1571 lines
77 KiB
Python
1571 lines
77 KiB
Python
"""
|
||
pipeline-service: platform_ability — 平台内部 agent 能力包（pipeline_id=platform_general）
|
||
|
||
产线平台自身的运维/管理 agent：拥有平台应用和各模块的技能，
|
||
核心能力是「通读模型 API 文档 → 自动生成模型治理配置（供应商/适配模板/模型/定价）」。
|
||
|
||
权限模型（代码层硬门禁，不依赖 prompt）：
|
||
所有工具 handler 入口校验调用者持有 owner 组织的角色（orgtypeid='owner'，含通配 'owner.*'）。
|
||
内部 agent 只服务 owner 组织的角色——其他组织的用户即使打开页面也调不动任何工具。
|
||
|
||
配置生成链路（用户给定 API 文档 → 完成配置）：
|
||
① fetch_model_doc：抓取文档页面（SSRF 防护：仅 http(s) 公网域名）
|
||
② extract_llm_api_spec：LLM 从文档提取端点/协议/请求响应格式/定价
|
||
③ apply_llm_config：写库——复用 pipeline-llm 模块的 CRUD 函数
|
||
（llm_vendor.endpoints + llm_api_profile 模板 + llm_model 含四价）
|
||
"""
|
||
|
||
import asyncio
|
||
import json
|
||
import logging
|
||
import re
|
||
import time
|
||
|
||
import yaml
|
||
|
||
from appPublic.uniqueID import getID
|
||
|
||
from pipeline_core import (
|
||
ToolDefinition,
|
||
PipelineAbility,
|
||
register_ability,
|
||
)
|
||
|
||
logger = logging.getLogger("pipeline.platform_ability")
|
||
|
||
PLATFORM_PIPELINE_ID = "platform_general"
|
||
|
||
# 组织门禁：内部 agent 仅 owner 组织的角色可操作（任一 owner.* 角色，含通配）
|
||
OWNER_ORG = "owner"
|
||
|
||
# 当前运行时调用链支持的协议（模式定义在模型的适配模板/执行器，非供应商）：
|
||
# openai_compat —— 同步 chat/completions 一次往返
|
||
# dashscope_async —— 异步执行器：提交任务→轮询 query_profile_ids→取结果（2026-09-05）
|
||
# 其他协议的适配模板会存入 llm_api_profile 备查，但运行时暂不渲染。
|
||
RUNTIME_PROTOCOLS = ("openai_compat", "dashscope_async")
|
||
|
||
# ────────────────────── 会话级规格缓存（extract 锚定，2026-09-05）──────────────────────
|
||
# 根因：apply_llm_config 要求 LLM 把 extract 返回的大 JSON 规格逐字复制进参数，
|
||
# 模型「转述」大 JSON 必然编造结构（实测三轮全编造 vendor{}/api_profile{} 等契约外字段）。
|
||
# 对策：extract 成功后把规格锚定到「会话级缓存」，apply/apply_model_pricing 用
|
||
# use_last_extract=true 直接取用 + 少量覆盖项（vendor_name 等），LLM 永不搬运大 JSON。
|
||
#
|
||
# 会话级隔离（用户硬要求）：缓存键 = pipeline_id:user_id:session_id，三段缺一不可，
|
||
# 防跨会话/跨用户/跨产线串数据。存 Redis db4（治理命名空间，gateway 已在用），
|
||
# 跨 worker 进程不丢；TTL 兜底过期。Redis 不可用时降级进程内字典（仍按同键隔离）。
|
||
_SPEC_CACHE_TTL = 7200 # 规格缓存 2 小时（一次配置会话足够）
|
||
_SPEC_CACHE_PREFIX = "llm_spec"
|
||
_spec_cache_local = {} # Redis 不可用时的进程内降级（同样按隔离键）
|
||
|
||
|
||
def _spec_cache_key(ctx):
|
||
"""会话级隔离键：pipeline_id:user_id:session_id。任一段缺失返回 ''（拒绝缓存）。"""
|
||
pl = (ctx.get("pipeline_id") or "").strip()
|
||
uid = (ctx.get("user_id") or "").strip()
|
||
sid = (ctx.get("session_id") or "").strip()
|
||
if not (pl and uid and sid):
|
||
return ''
|
||
return "%s:%s:%s:%s" % (_SPEC_CACHE_PREFIX, pl, uid, sid)
|
||
|
||
|
||
def _spec_redis():
|
||
"""复用治理命名空间 Redis db4（与 gateway 限流同库，会话态用 db3 不冲突）。"""
|
||
try:
|
||
import redis as _r
|
||
try:
|
||
from appPublic.jsonConfig import getConfig
|
||
url = (getConfig().website or {}).get('session_redis', {}).get('url', '')
|
||
except Exception:
|
||
url = ''
|
||
if not url:
|
||
url = 'redis://127.0.0.1:6379/3'
|
||
base = url.rsplit('/', 1)[0]
