fix(llm): call_llm 429 自动 fallback + 移除 compliance 硬编码 provider

- nvidia_client.py 新增 _get_provider_fallback_list() 从 DB 读提供商顺序 - call_llm 循环尝试提供商列表，HTTP 429 自动跳过到下一个 - 移除 compliance_optimizer.py 中硬编码 provider="opencode-go" - 新增 logging 模块引用
2026-06-01 13:23:35 +08:00
parent 09c467e56d
commit 31d8434010
2 changed files with 77 additions and 60 deletions
@@ -6,9 +6,12 @@ Unified LLM Client
 import os
 import requests
 import json
+import logging
 from typing import Optional, Dict, Any, List
 from pathlib import Path

+logger = logging.getLogger(__name__)
+
 from dotenv import load_dotenv

 env_path = Path(__file__).resolve().parents[2] / ".env"
@@ -88,6 +91,20 @@ def _get_db_defaults(provider: Optional[str] = None) -> dict:
        pass
    return {"temperature": 0.20, "max_tokens": 2048, "system_prompt": "你是一个专业的内容创作助手。"}

+def _get_provider_fallback_list() -> List[str]:
+    """返回提供商尝试顺序列表（活跃优先，其余按配置顺序）"""
+    try:
+        from ..database import SessionLocal
+        from ..models import LLMConfig
+        db = SessionLocal()
+        configs = db.query(LLMConfig).order_by(LLMConfig.is_active.desc(), LLMConfig.id).all()
+        db.close()
+        providers = [c.provider for c in configs if c.provider]
+        return providers
+    except Exception:
+        pass
+    return ["opencode-go", "nvidia"]
+
 def call_llm(
    prompt: str,
    model: Optional[str] = None,
@@ -105,66 +122,66 @@ def call_llm(
    temperature = temperature if temperature is not None else defaults["temperature"]
    max_tokens = max_tokens if max_tokens is not None else defaults["max_tokens"]
    system_prompt = system_prompt if system_prompt is not None else defaults["system_prompt"]
-    cfg = _get_provider_config(provider)
-    endpoint = f"{cfg['base_url'].rstrip('/')}/chat/completions"
-    headers = {
-        "Authorization": f"Bearer {cfg['api_key']}",
-        "Content-Type": "application/json"
-    }
-    payload = {
-        "model": model or cfg["model"],
-        "messages": [
-            {"role": "system", "content": system_prompt},
-            {"role": "user", "content": prompt}
-        ],
-        "temperature": temperature,
-        "max_tokens": max_tokens,
-        "top_p": top_p,
-        "frequency_penalty": frequency_penalty,
-        "presence_penalty": presence_penalty,
-        "stream": stream,
-    }
-    if additional_params:
-        payload.update(additional_params)

