code_pipeline/data_construction/utils/llm_interface.py

from __future__ import annotations
"""
统一 LLM 调用模块
================
仿照 mify_client.py 的调用方式，通过 Mify API 代理调用闭源模型。

支持模型:
  - gpt-5.5                      (Azure OpenAI)
  - gemini-3.1-pro-preview-pt    (Vertex AI)

所有数据构造脚本通过 get_llm() 获取实例，通过 .generate() / .generate_json() 调用。
"""
import json
import time
import logging
import os
from abc import ABC, abstractmethod
from typing import Optional

from openai import OpenAI

try:
    import anthropic
    _ANTHROPIC_AVAILABLE = True
except ImportError:
    anthropic = None
    _ANTHROPIC_AVAILABLE = False

logger = logging.getLogger(__name__)

# ==================== Mify API 配置 ====================
MIFY_API_KEY = os.getenv("MIFY_API_KEY", "")
MIFY_BASE_URL = os.getenv("MIFY_BASE_URL", "https://api.llm.mioffice.cn/v1")
# Claude 模型走单独的 Anthropic-compatible endpoint
MIFY_ANTHROPIC_BASE_URL = os.getenv("MIFY_ANTHROPIC_BASE_URL", "https://api.llm.mioffice.cn/anthropic")

# 模型名称 -> Provider ID 映射 (OpenAI-compatible models 才需要)
# 注: 不在此表的模型不发 provider header (如 ppio/pa/gpt-5.5, 名字自带路由, 带 header 会 400).
MODEL_PROVIDER = {
    "gpt-5.4": "azure_openai",
    "gemini-3.1-pro-preview-pt": "vertex_ai",
    "gemini-3-pro-preview-pt": "vertex_ai",
    "glm-5.2": "zhipuai",
}

# Claude 系列走 Anthropic SDK + Mify /anthropic endpoint
ANTHROPIC_MODELS = {
    "ppio/pa/claude-opus-4-7",
    "ppio/pa/claude-sonnet-4-6",
    "ppio/pa/claude-haiku-4-5",
}

# Reasoning / 受限模型: 不接受 temperature 参数 (会返回 400)
# - gpt-5.x (azure_openai 路由) 系列 是 reasoning 模型, 拒绝 temperature
# - claude-opus-4-7 (deprecated temperature)
# 注: ppio/pa/gpt-5.5 实测接受 temperature, 故不在此名单.
MODELS_NO_TEMPERATURE = {
    "gpt-5.4",
    "gpt-5.2",
    "o3",
    "ppio/pa/claude-opus-4-7",
}

# 模型的上下文窗口大小
MODEL_MAX_LENGTH = {
    "ppio/pa/gpt-5.5": 272000,
    "gemini-3.1-pro-preview-pt": 900000,
    "ppio/pa/claude-opus-4-7": 200000,
    "ppio/pa/claude-sonnet-4-6": 200000,
    "glm-5.2": 200000,
}

# 模型简短别名（文件命名用）
MODEL_ALIAS = {
    "ppio/pa/gpt-5.5": "gpt5.5",
    "gemini-3.1-pro-preview-pt": "gemini3.1pro",
    "ppio/pa/claude-opus-4-7": "claude-opus-4.7",
    "ppio/pa/claude-sonnet-4-6": "claude-sonnet-4.6",
    "glm-5.2": "glm5.2",
}


# ==================== Content-filter 检测 ====================
class ContentFilterError(Exception):
    """模型 backend 拒绝该 prompt (Azure content filter / Anthropic safety / etc).
    Deterministic, 不重试 — 由上层 fallback chain 接管."""