|
||
return _r.Redis.from_url(base + '/4', socket_timeout=2)
|
||
except Exception:
|
||
return None
|
||
|
||
|
||
def _spec_save(ctx, spec):
|
||
"""锚定规格到会话缓存。返回 True/False。无隔离键则不缓存（apply 会要求显式传 spec）。"""
|
||
key = _spec_cache_key(ctx)
|
||
if not key:
|
||
return False
|
||
try:
|
||
payload = json.dumps(spec, ensure_ascii=False)
|
||
except Exception:
|
||
return False
|
||
r = _spec_redis()
|
||
if r is not None:
|
||
try:
|
||
r.set(key, payload, ex=_SPEC_CACHE_TTL)
|
||
return True
|
||
except Exception:
|
||
pass
|
||
_spec_cache_local[key] = (payload, time.time() + _SPEC_CACHE_TTL)
|
||
return True
|
||
|
||
|
||
def _spec_load(ctx):
|
||
"""取本会话锚定的规格。无则返回 None。严格会话级隔离（键含 session_id）。"""
|
||
key = _spec_cache_key(ctx)
|
||
if not key:
|
||
return None
|
||
r = _spec_redis()
|
||
raw = None
|
||
if r is not None:
|
||
try:
|
||
v = r.get(key)
|
||
raw = v.decode('utf-8') if isinstance(v, (bytes, bytearray)) else v
|
||
except Exception:
|
||
raw = None
|
||
if raw is None:
|
||
ent = _spec_cache_local.get(key)
|
||
if ent and ent[1] > time.time():
|
||
raw = ent[0]
|
||
elif ent:
|
||
_spec_cache_local.pop(key, None)
|
||
if not raw:
|
||
return None
|
||
try:
|
||
return json.loads(raw)
|
||
except Exception:
|
||
return None
|
||
|
||
|
||
def _spec_overlay(base, overrides):
|
||
"""覆盖项合并：仅允许少量标量覆盖（vendor_name 等），不做大 JSON 搬运。
|
||
|
||
overrides 只认白名单键，防止 LLM 借覆盖项塞回编造结构。
|
||
"""
|
||
allowed = {'vendor_name', 'base_url', 'protocol', 'doc_url', 'doc_notes'}
|
||
out = dict(base or {})
|
||
for k in allowed:
|
||
if k in (overrides or {}) and str(overrides.get(k) or '').strip():
|
||
out[k] = str(overrides[k]).strip()
|
||
return out
|
||
|
||
|
||
# ────────────────────── 权限门禁（代码层） ──────────────────────
|
||
|
||
async def _require_owner(sor, ctx) -> str:
|
||
"""校验当前用户持有 owner 组织的角色。返回 ''=通过，否则错误信息。
|
||
|
||
判定：userrole→role JOIN 后，存在任一 orgtypeid='owner' 的角色（含通配 'owner.*'）。
|
||
"""
|
||
uid = ctx.get("user_id", "") or ""
|
||
if not uid:
|
||
return "无法识别当前用户身份（未登录），内部 agent 仅 owner 组织角色可用"
|
||
recs = await sor.sqlExe(
|
||
"SELECT r.orgtypeid, r.name FROM userrole ur JOIN role r ON ur.roleid=r.id "
|
||
"WHERE ur.userid=${u}$", {"u": uid})
|
||
await sor.sqlExe("COMMIT", {})
|
||
roles = []
|
||
for r in (recs or []):
|
||
o = getattr(r, "orgtypeid", "") or ""
|
||
n = getattr(r, "name", "") or ""
|
||
if o and n:
|
||
roles.append("%s.%s" % (o, n))
|
||
if any(fn.split(".", 1)[0] == OWNER_ORG for fn in roles):
|
||
return ""
|
||
return "权限不足：内部 agent 工具仅 %s 组织的角色可操作，当前角色 %s" % (
|
||
OWNER_ORG, sorted(set(roles)) or "无")
|
||
|
||
|
||
def _row(r):
|
||
"""sqlor row → dict（容忍属性访问）。"""
|
||
if hasattr(r, "__dict__"):
|
||
return {k: v for k, v in r.__dict__.items() if not k.startswith("_")}
|
||
return {}
|
||
|
||
|
||
# ────────────────────── 工具 1：模型治理状态 ──────────────────────
|
||
|
||
