODELS = {"deepseek-v4-flash", "deepseek-v4-pro"}
+
+
+def _openai_kwargs(
+    *,
+    temperature: float | None,
+    max_tokens: int | None,
+    model: str = "",
+    base_url: str = "",
+    prefer_final_answer: bool = False,
+) -> dict:
+    """Build OpenAI create() kwargs and map supported provider capabilities.
 
     max_tokens=None 时不传 — 由服务端默认上限管理(推理模型的思考 token 也
     计入该参数预算, 限制会挤占正文, 见 stream_ai_text 文档)。
@@ -494,6 +573,15 @@ def _openai_kwargs(*, temperature: float | None, max_tokens: int | None) -> dict
         reasoning_effort = current_openai_reasoning_effort()
         if reasoning_effort:
             kwargs["reasoning_effort"] = reasoning_effort
+    if (
+        prefer_final_answer
+        and model.strip().lower() in _DEEPSEEK_V4_MODELS
+        and urlsplit(base_url.strip()).hostname == "api.deepseek.com"
+    ):
+        # DeepSeek V4 defaults to thinking mode. For report-style tasks the
+        # hidden reasoning shares max_tokens with the final answer and can
+        # exhaust the budget before any visible content is emitted.
+        kwargs["extra_body"] = {"thinking": {"type": "disabled"}}
     return kwargs
 
 

backend/app/services/concept_rotation_analyzer.py

@@ -278,10 +278,10 @@ def _build_user_prompt(signals: dict, overview: dict, days: int, dates: list[str
         _build_signal_block("🎰 游资特征 (排名波动大)", signals.get("hot_money", [])),
     ]
 
-    from app.services.ai_provider import sanitize_focus
-    safe_focus = sanitize_focus(focus)
-    if safe_focus:
-        parts.extend(["", f"本次分析请特别关注: {safe_focus}"])
+    from app.services.ai_provider import build_focus_instruction
+    focus_instruction = build_focus_instruction(focus, report_name=f"{dim}轮动分析报告")
+    if focus_instruction:
+        parts.extend(["", focus_instruction])
 
     return "\n".join(parts)
 
@@ -374,6 +374,7 @@ async def analyze_rotation_stream(
             temperature=0.5,
             # 不限制输出(推理模型思考 token 计入预算, 见 ai_provider.stream_ai_text)
             max_tokens=None,
+            prefer_final_answer=True,
         ):
             got_content = True
             yield json.dumps({"type": "delta", "content": delta}, ensure_ascii=False)

backend/app/services/financial_analyzer.py

@@ -136,13 +136,10 @@ def _build_user_prompt(fins: dict[str, list[dict]], symbol: str, focus: str) ->
         data_json,
         "```",
     ]
-    from app.services.ai_provider import sanitize_focus
-    safe_focus = sanitize_focus(focus)
-    if safe_focus:
-        lines.extend([
-            "",
-            f"本次分析请特别关注: {safe_focus}",
-        ])
+    from app.services.ai_provider import build_focus_instruction
+    focus_instruction = build_focus_instruction(focus, report_name="财务分析报告")
+    if focus_instruction:
+        lines.extend(["", focus_instruction])
     return "\n".join(lines)
 
 
@@ -187,6 +184,7 @@ async def analyze_financials_stream(
             temperature=0.4,
             # 不限制输出(推理模型思考 token 计入预算, 见 ai_provider.stream_ai_text)
             max_tokens=None,
+            prefer_final_answer=True,
         ):
             got_content = True
             yield json.dumps({"type": "delta", "content": delta}, ensure_ascii=False)

backend/app/services/market_recap.py

@@ -237,10 +237,10 @@ def _build_user_prompt(overview: dict, news: list[dict], focus: str, lhb_context
             "消息催化一节请直接从量价异动给出可能的催化逻辑结论,不要编造具体消息,也不要复述本说明。)",
         ])
 
-    from app.services.ai_provider import sanitize_focus
-    safe_focus = sanitize_focus(focus)
-    if safe_focus:
-        parts.extend(["", f"本次复盘请特别关注: {safe_focus}"])
+    from app.services.ai_provider import build_focus_instruction
+    focus_instruction = build_focus_instruction(focus, report_name="大盘复盘报告")
+    if focus_instruction:
+        parts.extend(["", focus_instruction])
 
     return "\n".join(parts)
 