_CONTENT_FILTER_PATTERNS = (
    "content management policy",       # Azure OpenAI
    "response was filtered",           # Azure OpenAI
    "ResponsibleAIPolicyViolation",    # Azure
    "content filtering",               # Generic
    "safety filter",                   # Gemini
    "safety_violation",                # Gemini
    "blocked by safety",               # Anthropic
    "content_policy_violation",        # OpenAI moderation
)


def _is_content_filter_error(msg: str) -> bool:
    if not msg:
        return False
    low = msg.lower()
    return any(p.lower() in low for p in _CONTENT_FILTER_PATTERNS)


# ==================== 抽象接口 ====================
class LLMInterface(ABC):
    @abstractmethod
    def generate(
        self,
        messages: list[dict],
        n: int = 1,
        temperature: float = 0.7,
        max_tokens: int = 2048,
    ) -> list[str]:
        ...

    def generate_json(
        self,
        messages: list[dict],
        temperature: float = 0.7,
        max_tokens: int = 2048,
        retries: int = 3,
    ) -> dict:
        for attempt in range(retries):
            results = self.generate(
                messages, n=1, temperature=temperature, max_tokens=max_tokens
            )
            text = results[0].strip()
            # 清洗 markdown code block
            if text.startswith("```json"):
                text = text[7:]
            if text.startswith("```"):
                text = text[3:]
            if text.endswith("```"):
                text = text[:-3]
            text = text.strip()
            try:
                return json.loads(text)
            except json.JSONDecodeError:
                logger.warning(
                    f"JSON parse failed (attempt {attempt + 1}/{retries}): {text[:200]}"
                )
                if attempt < retries - 1:
                    time.sleep(1)
        raise ValueError(f"Failed to parse JSON after {retries} attempts")


# ==================== Mify API 调用实现 ====================
class MifyLLM(LLMInterface):
    """通过 Mify API 代理调用闭源模型 (GPT / Gemini / Claude 等).

    自动按模型名分发:
      - Claude 系列 (ANTHROPIC_MODELS): Anthropic SDK + /anthropic endpoint
      - 其他: OpenAI SDK + /v1 endpoint
    """

    def __init__(
        self,
        model: str,
        api_key: str = MIFY_API_KEY,
        base_url: str = MIFY_BASE_URL,
        anthropic_base_url: str = MIFY_ANTHROPIC_BASE_URL,
        max_retries: int = 3,
        retry_delay: int = 5,
        request_interval: float = 1.0,
    ):
        self.model = model
        self.api_key = api_key
        self.base_url = base_url
        self.anthropic_base_url = anthropic_base_url
        self.max_retries = max_retries
        self.retry_delay = retry_delay
        self.request_interval = request_interval

        self.is_anthropic = model in ANTHROPIC_MODELS
        if self.is_anthropic:
            if not _ANTHROPIC_AVAILABLE:
                raise RuntimeError(
                    f"Model {model} requires the 'anthropic' package. Install via `pip install anthropic`."
                )
            self.anthropic_client = anthropic.Anthropic(
                api_key=api_key, base_url=anthropic_base_url,
            )
            self.client = None
        else:
            # 只有在 MODEL_PROVIDER 显式配置时才发 provider header.
            # 名字自带路由的模型 (如 ppio/pa/gpt-5.5) 不能带 header, 否则网关 400.
            provider_id = MODEL_PROVIDER.get(model)
            headers = {"X-Model-Provider-Id": provider_id} if provider_id else {}
            self.client = OpenAI(
                api_key=api_key,
                base_url=base_url,
                default_headers=headers,
            )
            self.anthropic_client = None

    def _split_system(self, messages: list[dict]) -> tuple[Optional[str], list[dict]]:
        """Anthropic SDK 要求 system 走单独 kwarg, 这里把 system 消息抽出来."""
        system_chunks: list[str] = []
        rest: list[dict] = []
        for m in messages:
            if m.get("role") == "system":
                system_chunks.append(m.get("content", ""))
            else:
                rest.append(m)
        system = "\n".join(s for s in system_chunks if s) or None
        return system, rest