async def _h_platform_llm_status(sor, params, ctx):
|
||
"""供应商/账号/模型/用量概览（owner 组织诊断配置用）。"""
|
||
err = await _require_owner(sor, ctx)
|
||
if err:
|
||
return err
|
||
out = {}
|
||
for tbl in ("llm_vendor", "llm_account", "llm_model"):
|
||
recs = await sor.sqlExe(
|
||
"SELECT status, COUNT(*) AS c FROM " + tbl + " GROUP BY status", {})
|
||
await sor.sqlExe("COMMIT", {})
|
||
out[tbl] = {getattr(r, "status", ""): int(getattr(r, "c", 0)) for r in (recs or [])}
|
||
accs = await sor.sqlExe(
|
||
"SELECT name, balance, status FROM llm_account ORDER BY balance DESC LIMIT 20", {})
|
||
await sor.sqlExe("COMMIT", {})
|
||
out["accounts"] = [
|
||
{"name": getattr(r, "name", ""), "balance": float(getattr(r, "balance", 0) or 0),
|
||
"status": getattr(r, "status", "")} for r in (accs or [])]
|
||
models = await sor.sqlExe(
|
||
"SELECT name, vendor_model_id, capability, status "
|
||
"FROM llm_model ORDER BY created_at DESC LIMIT 30", {})
|
||
await sor.sqlExe("COMMIT", {})
|
||
out["models"] = [
|
||
{"name": getattr(r, "name", ""), "vendor_model_id": getattr(r, "vendor_model_id", ""),
|
||
"capability": getattr(r, "capability", ""),
|
||
"status": getattr(r, "status", "")} for r in (models or [])]
|
||
return json.dumps(out, ensure_ascii=False)
|
||
|
||
|
||
# ────────────────────── 工具 2：抓取模型 API 文档 ──────────────────────
|
||
|
||
_MAX_DOC_CHARS = 60000
|
||
_MAX_REDIRECTS = 5
|
||
|
||
|
||
def _validate_doc_url(url: str) -> str:
|
||
"""SSRF 防护：仅允许 http(s) + 公网域名。返回 ''=通过，否则错误。"""
|
||
if not re.match(r"^https?://", url or ""):
|
||
return "文档 URL 必须是 http/https 链接"
|
||
host = re.sub(r"^https?://", "", url).split("/")[0].split(":")[0].lower()
|
||
if not re.match(r"^[a-z0-9][a-z0-9.-]+\.[a-z]{2,}$", host):
|
||
return "文档 URL 域名非法（不接受 IP/内网地址）"
|
||
if host in ("localhost",) or host.endswith((".local", ".internal", ".localhost")):
|
||
return "文档 URL 不允许指向内网/本地地址"
|
||
return ""
|
||
|
||
|
||
def _html_to_text(html: str) -> str:
|
||
"""粗暴去标签提取文本（文档页面用，不追求完美排版）。"""
|
||
txt = re.sub(r"(?is)<(script|style|noscript)[^>]*>.*?</\1>", " ", html or "")
|
||
txt = re.sub(r"(?is)<[^>]+>", " ", txt)
|
||
txt = re.sub(r" ", " ", txt)
|
||
txt = re.sub(r"<", "<", txt)
|
||
txt = re.sub(r">", ">", txt)
|
||
txt = re.sub(r"&", "&", txt)
|
||
txt = re.sub(r""", '"', txt)
|
||
txt = re.sub(r"[ \t]+", " ", txt)
|
||
txt = re.sub(r"\n\s*\n+", "\n", txt)
|
||
return txt.strip()
|
||
|
||
|
||
def _is_private_ip(ip: str) -> bool:
|
||
"""DNS 解析后的二次防线：解析到私有/环回/链路本地地址一律拒绝（DNS rebinding 防护）。"""
|
||
import ipaddress
|
||
try:
|
||
a = ipaddress.ip_address(ip)
|
||
except ValueError:
|
||
return True # 解析不出就当私有（拒绝）
|
||
return (a.is_private or a.is_loopback or a.is_link_local
|
||
or a.is_reserved or a.is_multicast or a.is_unspecified)
|
||
|
||
|
||
async def _fetch_url_safe(url: str, max_redirects: int = _MAX_REDIRECTS):
|
||
"""带重定向逐跳校验的抓取（每一跳都过域名+DNS双重校验）。
