6953a27937
- _resolve_cfg_with_govern: 机构有策略→门禁链(限流/限额/主备容错/端点选择/预授权)； 无策略→__LEGACY__走旧llm表(向后兼容)；治理真实失败抛错禁止静默回退 - llm_call/llm_call_msgs/llm_call_msgs_native 三处接入；失败结算释放预授权写failed流水 - proxy_chat_completion(运行环境token路径)同样接入，按上游真实usage结算
371 lines
15 KiB
Python
371 lines
15 KiB
Python
"""
|
||
LLM bridge for pipeline handlers.
|
||
|
||
Provides a simple async interface for handlers to call LLM APIs.
|
||
Looks up model config from the llm database table first,
|
||
falls back to environment variables.
|
||
"""
|
||
|
||
import json
|
||
import logging
|
||
import os
|
||
|
||
logger = logging.getLogger("pipeline.llm_bridge")
|
||
|
||
# Cache: model_name -> {api_base, api_key, model_id}
|
||
_model_cache: dict = {}
|
||
|
||
|
||
def _decrypt_key(encrypted: str) -> str:
|
||
"""Decrypt api_key stored with password_encode. Falls back to plaintext."""
|
||
if not encrypted:
|
||
return ""
|
||
try:
|
||
from appPublic.rc4 import unpassword
|
||
from appPublic.jsonConfig import getConfig
|
||
config = getConfig()
|
||
# 与 ahserver.globalEnv.get_password_key() 一致：password_key 为空时用默认 key
|
||
key = config.password_key or 'QRIVSRHrthhwyjy176556332'
|
||
# unpassword(code, key)：code=密文，key=密钥（之前参数顺序写反了）
|
||
return unpassword(encrypted, key)
|
||
except Exception:
|
||
return encrypted # already plaintext or decrypt failed
|
||
|
||
|
||
async def _get_model_config(model_name: str = None, org_id: str = None) -> dict:
|
||
"""Look up model config from llm table. Returns dict with api_base, api_key, model_id.
|
||
|
||
org_id 多租户隔离：非空时只取本机构(org_id=${org}$)的模型，不含系统级兜底；
|
||
org_id 为空时不过滤（向后兼容）。
|
||
"""
|
||
global _model_cache
|
||
|
||
cache_key = f"{model_name or ''}:{org_id or ''}"
|
||
if model_name and cache_key in _model_cache:
|
||
return _model_cache[cache_key]
|
||
|
||
try:
|
||
from sqlor.dbpools import DBPools
|
||
db = DBPools()
|
||
dbname = "pipeline"
|
||
async with db.sqlorContext(dbname) as sor:
|
||
if model_name:
|
||
# name/model_id 双匹配：与 gateway._resolve_and_persist_model 语义一致——
|
||
# 调用方传 llm.name 或 API 模型名（model_id）都能解析，避免语义漂移。
|
||
sql = ("SELECT api_base, api_key, model_id FROM llm "
|
||
"WHERE status='active' AND (name=${name}$ OR model_id=${name}$)")
|
||
params = {"name": model_name}
|
||
if org_id:
|
||
# 与 gateway 一致：本机构模型 + 系统级共享（org_id 为空/0）
|
||
sql += " AND (org_id=${org}$ OR org_id='' OR org_id='0')"
|
||
params["org"] = org_id
|
||
sql += " LIMIT 1"
|
||
recs = await sor.sqlExe(sql, params)
|
||
else:
|
||
sql = "SELECT api_base, api_key, model_id, name FROM llm WHERE status='active'"
|
||
params = {}
|
||
if org_id:
|
||
sql += " AND org_id=${org}$"
|
||
params["org"] = org_id
|
||
sql += " ORDER BY id LIMIT 1"
|
||
recs = await sor.sqlExe(sql, params)
|
||
if recs:
|
||
r = recs[0]
|
||
cfg = {
|
||
"api_base": getattr(r, "api_base", "") or "",
|
||
"api_key": _decrypt_key(getattr(r, "api_key", "") or ""),
|
||
"model_id": getattr(r, "model_id", "") or "",
|
||
}
|
||
_model_cache[cache_key] = cfg
|
||
return cfg
|
||
except Exception as e:
|
||
logger.warning("llm_bridge: DB lookup failed: %s", e)
|
||
|
||
return {}
|
||
|
||
|
||
def _no_llm_error(model_name=None, org_id=None) -> ValueError:
|
||
"""模型不可用的错误必须真实可行动：写明模型名与原因，前端原样展示、运维据此处理。
|
||
|
||
禁止笼统的 'No LLM API configured'——用户看到它只会卡住干等。
|
||
"""
|
||
if model_name:
|
||
return ValueError(
|
||
f"模型「{model_name}」不可用：llm 表中未找到可用配置"
|
||
f"（name/model_id 不匹配、status 非 active"
|
||
