feat(llm): 增加 DeepSeek 纯文本适配器与 DEEPSEEK_*

MA_LLM_TEXT_PROVIDER=deepseek|deep_seek；.env.example 增加 ②-D；单测与 Kimi 同形。

Made-with: Cursor

.env.example

@@ -29,8 +29,8 @@ CSRF_TRUSTED_ORIGINS=http://localhost:5173,http://127.0.0.1:5173
 # OPENAI_TEXT_MODEL=
 # 别名：LLM_API_KEY、LLM_BASE_URL、LLM_MODEL
 
-# --- LLM ② 纯文本（报告/策略/关键词等 call_llm；与 ②-B、②-C 三选一，只让一条 MA_LLM_TEXT_PROVIDER 生效，换方案时先改/注释本行再启 ②-B 或 ②-C）---
-# 方案 A（默认）= 经 openai_gateway 调 ① 的 Base URL；要换 B=官站、C=Kimi 时，把下面这行改成 openai_official|chatgpt|kimi 等，并去对应小节填凭据
+# --- LLM ② 纯文本（报告/策略/关键词等 call_llm；与 ②-B～②-D 多选一，只让一条 MA_LLM_TEXT_PROVIDER 生效，换方案时先改/注释本行再启别节）---
+# 方案 A（默认）= 经 openai_gateway 调 ①；要换 B=官站、C=Kimi、D=DeepSeek 时，把下面这行改成 openai_official|kimi|deepseek 等，并去对应小节填凭据
 MA_LLM_TEXT_PROVIDER=crawler_openai_compatible
 
 # --- ②-B 纯文本用 OpenAI 官站（仅当 MA=openai_official / openai_chatgpt / chatgpt；与 ②-A、②-C 互斥；凭据与 ① 可不同）---
@@ -50,6 +50,15 @@ MA_LLM_TEXT_PROVIDER=crawler_openai_compatible
 # 别名：MOONSHOT_API_KEY、KIMI_MODEL、MOONSHOT_MODEL
 # KIMI_TIMEOUT=600
 
+# --- ②-D 纯文本用 DeepSeek 官方（仅当 MA=deepseek / deep_seek；与 ②-A～②-C 互斥；配料/视觉仍只走 ① 的 OPENAI_*）---
+# MA_LLM_TEXT_PROVIDER=deepseek
+# DEEPSEEK_API_KEY=sk-...
+# DEEPSEEK_BASE_URL=https://api.deepseek.com/v1
+# DEEPSEEK_TEXT_MODEL=deepseek-chat
+# DEEPSEEK_CONTEXT_WINDOW=64000
+# 别名：DEEPSEEK_MODEL
+# DEEPSEEK_TIMEOUT=600
+
 # --- 报告/策略：部分 LLM 块单独温度（默认 0.1；与上面 MA 无冲突）---
 # MA_STRATEGY_LLM_TEMPERATURE=0.1
 

backend/pipeline/llm/providers/__init__.py

@@ -4,6 +4,7 @@ from __future__ import annotations
 
 from .adapters import (
     CrawlerOpenAiCompatibleTextLlm,
+    DeepSeekTextLlm,
     KimiMoonshotTextLlm,
     OpenAiOfficialChatGptTextLlm,
 )
@@ -12,6 +13,7 @@ from .protocol import TextLlmClient
 
 __all__ = [
     "CrawlerOpenAiCompatibleTextLlm",
+    "DeepSeekTextLlm",
     "KimiMoonshotTextLlm",
     "OpenAiOfficialChatGptTextLlm",
     "TextLlmClient",

backend/pipeline/llm/providers/adapters/__init__.py

@@ -2,11 +2,13 @@
 from __future__ import annotations
 
 from .crawler_openai_compatible import CrawlerOpenAiCompatibleTextLlm
+from .deepseek_text import DeepSeekTextLlm
 from .kimi_moonshot_text import KimiMoonshotTextLlm
 from .openai_official_chatgpt import OpenAiOfficialChatGptTextLlm
 