|
||
|
||
返回 (text, content_type)；失败抛 ValueError。
|
||
"""
|
||
import aiohttp
|
||
current = url
|
||
for _ in range(max_redirects + 1):
|
||
verr = _validate_doc_url(current)
|
||
if verr:
|
||
raise ValueError(verr)
|
||
host = re.sub(r"^https?://", "", current).split("/")[0].split(":")[0].lower()
|
||
# DNS 解析校验（同步阻塞短调用，可接受；防 DNS rebinding 指向内网）
|
||
try:
|
||
infos = await asyncio.get_event_loop().getaddrinfo(host, None)
|
||
except Exception as e:
|
||
raise ValueError("文档域名无法解析：%s" % str(e)[:120])
|
||
for info in infos:
|
||
ip = str(info[4][0])
|
||
if _is_private_ip(ip):
|
||
raise ValueError("文档域名解析到内网地址（%s），拒绝访问" % ip)
|
||
timeout = aiohttp.ClientTimeout(total=30)
|
||
async with aiohttp.ClientSession(timeout=timeout) as sess:
|
||
async with sess.get(current, headers={"User-Agent": "Mozilla/5.0"},
|
||
allow_redirects=False, ssl=False) as resp:
|
||
if resp.status in (301, 302, 303, 307, 308):
|
||
loc = resp.headers.get("Location", "")
|
||
if not loc:
|
||
raise ValueError("重定向缺少 Location")
|
||
if loc.startswith("/"):
|
||
scheme = "https" if current.startswith("https") else "http"
|
||
loc = "%s://%s%s" % (scheme, host, loc)
|
||
current = loc
|
||
continue
|
||
if resp.status != 200:
|
||
raise ValueError("抓取失败：HTTP %d" % resp.status)
|
||
ctype = resp.headers.get("Content-Type", "")
|
||
raw = await resp.text(errors="replace")
|
||
return raw, ctype
|
||
raise ValueError("重定向次数超限（>%d）" % max_redirects)
|
||
|
||
|
||
async def _h_fetch_model_doc(sor, params, ctx):
|
||
"""抓取模型 API 文档页面 → 纯文本（供 LLM 提取配置规格）。"""
|
||
err = await _require_owner(sor, ctx)
|
||
if err:
|
||
return err
|
||
url = (params.get("url") or "").strip()
|
||
try:
|
||
raw, ctype = await _fetch_url_safe(url)
|
||
except ValueError as e:
|
||
return str(e)
|
||
except Exception as e:
|
||
return "抓取失败：%s" % str(e)[:200]
|
||
if "html" in ctype.lower() or raw.lstrip()[:15].lower().startswith(("<!doctype", "<html")):
|
||
text = _html_to_text(raw)
|
||
else:
|
||
text = raw
|
||
if len(text) > _MAX_DOC_CHARS:
|
||
text = text[:_MAX_DOC_CHARS] + "\n[文档过长已截断，共 %d 字符]" % len(text)
|
||
# 出处行：提取规格照抄进 doc_url，落库写进模型/定价描述（2026-09-05 用户规则）
|
||
text = text + "\n\n[出处URL] " + url
|
||
return text
|
||
|
||
|
||
# ────────────────────── 工具 3：提取配置规格（LLM） ──────────────────────
|
||
|
||
_EXTRACT_PROMPT = """你是大模型 API 配置专家。通读下面这份模型 API 文档，提取配置规格。
|
||
|
||
只输出一个 JSON 对象（不要 markdown 代码块），字段：
|
||
{
|
||
"vendor_name": "供应商名称",
|
||
"base_url": "API 基础地址（如 https://dashscope.aliyuncs.com/compatible-mode/v1）",
|
||
"protocol": "openai_compat | dashscope_async | custom（能走 /chat/completions 的填 openai_compat）",
|
||
"endpoints": [{"base_url": "...", "region": "domestic|international", "timeout": 60}],
|
||
"chat_path": "对话接口路径（如 /chat/completions）",
|
||
"auth_header": "认证头格式说明（如 Bearer API_KEY）",
|
||
"request_headers": {"说明": "文档调用示例里除认证外的必需请求头，逐字照抄（如 DashScope 异步的 X-DashScope-Async: enable）；无则 null"},
