diff --git a/.env.example b/.env.example index 4ec4661..8943d9d 100644 --- a/.env.example +++ b/.env.example @@ -29,9 +29,19 @@ 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 -# 报告/策略等「纯文本」任务的后端选择（默认经 crawler 副本内 AI_crawler 调 OpenAI 兼容网关；其它取值待扩展） +# 报告/策略等「纯文本」任务的后端选择： +# crawler_openai_compatible（默认）= 经 crawler 副本内 AI_crawler 调自建 OpenAI 兼容网关 +# openai_official / openai_chatgpt / chatgpt = 直连 OpenAI 官方 api.openai.com（与上项凭据独立） # MA_LLM_TEXT_PROVIDER=crawler_openai_compatible +# OpenAI 官方 Chat Completions（仅当 MA_LLM_TEXT_PROVIDER 为 openai_official / openai_chatgpt / chatgpt 时需要） +# OPENAI_OFFICIAL_API_KEY=sk-... +# OPENAI_OFFICIAL_BASE_URL=https://api.openai.com/v1 +# OPENAI_OFFICIAL_TEXT_MODEL=gpt-4o-mini +# 与预检一致；gpt-4o 等可设 128000 +# OPENAI_OFFICIAL_CONTEXT_WINDOW=128000 +# OPENAI_OFFICIAL_TIMEOUT=600 + # 独立策略稿 / 报告内「策略与机会」LLM 块采样温度（默认 0.1，低于 chat_completion_text 的 0.2，减轻同提示多轮漂移）。设 0 更稳、设 0.2 与旧默认接近。 # MA_STRATEGY_LLM_TEMPERATURE=0.1 diff --git a/backend/pipeline/demos/try_openai_official_llm.py b/backend/pipeline/demos/try_openai_official_llm.py new file mode 100644 index 0000000..11a1904 --- /dev/null +++ b/backend/pipeline/demos/try_openai_official_llm.py @@ -0,0 +1,27 @@ +""" +试跑 OpenAI 官方「ChatGPT」文本适配器（不经过完整报告管线）。 + +准备：在 .env 中设置 MA_LLM_TEXT_PROVIDER=chatgpt 与 OPENAI_OFFICIAL_API_KEY（见 .env.example）。 + +用法（在 backend 目录下）： + + .venv\\Scripts\\python.exe -m pipeline.demos.try_openai_official_llm +""" +from __future__ import annotations + +import os +import sys + +if __name__ == "__main__": + if not (os.environ.get("OPENAI_OFFICIAL_API_KEY") or "").strip(): + print("未设置 OPENAI_OFFICIAL_API_KEY，退出。", file=sys.stderr) + sys.exit(1) + # 与生产一致时可在 .env 中设 MA_LLM_TEXT_PROVIDER=chatgpt；本脚本也强制用官方适配器 + from pipeline.llm.providers.adapters.openai_official_chatgpt import OpenAiOfficialChatGptTextLlm + + out = OpenAiOfficialChatGptTextLlm().complete_text( + "You reply in one short English sentence only.", + "What is 2+2?", + temperature=0.0, + ) + print(out) diff --git a/backend/pipeline/llm/llm_client.py b/backend/pipeline/llm/llm_client.py index 88b18a4..dbbeae6 100644 --- a/backend/pipeline/llm/llm_client.py +++ b/backend/pipeline/llm/llm_client.py @@ -2,7 +2,7 @@ from __future__ import annotations from .providers.factory import get_text_llm -from .providers.output_normalize import strip_outer_markdown_fence +from .providers.shared.output_normalize import strip_outer_markdown_fence def call_llm( diff --git a/backend/pipeline/llm/providers/__init__.py b/backend/pipeline/llm/providers/__init__.py index e8032da..7eebc7f 100644 --- a/backend/pipeline/llm/providers/__init__.py +++ b/backend/pipeline/llm/providers/__init__.py @@ -1,10 +1,14 @@ """文本大模型调用的协议、适配器与工厂（与具体提示词/业务生成逻辑解耦）。""" + from __future__ import annotations +from .adapters import CrawlerOpenAiCompatibleTextLlm, OpenAiOfficialChatGptTextLlm from .factory import get_text_llm, reset_text_llm_client_for_tests from .protocol import TextLlmClient __all__ = [ + "CrawlerOpenAiCompatibleTextLlm", + "OpenAiOfficialChatGptTextLlm", "TextLlmClient", "get_text_llm", "reset_text_llm_client_for_tests", diff --git a/backend/pipeline/llm/providers/adapters/__init__.py b/backend/pipeline/llm/providers/adapters/__init__.py