    def _call_once_openai(self, messages: list[dict], temperature: float, max_tokens: int) -> str | None:
        kwargs = {
            "model": self.model,
            "messages": messages,
            "max_completion_tokens": max_tokens,
            "stream": False,
        }
        if self.model not in MODELS_NO_TEMPERATURE:
            kwargs["temperature"] = temperature
        for attempt in range(self.max_retries):
            try:
                response = self.client.chat.completions.create(**kwargs)
                if response and response.choices:
                    content = response.choices[0].message.content
                    if content is not None:
                        return content
                return None
            except Exception as e:
                msg = str(e)
                # Content filter 是 deterministic 的, 重试无意义, 直接 raise 让 fallback 接管
                if _is_content_filter_error(msg):
                    logger.warning(
                        f"  Content-filter rejected by {self.model}, abort retries: {msg[:200]}"
                    )
                    raise ContentFilterError(msg) from e
                logger.warning(
                    f"  [Retry {attempt + 1}/{self.max_retries}] {type(e).__name__}: {msg[:200]}"
                )
                if attempt < self.max_retries - 1:
                    time.sleep(self.retry_delay * (attempt + 1))
        return None

    def _call_once_anthropic(self, messages: list[dict], temperature: float, max_tokens: int) -> str | None:
        system, rest = self._split_system(messages)
        kwargs = {
            "model": self.model,
            "messages": rest,
            "max_tokens": max_tokens,
        }
        if self.model not in MODELS_NO_TEMPERATURE:
            kwargs["temperature"] = temperature
        if system:
            kwargs["system"] = system
        for attempt in range(self.max_retries):
            try:
                response = self.anthropic_client.messages.create(**kwargs)
                if response.content:
                    parts = [c.text for c in response.content if getattr(c, "type", None) == "text"]
                    if parts:
                        return "\n".join(parts)
                return None
            except Exception as e:
                msg = str(e)
                if _is_content_filter_error(msg):
                    logger.warning(
                        f"  Content-filter rejected by {self.model}, abort retries: {msg[:200]}"
                    )
                    raise ContentFilterError(msg) from e
                logger.warning(
                    f"  [Retry {attempt + 1}/{self.max_retries}] {type(e).__name__}: {msg[:200]}"
                )
                if attempt < self.max_retries - 1:
                    time.sleep(self.retry_delay * (attempt + 1))
        return None

    def _call_once(self, messages: list[dict], temperature: float, max_tokens: int) -> str | None:
        """单次 API 调用, 按模型分发到 OpenAI 或 Anthropic backend."""
        if self.is_anthropic:
            return self._call_once_anthropic(messages, temperature, max_tokens)
        return self._call_once_openai(messages, temperature, max_tokens)

    def generate(
        self,
        messages: list[dict],
        n: int = 1,
        temperature: float = 0.7,
        max_tokens: int = 2048,
    ) -> list[str]:
        results = []
        for i in range(n):
            result = self._call_once(messages, temperature, max_tokens)
            results.append(result or "")
            if i < n - 1 and self.request_interval > 0:
                time.sleep(self.request_interval)
        return results


# ==================== 直接 OpenAI API (非 Mify) ====================
class DirectOpenAILLM(LLMInterface):
    """直接调用 OpenAI 官方 API (非代理)。"""

    def __init__(self, model: str, api_key: Optional[str] = None, base_url: Optional[str] = None):
        kwargs = {}
        if api_key:
            kwargs["api_key"] = api_key
        if base_url:
            kwargs["base_url"] = base_url
        self.client = OpenAI(**kwargs)
        self.model = model

    def generate(
        self,
        messages: list[dict],
        n: int = 1,
        temperature: float = 0.7,
        max_tokens: int = 2048,
    ) -> list[str]:
        response = self.client.chat.completions.create(
            model=self.model,
            messages=messages,
            n=n,
            temperature=temperature,
            max_tokens=max_tokens,
        )
        return [choice.message.content or "" for choice in response.choices]


# ==================== vLLM (本地部署) ====================
class VLLMLLM(LLMInterface):
    """通过 vLLM 的 OpenAI-compatible API 调用本地模型。"""

    def __init__(self, model: str, base_url: str, api_key: str = "EMPTY"):
        self.client = OpenAI(base_url=base_url, api_key=api_key)
        self.model = model

    def generate(
        self,
        messages: list[dict],
        n: int = 1,
        temperature: float = 0.7,
        max_tokens: int = 2048,
    ) -> list[str]:
        response = self.client.chat.completions.create(
            model=self.model,
            messages=messages,
            n=n,
            temperature=temperature,
            max_tokens=max_tokens,
        )
        return [choice.message.content or "" for choice in response.choices]


# ==================== Fallback 链 ====================
class FallbackLLM(LLMInterface):
    """串行尝试多个 backend, 任一成功就返回. 用于绕过 Azure content filter
    等 deterministic 拒绝: gpt-5.5 命中 ContentFilterError -> gemini -> claude.