-    try:
-        resp = requests.post(endpoint, json=payload, headers=headers, timeout=120, stream=stream)
-        if resp.status_code != 200:
-            raise LLMError(f"HTTP {resp.status_code}: {resp.text[:200]}")
-        if stream:
-            content_parts = []
-            reasoning_parts = []
-            for line in resp.iter_lines():
-                if not line: continue
-                if line.startswith(b'data: '):
-                    data = line[6:]
-                    if data == b'[DONE]': break
-                    try:
-                        chunk = json.loads(data)
-                        delta = chunk['choices'][0]['delta']
-                        if delta.get('content'):
-                            content_parts.append(delta['content'])
-                        if delta.get('reasoning_content'):
-                            reasoning_parts.append(delta['reasoning_content'])
-                    except Exception: continue
-            if content_parts:
-                return "".join(content_parts)
-            return "".join(reasoning_parts)
-        else:
-            data = resp.json()
-            msg = data["choices"][0]["message"]
-            # 优先取 content（推理模型如 deepseek 的最终答案在此字段）
-            # 如果 content 为空但 reasoning_content 有值（说明 max_tokens 不够没输出完），取其末尾作为近似答案
-            content = msg.get('content') or ''
-            if not content.strip():
-                rc = msg.get('reasoning_content', '')
-                if rc:
-                    # 取 reasoning 末尾最可能包含答案的句子
-                    parts = [p.strip() for p in rc.replace('\n', '。').split('。') if p.strip()]
-                    content = parts[-1] if parts else rc
-            return content.strip() if content else ''
-    except requests.RequestException as e:
-        raise LLMError(f"Request failed: {e}")
+    providers_to_try = [provider] if provider else _get_provider_fallback_list()
+    last_error = None
+    for p in providers_to_try:
+        try:
+            cfg = _get_provider_config(p)
+            endpoint = f"{cfg['base_url'].rstrip('/')}/chat/completions"
+            headers = {"Authorization": f"Bearer {cfg['api_key']}", "Content-Type": "application/json"}
+            payload = {
+                "model": model or cfg["model"],
+                "messages": [{"role": "system", "content": system_prompt}, {"role": "user", "content": prompt}],
+                "temperature": temperature, "max_tokens": max_tokens, "top_p": top_p,
+                "frequency_penalty": frequency_penalty, "presence_penalty": presence_penalty,
+                "stream": stream,
+            }
+            if additional_params:
+                payload.update(additional_params)
+
+            resp = requests.post(endpoint, json=payload, headers=headers, timeout=120, stream=stream)
+            if resp.status_code == 429:
+                logger.warning(f"[LLM] {p} 429 配额超限 → 尝试下一个")
+                last_error = LLMError(f"HTTP {resp.status_code}: {resp.text[:200]}")
+                continue
+            if resp.status_code != 200:
+                raise LLMError(f"HTTP {resp.status_code}: {resp.text[:200]}")
+
+            if stream:
+                content_parts, reasoning_parts = [], []
+                for line in resp.iter_lines():
+                    if not line: continue
+                    if line.startswith(b'data: '):
+                        data = line[6:]
+                        if data == b'[DONE]': break
+                        try:
+                            chunk = json.loads(data)
+                            delta = chunk['choices'][0]['delta']
+                            if delta.get('content'): content_parts.append(delta['content'])
+                            if delta.get('reasoning_content'): reasoning_parts.append(delta['reasoning_content'])
+                        except Exception: continue
+                return "".join(content_parts) or "".join(reasoning_parts)
+            else:
+                data = resp.json()
+                msg = data["choices"][0]["message"]
+                content = msg.get('content') or ''
+                if not content.strip():
+                    rc = msg.get('reasoning_content', '')
+                    if rc:
+                        parts = [p.strip() for p in rc.replace('\n', '。').split('。') if p.strip()]
+                        content = parts[-1] if parts else rc
+                return content.strip() or ''
+        except LLMError as e:
+            if '429' in str(e):
+                logger.warning(f"[LLM] {p} 429 配额超限 → 尝试下一个")
+                last_error = e; continue
+            raise
+        except Exception as e:
+            logger.warning(f"[LLM] {p} 调用异常: {e} → 尝试下一个")
+            last_error = e; continue
+
+    raise last_error or LLMError("所有 LLM 提供商均不可用")

 def expand_content_with_llm(
    topic: dict,
@@ -182,7 +182,7 @@ def polish_with_llm(html: str, platform: str, remaining_issues: Optional[List[Di
                prompt = get_prompt("compliance_fix", issues_desc=issues_desc, html=html)
            else:
                prompt = get_prompt("compliance_polish", html=html)
-            polished = call_llm(prompt, provider="opencode-go", temperature=temperature, max_tokens=max_tokens, system_prompt=system_prompt)
+            polished = call_llm(prompt, provider=None, temperature=temperature, max_tokens=max_tokens, system_prompt=system_prompt)
            polished = clean_html_content(polished)
            polished = strip_ai_preface(polished)
            polished = strip_thinking_html(polished)