@@ -336,6 +336,7 @@ async def recap_market_stream(
             temperature=0.5,
             # 不限制输出(推理模型思考 token 计入预算, 见 ai_provider.stream_ai_text)
             max_tokens=None,
+            prefer_final_answer=True,
         ):
             got_content = True
             yield json.dumps({"type": "delta", "content": delta}, ensure_ascii=False)

backend/app/services/stock_analyzer.py

@@ -16,8 +16,8 @@ from __future__ import annotations
 
 import json
 import logging
+from collections.abc import AsyncIterator
 from pathlib import Path
-from typing import AsyncIterator
 
 import polars as pl
 
@@ -238,10 +238,10 @@ def _build_user_prompt(
             "请按系统提示词第 4 节的说明,在基本面/财务面维度给出\"接入中\"的友好提示,不要编造数据。)",
         ])
 
-    from app.services.ai_provider import sanitize_focus
-    safe_focus = sanitize_focus(focus)
-    if safe_focus:
-        parts.extend(["", f"本次分析请特别关注: {safe_focus}"])
+    from app.services.ai_provider import build_focus_instruction
+    focus_instruction = build_focus_instruction(focus, report_name="个股分析报告")
+    if focus_instruction:
+        parts.extend(["", focus_instruction])
     return "\n".join(parts)
 
 
@@ -325,11 +325,12 @@ async def analyze_stock_stream(
             # 不限制输出: 推理模型(deepseek reasoner 系)思考 token 计入 max_tokens
             # 预算, 固定上限会把正文挤光(实测 4500 全被推理吃掉 → 正文 0 字)。
             max_tokens=None,
+            prefer_final_answer=True,
         ):
             got_content = True
             yield json.dumps({"type": "delta", "content": delta}, ensure_ascii=False)
 
-    except Exception as e:  # noqa: BLE001
+    except Exception as e:
         logger.exception("AI stock analysis failed for %s: %s", symbol, e)
         yield json.dumps({"type": "error", "message": f"AI 分析失败: {e}"}, ensure_ascii=False)
         return

backend/tests/test_ai_analysis_focus.py

@@ -0,0 +1,62 @@
+"""AI 分析关注重点的统一 Prompt 契约测试。"""
+
+from app.services.ai_provider import build_focus_instruction, sanitize_focus
+from app.services.concept_rotation_analyzer import (
+    _build_user_prompt as build_rotation_user_prompt,
+)
+from app.services.financial_analyzer import (
+    _build_user_prompt as build_financial_user_prompt,
+)
+from app.services.market_recap import (
+    _build_user_prompt as build_recap_user_prompt,
+)
+from app.services.stock_analyzer import (
+    _build_user_prompt as build_stock_user_prompt,
+)
+
+
+def test_focus_whitespace_is_normalized() -> None:
+    assert sanitize_focus("  支撑位\n  多少  ") == "支撑位 多少"
+
+
+def test_trade_wording_is_reframed_instead_of_silently_dropped() -> None:
+    instruction = build_focus_instruction("现在能买吗,目标价多少", report_name="个股分析报告")
+
+    assert "用户关注: 现在能买吗,目标价多少" in instruction
+    assert "关注重点回应" in instruction
+    assert "不得给出相应操作结论" in instruction
+    assert "关键价位" in instruction
+
+
+def test_safe_focus_requires_a_direct_answer_without_extra_warning() -> None:
+    instruction = build_focus_instruction("支撑位多少", report_name="个股分析报告")
+
+    assert "用户关注: 支撑位多少" in instruction
+    assert "用 2-4 条带具体数据的结论直接回应" in instruction
+    assert "不得给出相应操作结论" not in instruction
+
+
+def test_all_focus_enabled_analyzers_share_the_priority_contract() -> None:
+    focus = "支撑位多少"
+    prompts = [
+        build_stock_user_prompt([], {}, {}, 10.0, "600000.SH", focus),
+        build_financial_user_prompt({}, "600000.SH", focus),
+        build_recap_user_prompt({}, [], focus),
+        build_rotation_user_prompt({}, {}, 12, [], focus),
+    ]
+
+    for prompt in prompts:
+        assert "## 用户关注重点(必须优先回应)" in prompt
+        assert f"用户关注: {focus}" in prompt
+        assert "### 0. 🔎 关注重点回应" in prompt
+
+
+def test_empty_focus_does_not_add_focus_section() -> None:
+    prompts = [
+        build_stock_user_prompt([], {}, {}, 10.0, "600000.SH", ""),
+        build_financial_user_prompt({}, "600000.SH", ""),
+        build_recap_user_prompt({}, [], ""),
+        build_rotation_user_prompt({}, {}, 12, [], ""),
+    ]
+
+    assert all("用户关注重点" not in prompt for prompt in prompts)

backend/tests/test_ai_provider.py

@@ -1,6 +1,7 @@
 from __future__ import annotations
 
 import tomllib
+from types import SimpleNamespace
 
 import httpx
 import openai
@@ -48,6 +49,41 @@ def test_normalize_openai_base_url_strips_trailing_slash():
     assert normalize_openai_base_url("https://open.bigmodel.cn/api/paas/v4/") == "https://open.bigmodel.cn/api/paas/v4"
 