    顺序按构造时传入的 backends 排序. 只有 primary 失败才付出 fallback 成本.
    """

    def __init__(self, backends: list[LLMInterface]):
        if not backends:
            raise ValueError("FallbackLLM requires at least one backend")
        self.backends = backends

    def generate(
        self,
        messages: list[dict],
        n: int = 1,
        temperature: float = 0.7,
        max_tokens: int = 2048,
    ) -> list[str]:
        last_error: Optional[Exception] = None
        for i, backend in enumerate(self.backends):
            tag = getattr(backend, "model", type(backend).__name__)
            try:
                results = backend.generate(messages, n=n, temperature=temperature, max_tokens=max_tokens)
                # 全空才算失败 (单个空字符串视为这次没拿到内容, 但若有 n>1 部分成功仍算 OK)
                if results and any(r for r in results):
                    if i > 0:
                        logger.info(f"FallbackLLM: succeeded via fallback backend [{i}] {tag}")
                    return results
                logger.warning(f"FallbackLLM: backend [{i}] {tag} returned empty, trying next")
            except ContentFilterError as e:
                last_error = e
                logger.info(f"FallbackLLM: backend [{i}] {tag} hit content filter, trying next")
            except Exception as e:
                last_error = e
                logger.warning(f"FallbackLLM: backend [{i}] {tag} raised {type(e).__name__}: {str(e)[:200]}")
        # 全部 backend 都失败
        if last_error is not None:
            raise last_error
        return [""] * n


# ==================== 工厂函数 ====================
def get_llm(
    provider: str,
    model: str,
    api_key: Optional[str] = None,
    base_url: Optional[str] = None,
) -> LLMInterface:
    """
    根据 provider 创建 LLM 实例.

    Args:
        provider: "mify" | "openai" | "vllm"
            - mify: 通过 Mify 代理调用 GPT/Gemini/Claude (默认推荐)
            - openai: 直接调用 OpenAI 官方 API
            - vllm: 调用本地 vLLM 服务
        model: 模型名 (如 "gpt-5.5"). 支持 fallback 链, 用逗号分隔多个模型,
               primary 命中 content filter 时自动顺移到下一个,
               例如 "gpt-5.5,gemini-3.1-pro-preview-pt".
        api_key: API 密钥 (mify 模式下可不传, 使用默认)
        base_url: API 地址 (vllm 模式必传)
    """
    # 解析逗号分隔的 fallback 链
    model_names = [m.strip() for m in model.split(",") if m.strip()]
    if not model_names:
        raise ValueError(f"empty model spec: {model!r}")

    def make_one(name: str) -> LLMInterface:
        if provider == "mify":
            return MifyLLM(
                model=name,
                api_key=api_key or MIFY_API_KEY,
                base_url=base_url or MIFY_BASE_URL,
            )
        elif provider == "openai":
            return DirectOpenAILLM(model=name, api_key=api_key, base_url=base_url)
        elif provider == "vllm":
            if not base_url:
                raise ValueError("base_url required for vllm provider")
            return VLLMLLM(model=name, base_url=base_url, api_key=api_key or "EMPTY")
        else:
            raise ValueError(f"Unknown provider: {provider}. Use 'mify', 'openai', or 'vllm'.")

    if len(model_names) == 1:
        return make_one(model_names[0])
    backends = [make_one(n) for n in model_names]
    logger.info(f"get_llm: built FallbackLLM chain {model_names}")
    return FallbackLLM(backends)