|
||
"request_fields": ["请求体字段名列表"],
|
||
"request_example": {"说明": "文档调用示例(cURL/代码)中的完整请求体 JSON，逐字照抄结构与各字段示例值，不要修改；文档无示例填 null"},
|
||
"response_format": "响应格式说明（content 字段路径 + usage 字段路径）",
|
||
"response_example": {"说明": "文档中的响应示例 JSON（或结果字段说明），逐字照抄；无则 null"},
|
||
"async_steps": [{"purpose": "query|download", "path": "该步骤接口路径", "method": "GET|POST", "request_fields": ["请求字段名列表"], "response_format": "该步骤响应格式说明"}],
|
||
"models": [{"vendor_model_id": "供应商侧模型ID", "capability": "t2t|t2i|i2t|t2v|i2v|embedding|rerank|tts|asr",
|
||
"sync_mode": "sync|async（提交后需轮询查询结果的填 async）",
|
||
"description": "一句话说明"}],
|
||
"pricing": {
|
||
"currency": "CNY|USD（文档标注的币种）",
|
||
"items": [{"vendor_model_id": "对应模型ID",
|
||
"factor": "duration|flat|prompt_tokens|completion_tokens（计价因子：视频按秒=duration，按次=flat，token计价=对应token因子）",
|
||
"unit_price": 0.45,
|
||
"unit": "秒|次|百万（文档标注的计价单位名）",
|
||
"dimensions": {"resolution": "480P（影响价格档位的维度，按文档原文值；禁止放 model）"},
|
||
"doc_quote": "文档原文定价句（逐字照抄，标明出处用）"}]
|
||
},
|
||
"doc_url": "从文档文本末尾的 [出处URL] 行逐字照抄（配置出处，必填）",
|
||
"doc_notes": "文档中影响配置的关键注意点"
|
||
}
|
||
|
||
异步模型规则：
|
||
- 文档描述「先提交任务、再轮询查询结果」的模型，sync_mode 填 async，并提取 async_steps：
|
||
至少一条 purpose=query（查询任务状态/结果）；若文档另有独立下载/取文件接口，再加一条 purpose=download。
|
||
- 任务查询接口通常是**供应商级共用**的（如 DashScope 全系生成模型都是 GET /tasks/{task_id}），
|
||
path 照文档逐字抄；同供应商多个模型会复用同一份查询模板，不要因模型而异。
|
||
- 同步模型 async_steps 填空数组 []。
|
||
|
||
定价规则：
|
||
- pricing.items 每条必须带 doc_quote（文档原文定价句逐字照抄）——定价只允许来自文档原文
|
||
- 价格/单位按文档原文记录，禁止自行换算、推导或统一单位
|
||
- 同一模型多档价格（如不同分辨率）拆成多条 item，各带自己的 dimensions 与 doc_quote
|
||
- dimensions 禁止放 model——一个定价方案只服务一个模型；定价完全相同的多个模型
|
||
共享同一个定价方案（同一 ppid），不是往定价里加 model 过滤
|
||
- 文档没写价格：items 填空数组 []（禁止编造价格）
|
||
|
||
文档内容：
|
||
"""
|
||
|
||
|
||
async def _h_extract_llm_api_spec(sor, params, ctx):
|
||
"""LLM 通读文档文本 → 结构化配置规格（JSON）。"""
|
||
err = await _require_owner(sor, ctx)
|
||
if err:
|
||
return err
|
||
doc_text = (params.get("doc_text") or "").strip()
|
||
if len(doc_text) < 100:
|
||
return "文档文本太短（<100字符），无法提取配置——先用 fetch_model_doc 抓取"
|
||
try:
|
||
from pipeline_service.llm_bridge import llm_call_msgs
|
||
# purpose='utility'：辅助任务走辅助模型链（机构策略配置则优先，便宜快）
|
||
# timeout=300：长文档提取慢，端点默认 60 秒实测跑不完（2026-09-04 根因）
|
||
raw = await llm_call_msgs(
|
||
[{"role": "system", "content": _EXTRACT_PROMPT},
|
||
{"role": "user", "content": doc_text[:48000]}],
|
||
temperature=0.1, org_id="0", user_id=ctx.get("user_id", ""),
|
||
purpose="utility", timeout=300)
|
||
except Exception as e:
|
||
return "LLM 提取失败：%s" % str(e)[:200]
|
||
txt = (raw or "").strip()
|
||
if txt.startswith("```"):
|
||
txt = txt.split("\n", 1)[1] if "\n" in txt else txt
|
||
txt = txt.rsplit("```", 1)[0]
|
||
try:
|
||
spec = json.loads(txt)
|
||
except Exception:
|
||
m = re.search(r"\{.*\}", txt, re.S)
|
||
if not m:
|
||