+ (f"，或不属于当前机构 org={org_id}" if org_id else "")
|
||
+ "）。请在模型下拉里改选可用模型，或联系管理员在模型管理中补齐配置。"
|
||
)
|
||
return ValueError(
|
||
"没有可用模型：当前会话未指定模型，且 llm 表无启用模型。"
|
||
"请在模型下拉中选择模型，或联系管理员配置模型。"
|
||
)
|
||
|
||
|
||
# 瞬时错误重试：超时/连接错误/限流/服务端5xx 均重试；配置错误(无key)、鉴权/参数4xx 不重试
|
||
_RETRYABLE_STATUS = (429, 500, 502, 503, 504)
|
||
_LLM_MAX_ATTEMPTS = 3
|
||
# 大上下文（角色 agent 多轮累积）生成偏慢，180s 曾触发超时（空 err=），放宽到 5 分钟；
|
||
# connect 单独设 30s，连接建立失败能快速失败并重试，而不是干等 5 分钟。
|
||
_LLM_TOTAL_TIMEOUT = 300
|
||
_LLM_CONNECT_TIMEOUT = 30
|
||
|
||
|
||
async def _post_chat_completion(url: str, headers: dict, payload: dict) -> dict:
|
||
"""POST /chat/completions，带瞬时错误重试。返回解析后的 JSON dict。"""
|
||
import aiohttp
|
||
import asyncio
|
||
|
||
last_exc = None
|
||
for attempt in range(_LLM_MAX_ATTEMPTS):
|
||
try:
|
||
async with aiohttp.ClientSession() as session:
|
||
async with session.post(
|
||
url, headers=headers, json=payload,
|
||
timeout=aiohttp.ClientTimeout(total=_LLM_TOTAL_TIMEOUT, connect=_LLM_CONNECT_TIMEOUT),
|
||
) as resp:
|
||
if resp.status != 200:
|
||
text = await resp.text()
|
||
err = ValueError("LLM API error %d: %s" % (resp.status, text[:300]))
|
||
if resp.status in _RETRYABLE_STATUS and attempt < _LLM_MAX_ATTEMPTS - 1:
|
||
last_exc = err
|
||
await asyncio.sleep(2 * (attempt + 1))
|
||
continue
|
||
raise err
|
||
data = await resp.json()
|
||
# 网关偶发返回 200 但内容无 choices（错误 JSON），视为可重试的瞬时异常；
|
||
# 不处理会导致上层 llm_call_msgs_native 的 data["choices"] 抛 KeyError。
|
||
if not isinstance(data, dict) or "choices" not in data:
|
||
err = ValueError("LLM 响应缺 choices: %s" % json.dumps(data, ensure_ascii=False)[:300])
|
||
if attempt < _LLM_MAX_ATTEMPTS - 1:
|
||
last_exc = err
|
||
await asyncio.sleep(2 * (attempt + 1))
|
||
continue
|
||
raise err
|
||
return data
|
||
except (asyncio.TimeoutError, aiohttp.ClientError) as e:
|
||
last_exc = e
|
||
if attempt < _LLM_MAX_ATTEMPTS - 1:
|
||
logger.warning("llm_bridge: 瞬时错误重试 %d/%d: %s", attempt + 1, _LLM_MAX_ATTEMPTS, e)
|
||
await asyncio.sleep(2 * (attempt + 1))
|
||
continue
|
||
raise
|
||
|
||
raise last_exc if last_exc else ValueError("LLM call failed")
|
||
|
||
|
||
# ────────────────────────── pipeline_llm 治理钩子 ──────────────────────────
|
||
# 机构配置了治理（llm_org_policy / llm_model 有记录）→ 调用走门禁链（限流/限额/
|
||
# 主备容错/端点选择/预授权），调用后按实际用量结算（双维度记账）。
|
||
# 未配置治理 → 返回 __LEGACY__ 走下方旧 llm 表逻辑（向后兼容，现有调用零改动）。
|
||
# 治理真实失败（限流/余额不足/候选耗尽）→ 抛 ValueError，消息真实可行动，禁止静默回退
|
||
# （否则治理被绕过，限额形同虚设）。
|
||
|
||
async def _resolve_cfg_with_govern(model, org_id, user_id, est_text):
|
||
"""返回 (cfg, gctx)。gctx 非空 = 本次调用受治理（调用后须结算）。"""
|
||
gctx = None
|
||
try:
|
||
from pipeline_llm.gateway import govern_resolve
|
||
except ImportError:
|
||
govern_resolve = None
|
||
if govern_resolve is not None and org_id:
|
||
try:
|
||
est = max((len(est_text) if est_text else 0) // 2, 200)
|
||
ok, res = await govern_resolve(
|
||
org_id=org_id, user_id=user_id or '',
|
||
model_name=model or '', est_tokens=est)
|
||
if ok and isinstance(res, dict):
|
||
gctx = res
|
||
elif res != '__LEGACY__':
|
||
raise ValueError(res)
|
||
except ValueError:
|
||
raise
|
||
except Exception as e:
|
||
logger.warning("llm_bridge: 治理前置异常（回退旧表）: %s", e)
|
||
if gctx:
|
||
cfg = {"api_base": gctx["api_base"], "api_key": gctx["api_key"],
|
||
"model_id": gctx["model_id"]}
|
||
return cfg, gctx
|
||
cfg = await _get_model_config(model, org_id=org_id)
|
||
return cfg, None
|
||
|
||
|
||