# ==================== 便捷函数 (兼容 mify_client.py 风格) ====================
def _split_system_for_anthropic(messages: list[dict]) -> tuple[Optional[str], list[dict]]:
    system_chunks: list[str] = []
    rest: list[dict] = []
    for m in messages:
        if m.get("role") == "system":
            system_chunks.append(m.get("content", ""))
        else:
            rest.append(m)
    system = "\n".join(s for s in system_chunks if s) or None
    return system, rest


def _call_anthropic_direct(
    model_name: str,
    messages: list[dict],
    max_tokens: int,
    temperature: float,
    max_retries: int,
    retry_delay: int,
) -> str | None:
    if not _ANTHROPIC_AVAILABLE:
        raise RuntimeError(
            f"Model {model_name} requires the 'anthropic' package. Install via `pip install anthropic`."
        )
    client = anthropic.Anthropic(api_key=MIFY_API_KEY, base_url=MIFY_ANTHROPIC_BASE_URL)
    system, rest = _split_system_for_anthropic(messages)
    kwargs = {
        "model": model_name,
        "messages": rest,
        "max_tokens": max_tokens,
    }
    if model_name not in MODELS_NO_TEMPERATURE:
        kwargs["temperature"] = temperature
    if system:
        kwargs["system"] = system

    for attempt in range(max_retries):
        try:
            response = client.messages.create(**kwargs)
            if response.content:
                parts = [c.text for c in response.content if getattr(c, "type", None) == "text"]
                if parts:
                    return "\n".join(parts)
            return None
        except Exception as e:
            logger.warning(
                f"  [Retry {attempt + 1}/{max_retries}] {type(e).__name__}: {str(e)[:200]}"
            )
            if attempt < max_retries - 1:
                import random
                # 指数退避 + 随机抖动，避免多 worker 同时重试撞车
                backoff = retry_delay * (2 ** attempt) + random.uniform(0, 2)
                time.sleep(backoff)
    return None


def call_model(
    model_name: str,
    messages: list[dict],
    max_tokens: int = 4096,
    temperature: float = 0,
    max_retries: int = 3,
    retry_delay: int = 5,
) -> str | None:
    """
    兼容 mify_client.py 的 call_model() 接口.
    自动按模型分发: Claude 走 Anthropic SDK + /anthropic, 其他走 OpenAI SDK + /v1.
    """
    if model_name in ANTHROPIC_MODELS:
        return _call_anthropic_direct(
            model_name, messages, max_tokens, temperature, max_retries, retry_delay,
        )

    # 只有在 MODEL_PROVIDER 显式配置时才发 provider header.
    # 名字自带路由的模型 (如 ppio/pa/gpt-5.5) 不能带 header, 否则网关 400.
    provider_id = MODEL_PROVIDER.get(model_name)
    headers = {"X-Model-Provider-Id": provider_id} if provider_id else {}
    client = OpenAI(
        api_key=MIFY_API_KEY,
        base_url=MIFY_BASE_URL,
        default_headers=headers,
    )

    kwargs = {
        "model": model_name,
        "messages": messages,
        "max_completion_tokens": max_tokens,
        "stream": False,
    }
    if model_name not in MODELS_NO_TEMPERATURE:
        kwargs["temperature"] = temperature

    for attempt in range(max_retries):
        try:
            response = client.chat.completions.create(**kwargs)
            if response and response.choices:
                content = response.choices[0].message.content
                if content is not None:
                    return content
            return None
        except Exception as e:
            logger.warning(
                f"  [Retry {attempt + 1}/{max_retries}] {type(e).__name__}: {str(e)[:200]}"
            )
            if attempt < max_retries - 1:
                import random
                backoff = retry_delay * (2 ** attempt) + random.uniform(0, 2)
                time.sleep(backoff)

    return None


def get_model_alias(model_name: str) -> str:
    return MODEL_ALIAS.get(model_name, model_name.replace("/", "_").replace(":", "_"))


def get_model_max_length(model_name: str) -> int:
    return MODEL_MAX_LENGTH.get(model_name, 32768)