return "LLM 输出不是合法 JSON：%s" % txt[:300]
|
||
try:
|
||
spec = json.loads(m.group(0))
|
||
except Exception:
|
||
return "LLM 输出不是合法 JSON：%s" % txt[:300]
|
||
# 会话级锚定（2026-09-05）：规格存会话缓存（键含 session_id，跨会话隔离），
|
||
# 后续 apply_llm_config / apply_model_pricing 用 use_last_extract=true 取用——
|
||
# LLM 不再搬运/转述大 JSON（转述必编造结构，实测三轮全编造）。
|
||
anchored = _spec_save(ctx, spec)
|
||
out = dict(spec)
|
||
if anchored:
|
||
out["__anchored__"] = ("规格已锚定到本会话缓存。下一步调用 apply_llm_config 与 "
|
||
"apply_model_pricing 时传 {\"use_last_extract\": true}，"
|
||
"需改动只传覆盖项（overrides，仅 vendor_name/base_url/"
|
||
"protocol/doc_url/doc_notes 白名单键）。"
|
||
"禁止把本规格复制进 spec 参数——复制即编造。")
|
||
return json.dumps(out, ensure_ascii=False)
|
||
|
||
|
||
# ────────────────────── 工具 4：写入配置（幂等） ──────────────────────
|
||
|
||
def _f(v, default=0.0):
|
||
try:
|
||
return float(v)
|
||
except (TypeError, ValueError):
|
||
return default
|
||
|
||
|
||
async def _h_apply_llm_config(sor, params, ctx):
|
||
"""按提取的规格写库：供应商（复用/新建）+ 适配模板 + 模型。
|
||
|
||
幂等：供应商按名称复用；模型按 name/vendor_model_id 复用（name 全局唯一键）——
|
||
已存在则更新元数据；供应商归属不一致时不改挂、如实报 vendor_conflicts。
|
||
|
||
规格来源（2026-09-05 防编造改造）：
|
||
use_last_extract=true（推荐）→ 取本会话 extract 锚定的规格，
|
||
overrides 白名单覆盖（vendor_name 等少量标量）；
|
||
否则用 params.spec（LLM 手传，历史路径）——结构不符时报错列出期望键。
|
||
"""
|
||
err = await _require_owner(sor, ctx)
|
||
if err:
|
||
return err
|
||
use_last = str(params.get("use_last_extract") or "").lower() in ("true", "1", "yes")
|
||
overrides = params.get("overrides") or {}
|
||
if isinstance(overrides, str):
|
||
try:
|
||
overrides = json.loads(overrides)
|
||
except Exception:
|
||
overrides = {}
|
||
spec = None
|
||
if use_last:
|
||
spec = _spec_load(ctx)
|
||
if spec is None:
|
||
return ("本会话没有锚定的提取规格——先调 extract_llm_api_spec（成功后规格自动"
|
||
"锚定本会话），再传 {\"use_last_extract\": true}。"
|
||
"注意会话级隔离：其他会话提取的规格本会话不可见，需在本会话重新提取。")
|
||
spec = _spec_overlay(spec, overrides)
|
||
else:
|
||
try:
|
||
spec = json.loads(params.get("spec") or "{}")
|
||
except Exception:
|
||
return "spec 不是合法 JSON"
|
||
# 结构校验（修B：报错给出路，不让模型盲改重试）
|
||
if not isinstance(spec, dict) or "vendor_name" not in spec or "models" not in spec:
|
||
got = sorted(spec.keys())[:12] if isinstance(spec, dict) else "(非对象)"
|
||
return ("spec 结构不符——不要自己编写/改写规格结构。期望顶层键："
|
||
"vendor_name / base_url / endpoints / protocol / chat_path / models / "
|
||
"pricing / doc_url / doc_notes（extract_llm_api_spec 的原样输出）。"
|
||
"你传入的顶层键：%s。正确做法：调 extract_llm_api_spec 后传 "
|
||
"{\"use_last_extract\": true}（规格已锚定本会话，禁止复制转述），"
|
||
"需改动加 overrides（仅 vendor_name/base_url/protocol/doc_url/doc_notes）。"
|
||
% got)
|
||
vendor_name = (spec.get("vendor_name") or "").strip()
|
||
if not vendor_name:
|
||
return "spec 缺少 vendor_name（可用 overrides.vendor_name 指定）"
|
||
endpoints = spec.get("endpoints") or []
|
||
if not endpoints and spec.get("base_url"):
|
||