 __all__ = [
     "CrawlerOpenAiCompatibleTextLlm",
+    "DeepSeekTextLlm",
     "KimiMoonshotTextLlm",
     "OpenAiOfficialChatGptTextLlm",
 ]

backend/pipeline/llm/providers/adapters/deepseek_text.py

@@ -0,0 +1,140 @@
+"""
+DeepSeek 官方 OpenAI 兼容 `chat/completions`（`https://api.deepseek.com`），**仅用于纯文本**（`call_llm` / 报告 / 策略）。
+
+与 `OPENAI_*` / 配料多模 分离，独立 `DEEPSEEK_*` 凭据。
+
+启用：`MA_LLM_TEXT_PROVIDER=deepseek`（或 `deep_seek`）。
+
+环境变量：见 `.env.example`（`DEEPSEEK_API_KEY`、基址、模型、上下文字数预检 等）。
+"""
+from __future__ import annotations
+
+import os
+from typing import Any
+
+import requests
+
+from pipeline.openai_gateway.chat_content import normalize_message_content
+from pipeline.openai_gateway.estimate import (
+    estimate_chat_input_tokens as estimate_crawler_style_input_tokens,
+)
+
+_DEFAULT_BASE = "https://api.deepseek.com/v1"
+_DEFAULT_MODEL = "deepseek-chat"
+# 常见 64k 级；若用长上下文/官方调整上限可改 DEEPSEEK_CONTEXT_WINDOW
+_DEFAULT_CTX = 64_000
+
+_BUF = 256
+_WANT_MAX = 8192
+
+
+def _read_timeout() -> tuple[float, float]:
+    read = 600
+    raw = (
+        os.environ.get("DEEPSEEK_TIMEOUT")
+        or os.environ.get("LLM_CHAT_TIMEOUT")
+        or os.environ.get("OPENAI_TIMEOUT")
+        or ""
+    ).strip()
+    if raw:
+        try:
+            read = max(60, int(raw))
+        except ValueError:
+            pass
+    conn = 30.0
+    raw_c = (os.environ.get("LLM_CHAT_CONNECT_TIMEOUT") or "").strip()
+    if raw_c:
+        try:
+            conn = max(5.0, float(raw_c))
+        except ValueError:
+            pass
+    return (conn, float(read))
+
+
+def _resolve_deepseek_credentials() -> tuple[str, str, str]:
+    key = (os.environ.get("DEEPSEEK_API_KEY") or "").strip()
+    if not key:
+        raise ValueError(
+            "使用 deepseek 文本适配器需设置 DEEPSEEK_API_KEY，"
+            "与配料/视觉所用 OPENAI_API_KEY 分开配置。"
+        )
+    base = (os.environ.get("DEEPSEEK_BASE_URL") or _DEFAULT_BASE).strip().rstrip("/")
+    model = (
+        (os.environ.get("DEEPSEEK_TEXT_MODEL") or os.environ.get("DEEPSEEK_MODEL") or _DEFAULT_MODEL)
+    ).strip()
+    return key, base, model
+
+
+def _context_window() -> int:
+    raw = (os.environ.get("DEEPSEEK_CONTEXT_WINDOW") or str(_DEFAULT_CTX)).strip()
+    try:
+        return max(4096, int(raw))
+    except ValueError:
+        return _DEFAULT_CTX
+
+
+def _default_temperature() -> float:
+    return 0.2
+
+
+class DeepSeekTextLlm:
+    """DeepSeek `chat/completions`；与 `KimiMoonshotTextLlm` 同形（max_tokens 预检）。"""
+
+    def complete_text(
+        self,
+        system_prompt: str,
+        user_prompt: str,
+        *,
+        temperature: float | None = None,
+    ) -> str:
+        api_key, base, model = _resolve_deepseek_credentials()
+        body: dict[str, Any] = {
+            "model": model,
+            "messages": [
+                {"role": "system", "content": system_prompt},
+                {"role": "user", "content": user_prompt},
+            ],
+            "temperature": _default_temperature() if temperature is None else float(temperature),
+            "max_tokens": _WANT_MAX,
+        }
+        est = estimate_crawler_style_input_tokens(system_prompt, user_prompt)
+        context_window = _context_window()
+        if est >= context_window - _BUF - 256:
+            raise ValueError(
+                f"提示词过长（估算输入约 {est} tokens，DEEPSEEK_CONTEXT_WINDOW={context_window}），"
+                "请缩小输入或调大 DEEPSEEK_TEXT_MODEL / DEEPSEEK_CONTEXT_WINDOW。"
+            )
+        avail = context_window - est - _BUF
+        want = int(body.get("max_tokens") or _WANT_MAX)
+        body["max_tokens"] = max(256, min(want, max(avail, 256)))
+        r = requests.post(
+            f"{base}/chat/completions",
+            headers={
+                "Authorization": f"Bearer {api_key}",
+                "Content-Type": "application/json",
+            },
+            json=body,
+            timeout=_read_timeout(),
+        )
+        try:
+            r.raise_for_status()
+        except requests.HTTPError as e:
+            snippet = ""
+            if e.response is not None:
+                snippet = (e.response.text or "")[:1200].replace("\r\n", "\n").replace("\n", " ")
+            if snippet:
+                raise requests.HTTPError(
+                    f"{e!s} | body: {snippet}",
+                    response=e.response,
+                    request=e.request,
+                ) from e
+            raise
+        data = r.json()
+        msg = (data.get("choices") or [{}])[0].get("message") or {}
+        return normalize_message_content(msg.get("content"))
+
+    def estimate_input_tokens(self, system_prompt: str, user_prompt: str) -> int:
+        return estimate_crawler_style_input_tokens(system_prompt, user_prompt)
+
+    def context_window_tokens(self) -> int:
+        return _context_window()