new file mode 100644 index 0000000..9f0d948 --- /dev/null +++ b/backend/pipeline/llm/providers/adapters/__init__.py @@ -0,0 +1,11 @@ +"""具体大模型通道实现：经统一协议暴露给 `factory`。""" +from __future__ import annotations + +from .crawler_openai_compatible import CrawlerOpenAiCompatibleTextLlm, ensure_ai_crawler_path +from .openai_official_chatgpt import OpenAiOfficialChatGptTextLlm + +__all__ = [ + "CrawlerOpenAiCompatibleTextLlm", + "OpenAiOfficialChatGptTextLlm", + "ensure_ai_crawler_path", +] diff --git a/backend/pipeline/llm/providers/crawler_openai_compatible.py b/backend/pipeline/llm/providers/adapters/crawler_openai_compatible.py similarity index 96% rename from backend/pipeline/llm/providers/crawler_openai_compatible.py rename to backend/pipeline/llm/providers/adapters/crawler_openai_compatible.py index ac47146..64c9a7f 100644 --- a/backend/pipeline/llm/providers/crawler_openai_compatible.py +++ b/backend/pipeline/llm/providers/adapters/crawler_openai_compatible.py @@ -9,7 +9,7 @@ from pathlib import Path from django.conf import settings -from .token_heuristics import estimate_crawler_style_input_tokens +from ..shared.token_heuristics import estimate_crawler_style_input_tokens def ensure_ai_crawler_path() -> None: diff --git a/backend/pipeline/llm/providers/adapters/openai_official_chatgpt.py b/backend/pipeline/llm/providers/adapters/openai_official_chatgpt.py new file mode 100644 index 0000000..ca938b1 --- /dev/null +++ b/backend/pipeline/llm/providers/adapters/openai_official_chatgpt.py @@ -0,0 +1,133 @@ +""" +OpenAI 官方 `https://api.openai.com`（ChatGPT 系列）`chat/completions` 直连接口。 + +**凭据与网关与爬虫副本中的自建网关独立**，避免与 `OPENAI_BASE_URL` 指向的兼容网关共用时互相串环境。 +通过 `MA_LLM_TEXT_PROVIDER=openai_official`（或 `openai_chatgpt` / `chatgpt`）启用。 +""" +from __future__ import annotations + +import os +from typing import Any + +import requests + +from ..shared.openai_message_content import normalize_message_content +from ..shared.token_heuristics import estimate_crawler_style_input_tokens + +_DEFAULT_BASE = "https://api.openai.com/v1" +_DEFAULT_MODEL = "gpt-4o-mini" +# gpt-4o / 4.1 等常见上限；可按模型在 .env 中覆盖 +_DEFAULT_CTX = 128_000 + +_BUF = 256 +_WANT_MAX = 8192 + + +def _read_timeout() -> tuple[float, float]: + read = 600 + raw = (os.environ.get("OPENAI_OFFICIAL_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_credentials() -> tuple[str, str, str]: + key = (os.environ.get("OPENAI_OFFICIAL_API_KEY") or "").strip() + if not key: + msg = "使用 openai_official 适配器需设置环境变量 OPENAI_OFFICIAL_API_KEY（与自建网关/爬虫副本的 key 可分开）。" + raise ValueError(msg) + base = (os.environ.get("OPENAI_OFFICIAL_BASE_URL") or _DEFAULT_BASE).strip().rstrip("/") + model = ( + os.environ.get("OPENAI_OFFICIAL_TEXT_MODEL") + or os.environ.get("OPENAI_OFFICIAL_MODEL") + or _DEFAULT_MODEL + ).strip() + return key, base, model + + +def _context_window() -> int: + raw = (os.environ.get("OPENAI_OFFICIAL_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 OpenAiOfficialChatGptTextLlm: + """ + 直连 OpenAI 官方「Chat Completions」；请求体与 `AI_crawler.chat_completion_text` 同形， + 并在本地做与爬虫网关一致的 `max_tokens` 收紧，减少 400。 + """ + + def complete_text( + self, + system_prompt: str, + user_prompt: str, + *, + temperature: float | None = None, + ) -> str: + api_key, base, model = _resolve_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，OPENAI_OFFICIAL_CONTEXT_WINDOW={context_window}），" + "请缩小输入或调大 OPENAI_OFFICIAL_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() diff --git a/backend/pipeline/llm/providers/factory.py b/backend/pipeline/llm/providers/factory.py index 42a7119..533d419 100644 --- a/backend/pipeline/llm/providers/factory.py +++ b/backend/pipeline/llm/providers/factory.py @@ -5,7 +5,8 @@ from __future__ import annotations import os -from .crawler_openai_compatible import CrawlerOpenAiCompatibleTextLlm +from .adapters.crawler_openai_compatible import CrawlerOpenAiCompatibleTextLlm +from .adapters.openai_official_chatgpt import OpenAiOfficialChatGptTextLlm from .protocol import TextLlmClient # 模块级单例：避免重复构造；测试可用 `reset_text_llm_client_for_tests` 切换实现。 @@ -29,9 +30,14 @@ def _build_client(pid: str) -> TextLlmClient: "openai_compatible", ): return CrawlerOpenAiCompatibleTextLlm() + if pid in ("openai_official", "openai_chatgpt", "chatgpt"): + return OpenAiOfficialChatGptTextLlm() + known = ( + "crawler_openai_compatible, crawler, default, openai_compatible, " + "openai_official, openai_chatgpt, chatgpt" + ) raise ValueError( - f"不支持的 {_ENV_KEY}={pid!r}；" - f"当前仅实现 {_DEFAULT_ID}（经 AI_crawler 的 OpenAI 兼容 `chat/completions`）。" + f"不支持的 {_ENV_KEY}={pid!r}；已知取值：{known}。", ) diff --git a/backend/pipeline/llm/providers/shared/__init__.py b/backend/pipeline/llm/providers/shared/__init__.py new file mode 100644 index 0000000..9e132ea --- /dev/null +++ b/backend/pipeline/llm/providers/shared/__init__.py @@ -0,0 +1,12 @@ +"""与具体后端无关的轻量工具：去围栏、token 启发式、OpenAI 风格 message 正文解析。""" +from __future__ import annotations + +from .openai_message_content import normalize_message_content +from .output_normalize import strip_outer_markdown_fence +from .token_heuristics import estimate_crawler_style_input_tokens + +__all__ = [ + "estimate_crawler_style_input_tokens", + "normalize_message_content", + "strip_outer_markdown_fence", +] diff --git a/backend/pipeline/llm/providers/shared/openai_message_content.py b/backend/pipeline/llm/providers/shared/openai_message_content.py new file mode 100644 index 0000000..712c7cb --- /dev/null +++ b/backend/pipeline/llm/providers/shared/openai_message_content.py @@ -0,0 +1,26 @@ +""" +解析 OpenAI chat.completions 返回的 `message.content`：可能是 str 或 part 列表。 +与 `AI_crawler._normalize_chat_content` 行为一致。 +""" +from __future__ import annotations + +from typing import Any + + +def normalize_message_content(content: Any) -> str: + if content is None: + return "" + if isinstance(content, str): + return content.strip() + if isinstance(content, list): + parts: list[str] = [] + for item in content: + if isinstance(item, dict): + if item.get("type") == "text": + parts.append(str(item.get("text") or "")) + elif "text" in item: + parts.append(str(item.get("text") or "")) + elif isinstance(item, str): + parts.append(item) + return "".join(parts).strip() + return str(content).strip() diff --git a/backend/pipeline/llm/providers/output_normalize.py b/backend/pipeline/llm/providers/shared/output_normalize.py similarity index 100% rename from backend/pipeline/llm/providers/output_normalize.py