async def _settle_govern(gctx, data, est_text):
|
||
"""调用成功后结算：优先上游真实 usage，缺失则按文本长度估算（note=est）。"""
|
||
if not gctx:
|
||
return
|
||
try:
|
||
from pipeline_llm.gateway import govern_settle
|
||
usage = (data or {}).get('usage') or {}
|
||
rt = int(usage.get('prompt_tokens') or 0)
|
||
ct = int(usage.get('completion_tokens') or 0)
|
||
note = ''
|
||
if not usage:
|
||
rt = rt or max((len(est_text) if est_text else 0) // 2, 100)
|
||
ct = ct or 200
|
||
note = 'est'
|
||
await govern_settle(gctx, True, rt, ct, note)
|
||
except Exception as e:
|
||
logger.warning("llm_bridge: 治理结算失败（不阻断调用）: %s", e)
|
||
|
||
|
||
async def _settle_govern_failed(gctx, note):
|
||
"""调用失败时结算：释放预授权，写 failed 流水。"""
|
||
if not gctx:
|
||
return
|
||
try:
|
||
from pipeline_llm.gateway import govern_settle
|
||
await govern_settle(gctx, False, 0, 0, str(note)[:200])
|
||
except Exception as e:
|
||
logger.warning("llm_bridge: 治理失败结算异常: %s", e)
|
||
|
||
|
||
def _msgs_text(messages):
|
||
try:
|
||
return ' '.join(str(m.get('content', '')) for m in (messages or []) if isinstance(m, dict))
|
||
except Exception:
|
||
return ''
|
||
|
||
|
||
async def llm_call(prompt: str, model: str = None, temperature: float = 0.7,
|
||
org_id: str = None, user_id: str = None) -> str:
|
||
"""Call LLM and return text response.
|
||
|
||
Backend priority:
|
||
1. harnessed_agent.llm_chat (if loaded in ServerEnv)
|
||
2. DB llm table (api_base + api_key)
|
||
3. Environment variables (LLM_API_BASE, LLM_API_KEY, LLM_MODEL)
|
||
|
||
org_id 多租户隔离：非空时 llm 表查询只取「本机构 + 系统级」模型。
|
||
"""
|
||
# Priority 1: harnessed_agent
|
||
try:
|
||
from ahserver.serverenv import ServerEnv
|
||
env = ServerEnv()
|
||
if hasattr(env, 'llm_chat'):
|
||
result = await env.llm_chat(prompt, model=model, temperature=temperature)
|
||
if isinstance(result, dict):
|
||
return result.get("content", result.get("text", str(result)))
|
||
return str(result)
|
||
except Exception:
|
||
pass
|
||
|
||
# Priority 2: DB llm table（治理启用时前置走门禁链）
|
||
cfg, gctx = await _resolve_cfg_with_govern(model, org_id, user_id, prompt)
|
||
if cfg.get("api_key") and cfg.get("api_base"):
|
||
api_base = cfg["api_base"]
|
||
api_key = cfg["api_key"]
|
||
model_id = cfg.get("model_id") or model or "default"
|
||
logger.info("llm_bridge: using %s model config for %s -> %s",
|
||
"governed" if gctx else "DB", model, api_base)
|
||
else:
|
||
# Priority 3: Environment variables
|
||
api_base = os.environ.get("LLM_API_BASE", "https://api.openai.com/v1")
|
||
api_key = os.environ.get("LLM_API_KEY", "")
|
||
model_id = model or os.environ.get("LLM_MODEL", "gpt-4o-mini")
|
||
|
||
if not api_key:
|
||
raise _no_llm_error(model, org_id)
|
||
|
||
import aiohttp
|
||
|
||
headers = {
|
||
"Authorization": f"Bearer {api_key}",
|
||
"Content-Type": "application/json",
|
||
}
|
||
payload = {
|
||
"model": model_id,
|
||
"messages": [{"role": "user", "content": prompt}],
|
||
"temperature": temperature,
|
||
}
|
||
|
||
url = api_base.rstrip("/") + "/chat/completions"
|
||
try:
|
||
data = await _post_chat_completion(url, headers, payload)
|
||
except Exception as e:
|
||
await _settle_govern_failed(gctx, e)
|
||
raise
|
||
await _settle_govern(gctx, data, prompt)
|
||
return data["choices"][0]["message"]["content"]
|
||
|
||
|
||
async def call_llm(tenant_id: str, prompt: str, model: str = None, temperature: float = 0.7) -> str:
|
||
"""SDLC handler interface — delegates to llm_call."""