 
+def test_openai_kwargs_prefers_final_answer_for_official_deepseek_v4():
+    kwargs = ai_provider._openai_kwargs(
+        temperature=0.5,
+        max_tokens=8192,
+        model="deepseek-v4-pro",
+        base_url="https://api.deepseek.com/v1",
+        prefer_final_answer=True,
+    )
+
+    assert kwargs["extra_body"] == {"thinking": {"type": "disabled"}}
+
+
+def test_openai_kwargs_does_not_send_deepseek_option_to_other_providers():
+    kwargs = ai_provider._openai_kwargs(
+        temperature=0.5,
+        max_tokens=8192,
+        model="gpt-5.5",
+        base_url="https://api.openai.com/v1",
+        prefer_final_answer=True,
+    )
+
+    assert "extra_body" not in kwargs
+
+
+def test_openai_kwargs_keeps_deepseek_default_without_final_answer_preference():
+    kwargs = ai_provider._openai_kwargs(
+        temperature=0.5,
+        max_tokens=8192,
+        model="deepseek-v4-pro",
+        base_url="https://api.deepseek.com/v1",
+    )
+
+    assert "extra_body" not in kwargs
+
+
 def test_format_openai_error_hides_html_gateway_body():
     response = httpx.Response(
         504,
@@ -400,6 +436,67 @@ def test_save_ai_settings_rejects_non_positive(monkeypatch):
         settings_api.save_ai_settings(req2)
 
 
+async def _fake_openai_stream(*chunks):
+    for chunk in chunks:
+        yield chunk
+
+
+def _stream_chunk(*, content=None, reasoning_content=None, finish_reason=None):
+    delta = SimpleNamespace(content=content, reasoning_content=reasoning_content)
+    choice = SimpleNamespace(delta=delta, finish_reason=finish_reason)
+    return SimpleNamespace(choices=[choice])
+
+
+@pytest.mark.asyncio
+async def test_iter_openai_text_rejects_reasoning_only_length_exhaustion():
+    stream = _fake_openai_stream(
+        _stream_chunk(reasoning_content="内部推理"),
+        _stream_chunk(finish_reason="length"),
+    )
+
+    with pytest.raises(RuntimeError, match="推理达到输出长度上限"):
+        async for _ in ai_provider._iter_openai_text(stream):
+            pass
+
+
+@pytest.mark.asyncio
+async def test_iter_openai_text_rejects_truncated_partial_content():
+    stream = _fake_openai_stream(
+        _stream_chunk(content="未完成正文"),
+        _stream_chunk(finish_reason="length"),
+    )
+    pieces = []
+
+    with pytest.raises(RuntimeError, match="输出达到长度上限"):
+        async for piece in ai_provider._iter_openai_text(stream):
+            pieces.append(piece)
+
+    assert pieces == ["未完成正文"]
+
+
+@pytest.mark.asyncio
+async def test_iter_openai_text_yields_complete_content_and_ignores_reasoning():
+    stream = _fake_openai_stream(
+        _stream_chunk(reasoning_content="内部推理"),
+        _stream_chunk(content="完整"),
+        _stream_chunk(content="正文"),
+        _stream_chunk(finish_reason="stop"),
+    )
+
+    pieces = [piece async for piece in ai_provider._iter_openai_text(stream)]
+
+    assert pieces == ["完整", "正文"]
+
+
+@pytest.mark.asyncio
+async def test_iter_openai_text_rejects_stream_without_content():
+    stream = _fake_openai_stream(_stream_chunk(finish_reason="stop"))
+
+    with pytest.raises(RuntimeError, match="未返回正文内容"):
+        async for _ in ai_provider._iter_openai_text(stream):
+            pass
+
+
 def test_codex_process_env_excludes_application_secrets(monkeypatch, tmp_path):
     monkeypatch.setenv("PATH", "test-path")
     monkeypatch.setenv("HTTPS_PROXY", "http://proxy.example")