endpoints = [{"base_url": spec["base_url"], "region": "domestic", "timeout": 60}]
|
||
if not endpoints:
|
||
return "spec 缺少 endpoints/base_url"
|
||
protocol = (spec.get("protocol") or "openai_compat").strip() or "openai_compat"
|
||
models = spec.get("models") or []
|
||
if not models:
|
||
return "spec.models 为空——文档里没有可配置的模型？"
|
||
|
||
# 能力分类校验（2026-09-05 用户定夺的规则）：模型能力必须是字典已登记的
|
||
# 分类；不存在则拒绝落库——先加能力分类（appcodes_kv llm_capability，
|
||
# 含种子/提示词/端点注释四处同步），再配模型。防 LLM 静默塞进近似分类。
|
||
recs = await sor.sqlExe(
|
||
"SELECT k FROM appcodes_kv WHERE parentid='llm_capability'", {})
|
||
await sor.sqlExe("COMMIT", {})
|
||
known_caps = set(getattr(r, "k", "") for r in (recs or []))
|
||
unknown = sorted(set(
|
||
(m.get("capability") or "t2t").strip() for m in models) - known_caps)
|
||
if unknown:
|
||
return ("能力分类未登记，拒绝落库：%s。请先在能力分类字典"
|
||
"（appcodes_kv parentid=llm_capability，含种子数据/提取提示词/"
|
||
"模型列表端点注释同步）中添加该分类，再重新执行本工具。"
|
||
"现有分类：%s" % ("、".join(unknown), "、".join(sorted(known_caps))))
|
||
|
||
from appPublic.uniqueID import getID
|
||
|
||
# 1. 供应商：按名称复用
|
||
recs = await sor.sqlExe(
|
||
"SELECT id, endpoints FROM llm_vendor WHERE name=${n}$", {"n": vendor_name})
|
||
await sor.sqlExe("COMMIT", {})
|
||
if recs:
|
||
vendor_id = getattr(recs[0], "id", "")
|
||
eps_old = []
|
||
try:
|
||
eps_old = json.loads(getattr(recs[0], "endpoints", "") or "[]")
|
||
except Exception:
|
||
eps_old = []
|
||
merged = list(eps_old)
|
||
added_eps = []
|
||
for ep in endpoints:
|
||
bu = (ep.get("base_url") or "").rstrip("/")
|
||
if bu and bu not in [(e.get("base_url") or "").rstrip("/") for e in merged]:
|
||
merged.append({"base_url": bu, "region": ep.get("region") or "domestic",
|
||
"timeout": int(_f(ep.get("timeout"), 60))})
|
||
added_eps.append(bu)
|
||
# 2026-09-05：供应商表不再有 protocol——请求形态由模型挂的适配模板决定
|
||
await sor.sqlExe(
|
||
"UPDATE llm_vendor SET endpoints=${e}$, updated_at=NOW() WHERE id=${i}$",
|
||
{"e": json.dumps(merged, ensure_ascii=False), "i": vendor_id})
|
||
await sor.sqlExe("COMMIT", {})
|
||
vendor_action = "复用供应商 %s（新增端点 %d 个）" % (vendor_id, len(added_eps))
|
||
else:
|
||
vendor_id = getID()
|
||
eps_norm = [{"base_url": (ep.get("base_url") or "").rstrip("/"),
|
||
"region": ep.get("region") or "domestic",
|
||
"timeout": int(_f(ep.get("timeout"), 60))} for ep in endpoints]
|
||
await sor.C("llm_vendor", {
|
||
"id": vendor_id, "name": vendor_name,
|
||
"endpoints": json.dumps(eps_norm, ensure_ascii=False),
|
||
"description": spec.get("doc_notes", "") or "",
|
||
"status": "active", "org_id": ctx.get("org_id", "") or "0"})
|
||
vendor_action = "新建供应商 %s（%s）" % (vendor_id, vendor_name)
|
||
|
||
# 2. 适配模板：按 (协议×能力) 复用；旧骨架模板（硬编码 xxx_file/__from_doc__，
|
||
# 2026-09-05 实测 t2v 因它渲染崩）视为不可用，按文档示例自愈重建
|
||
profile_ids = {}
|
||
tpl_errors = []
|
||
tpl_notes = []
|
||
caps = sorted(set((m.get("capability") or "t2t") for m in models))