backend/pipeline/llm/providers/factory.py

@@ -6,6 +6,7 @@ from __future__ import annotations
 import os
 
 from .adapters.crawler_openai_compatible import CrawlerOpenAiCompatibleTextLlm
+from .adapters.deepseek_text import DeepSeekTextLlm
 from .adapters.kimi_moonshot_text import KimiMoonshotTextLlm
 from .adapters.openai_official_chatgpt import OpenAiOfficialChatGptTextLlm
 from .protocol import TextLlmClient
@@ -35,10 +36,12 @@ def _build_client(pid: str) -> TextLlmClient:
         return OpenAiOfficialChatGptTextLlm()
     if pid in ("kimi", "moonshot", "kimi_moonshot", "moonshot_kimi"):
         return KimiMoonshotTextLlm()
+    if pid in ("deepseek", "deep_seek"):
+        return DeepSeekTextLlm()
     known = (
         "crawler_openai_compatible, crawler, default, openai_compatible, "
         "openai_official, openai_chatgpt, chatgpt, "
-        "kimi, moonshot, kimi_moonshot"
+        "kimi, moonshot, kimi_moonshot, deepseek, deep_seek"
     )
     raise ValueError(
         f"不支持的 {_ENV_KEY}={pid!r}；已知取值：{known}。",

backend/pipeline/tests/test_text_llm_providers.py

@@ -6,6 +6,7 @@ import pytest
 
 from pipeline.llm.providers import (
     CrawlerOpenAiCompatibleTextLlm,
+    DeepSeekTextLlm,
     KimiMoonshotTextLlm,
     OpenAiOfficialChatGptTextLlm,
     get_text_llm,
@@ -64,6 +65,46 @@ def test_kimi_complete_text_uses_post_and_returns_content(monkeypatch: pytest.Mo
     assert out == "kimi_out"
 
 
+def test_deepseek_complete_text_uses_post_and_returns_content(monkeypatch: pytest.MonkeyPatch) -> None:
+    monkeypatch.setenv("DEEPSEEK_API_KEY", "sk-ds-test")
+
+    def _fake_post(
+        url: str,
+        headers: dict,
+        json: dict,
+        timeout: object,
+    ) -> MagicMock:
+        assert "deepseek" in url
+        assert "/chat/completions" in url
+        assert json["messages"][0]["role"] == "system"
+        assert headers.get("Authorization", "").startswith("Bearer sk-ds-")
+        r = MagicMock()
+        r.json.return_value = {"choices": [{"message": {"content": "ds_out"}}]}
+        r.raise_for_status = MagicMock()
+        return r
+
+    with patch(
+        "pipeline.llm.providers.adapters.deepseek_text.requests.post",
+        side_effect=_fake_post,
+    ):
+        llm = DeepSeekTextLlm()
+        out = llm.complete_text("S", "U", temperature=0.1)
+    assert out == "ds_out"
+
+
+def test_factory_selects_deepseek_when_env_set(monkeypatch: pytest.MonkeyPatch) -> None:
+    reset_text_llm_client_for_tests()
+    monkeypatch.setenv("MA_LLM_TEXT_PROVIDER", "deepseek")
+    monkeypatch.setenv("DEEPSEEK_API_KEY", "sk-x")
+    try:
+        c = get_text_llm()
+        assert isinstance(c, DeepSeekTextLlm)
+    finally:
+        reset_text_llm_client_for_tests()
+        monkeypatch.delenv("MA_LLM_TEXT_PROVIDER", raising=False)
+        monkeypatch.delenv("DEEPSEEK_API_KEY", raising=False)
+
+
 def test_factory_selects_kimi_when_env_set(monkeypatch: pytest.MonkeyPatch) -> None:
     reset_text_llm_client_for_tests()
     monkeypatch.setenv("MA_LLM_TEXT_PROVIDER", "kimi")