rename to backend/pipeline/llm/providers/shared/output_normalize.py diff --git a/backend/pipeline/llm/providers/token_heuristics.py b/backend/pipeline/llm/providers/shared/token_heuristics.py similarity index 76% rename from backend/pipeline/llm/providers/token_heuristics.py rename to backend/pipeline/llm/providers/shared/token_heuristics.py index 302fe85..997d3e4 100644 --- a/backend/pipeline/llm/providers/token_heuristics.py +++ b/backend/pipeline/llm/providers/shared/token_heuristics.py @@ -1,6 +1,6 @@ """ 与 `crawler_copy/.../AI_crawler` 中 `_estimate_chat_input_tokens` 同口径的保守估算， -供预检、策略档位与 OpenAI 兼容适配器共用（无 tiktoken 时避免 max_tokens 400）。 +供预检、策略档位与各适配器共用（无 tiktoken 时避免 max_tokens 400）。 """ from __future__ import annotations diff --git a/backend/pipeline/tests/test_text_llm_providers.py b/backend/pipeline/tests/test_text_llm_providers.py new file mode 100644 index 0000000..c36b008 --- /dev/null +++ b/backend/pipeline/tests/test_text_llm_providers.py @@ -0,0 +1,70 @@ +from __future__ import annotations + +from unittest.mock import MagicMock, patch + +import pytest + +from pipeline.llm.providers import ( + CrawlerOpenAiCompatibleTextLlm, + OpenAiOfficialChatGptTextLlm, + get_text_llm, + reset_text_llm_client_for_tests, +) + + +def test_openai_official_complete_text_uses_post_and_returns_content(monkeypatch: pytest.MonkeyPatch) -> None: + monkeypatch.setenv("OPENAI_OFFICIAL_API_KEY", "sk-test") + + def _fake_post( + url: str, + headers: dict, + json: dict, + timeout: object, + ) -> MagicMock: + assert "/chat/completions" in url + assert json["messages"][0]["role"] == "system" + r = MagicMock() + r.json.return_value = {"choices": [{"message": {"content": "ok_out"}}]} + r.raise_for_status = MagicMock() + return r + + with patch( + "pipeline.llm.providers.adapters.openai_official_chatgpt.requests.post", + side_effect=_fake_post, + ): + llm = OpenAiOfficialChatGptTextLlm() + out = llm.complete_text("S", "U", temperature=0.1) + assert out == "ok_out" + + +def test_factory_selects_openai_official_when_env_set(monkeypatch: pytest.MonkeyPatch) -> None: + reset_text_llm_client_for_tests() + monkeypatch.setenv("MA_LLM_TEXT_PROVIDER", "chatgpt") + monkeypatch.setenv("OPENAI_OFFICIAL_API_KEY", "sk-x") + try: + c = get_text_llm() + assert isinstance(c, OpenAiOfficialChatGptTextLlm) + finally: + reset_text_llm_client_for_tests() + monkeypatch.delenv("MA_LLM_TEXT_PROVIDER", raising=False) + monkeypatch.delenv("OPENAI_OFFICIAL_API_KEY", raising=False) + + +def test_factory_unknown_provider_raises(monkeypatch: pytest.MonkeyPatch) -> None: + reset_text_llm_client_for_tests() + monkeypatch.setenv("MA_LLM_TEXT_PROVIDER", "no_such_provider") + try: + with pytest.raises(ValueError, match="不支持的"): + get_text_llm() + finally: + reset_text_llm_client_for_tests() + monkeypatch.delenv("MA_LLM_TEXT_PROVIDER", raising=False) + + +def test_default_provider_is_crawler_when_env_cleared(monkeypatch: pytest.MonkeyPatch) -> None: + """未设置 MA_LLM_TEXT_PROVIDER 时仍为爬虫副本网关；此处只断言类型，不调真实网络。""" + reset_text_llm_client_for_tests() + monkeypatch.delenv("MA_LLM_TEXT_PROVIDER", raising=False) + c = get_text_llm() + assert isinstance(c, CrawlerOpenAiCompatibleTextLlm) + reset_text_llm_client_for_tests()