|
||
return await llm_call(prompt, model=model, temperature=temperature)
|
||
|
||
|
||
async def llm_call_msgs(messages: list, model: str = None, temperature: float = 0.7,
|
||
org_id: str = None, user_id: str = None) -> str:
|
||
"""Call LLM with full message array (system/user/assistant)."""
|
||
import aiohttp
|
||
|
||
cfg, gctx = await _resolve_cfg_with_govern(model, org_id, user_id, _msgs_text(messages))
|
||
if cfg.get("api_key") and cfg.get("api_base"):
|
||
api_base = cfg["api_base"]
|
||
api_key = cfg["api_key"]
|
||
model_id = cfg.get("model_id") or model or "default"
|
||
else:
|
||
api_base = os.environ.get("LLM_API_BASE", "https://api.openai.com/v1")
|
||
api_key = os.environ.get("LLM_API_KEY", "")
|
||
model_id = model or os.environ.get("LLM_MODEL", "gpt-4o-mini")
|
||
|
||
if not api_key:
|
||
raise _no_llm_error(model, org_id)
|
||
|
||
headers = {"Authorization": f"Bearer {api_key}", "Content-Type": "application/json"}
|
||
payload = {"model": model_id, "messages": messages, "temperature": temperature}
|
||
|
||
url = api_base.rstrip("/") + "/chat/completions"
|
||
try:
|
||
data = await _post_chat_completion(url, headers, payload)
|
||
except Exception as e:
|
||
await _settle_govern_failed(gctx, e)
|
||
raise
|
||
await _settle_govern(gctx, data, _msgs_text(messages))
|
||
return data["choices"][0]["message"]["content"]
|
||
|
||
|
||
async def llm_call_msgs_native(messages: list, tools: list = None, model: str = None,
|
||
temperature: float = 0.7, org_id: str = None,
|
||
user_id: str = None) -> dict:
|
||
"""Native function calling. 传入 tools JSON schema，返回 message dict。
|
||
|
||
Returns:
|
||
{"content": str, "tool_calls": [{"id","type","function":{"name","arguments"}}]}
|
||
当模型返回 tool_calls 时，content 通常为空字符串。
|
||
"""
|
||
import aiohttp
|
||
|
||
cfg, gctx = await _resolve_cfg_with_govern(model, org_id, user_id, _msgs_text(messages))
|
||
if cfg.get("api_key") and cfg.get("api_base"):
|
||
api_base = cfg["api_base"]
|
||
api_key = cfg["api_key"]
|
||
model_id = cfg.get("model_id") or model or "default"
|
||
else:
|
||
api_base = os.environ.get("LLM_API_BASE", "https://api.openai.com/v1")
|
||
api_key = os.environ.get("LLM_API_KEY", "")
|
||
model_id = model or os.environ.get("LLM_MODEL", "gpt-4o-mini")
|
||
|
||
if not api_key:
|
||
raise _no_llm_error(model, org_id)
|
||
|
||
headers = {"Authorization": f"Bearer {api_key}", "Content-Type": "application/json"}
|
||
payload = {"model": model_id, "messages": messages, "temperature": temperature}
|
||
if tools:
|
||
payload["tools"] = tools
|
||
payload["tool_choice"] = "auto"
|
||
|
||
url = api_base.rstrip("/") + "/chat/completions"
|
||
try:
|
||
data = await _post_chat_completion(url, headers, payload)
|
||
except Exception as e:
|
||
await _settle_govern_failed(gctx, e)
|
||
raise
|
||
await _settle_govern(gctx, data, _msgs_text(messages))
|
||
msg = data["choices"][0]["message"]
|
||
return {
|
||
"content": msg.get("content") or "",
|
||
"tool_calls": msg.get("tool_calls") or [],
|
||
}
|
||
|