|
||
for cap in caps:
|
||
recs = await sor.sqlExe(
|
||
"SELECT id, request_template, response_template FROM llm_api_profile "
|
||
"WHERE protocol=${p}$ AND capability=${c}$ AND status='active' "
|
||
"ORDER BY created_at DESC LIMIT 1",
|
||
{"p": protocol, "c": cap})
|
||
await sor.sqlExe("COMMIT", {})
|
||
if recs:
|
||
pid_old = getattr(recs[0], "id", "")
|
||
old_tpl = (getattr(recs[0], "request_template", "") or "") + \
|
||
(getattr(recs[0], "response_template", "") or "")
|
||
if not any(mk in old_tpl for mk in _SKELETON_MARKERS):
|
||
profile_ids[cap] = pid_old
|
||
continue
|
||
await sor.sqlExe("UPDATE llm_api_profile SET status='deprecated' "
|
||
"WHERE id=${i}$", {"i": pid_old})
|
||
await sor.sqlExe("COMMIT", {})
|
||
tpl_notes.append("旧骨架模板 %s（%s）含硬编码占位符，已弃用并按文档示例重建"
|
||
% (pid_old, cap))
|
||
tpl = _gen_templates(protocol, cap, spec)
|
||
dry_err = _dry_render_check(tpl["req"], cap, tpl["biz_params"], tpl["media_params"])
|
||
if dry_err:
|
||
# 宁可不建，不埋雷：渲染不了的模板落库=运行时必崩，如实报错由助手修正规格重跑
|
||
tpl_errors.append("%s: %s" % (cap, dry_err))
|
||
continue
|
||
pid = getID()
|
||
await sor.C("llm_api_profile", {
|
||
"id": pid, "name": "%s-%s-自动配置" % (protocol, cap),
|
||
"protocol": protocol, "capability": cap,
|
||
"path": tpl["path"],
|
||
"headers": json.dumps(tpl["headers"], ensure_ascii=False),
|
||
"request_template": tpl["req"], "response_template": tpl["resp"],
|
||
"param_schema": json.dumps(tpl["param_schema"], ensure_ascii=False),
|
||
"status": "active"})
|
||
profile_ids[cap] = pid
|
||
for nt in tpl["notes"]:
|
||
tpl_notes.append("%s: %s" % (cap, nt))
|
||
|
||
# 3. 模型：按 name/vendor_model_id 幂等（name 全局唯一键）
|
||
doc_url = (spec.get("doc_url") or "").strip()
|
||
created, updated, skipped = [], [], []
|
||
conflicts = [] # 供应商归属冲突（模型已挂别的供应商）——如实报告不静默迁移
|
||
for m in models:
|
||
vmid = (m.get("vendor_model_id") or "").strip()
|
||
if not vmid:
|
||
skipped.append("(缺 vendor_model_id)")
|
||
continue
|
||
cap = m.get("capability") or "t2t"
|
||
sync_mode = "async" if str(m.get("sync_mode") or "").strip() == "async" else "sync"
|
||
desc = (m.get("description", "") or "").strip()
|
||
# 出处标注（2026-09-05 用户规则）：模型注册表描述字段必须保存文档 URL
|
||
if doc_url and ("[出处:" + doc_url + "]") not in desc:
|
||
desc = (desc + " " if desc else "") + "[出处:%s]" % doc_url
|
||
# 异步模型：后续步骤模板链（提交后顺序执行 query→[download]）
|
||
query_ids = []
|
||
if sync_mode == "async":
|
||
steps = spec.get("async_steps") or []
|
||
if not any((s.get("purpose") or "") == "query" for s in steps):
|
||
desc = "⚠️异步模型但文档未提取到查询步骤（query_profile_ids 需人工补录）。" + desc
|
||
for s in steps:
|
||
qid = await _ensure_step_profile(sor, protocol, cap, vmid, s, spec)
|
||
if qid:
|
||
query_ids.append(qid)
|
||
query_ids_json = json.dumps(query_ids) if query_ids else ""
|
||
# 幂等键=name（uk_llm_model_name 全局唯一）。不能按 (vendor_id, vmid) 查：
|
||
# 模型换挂供应商后（如迁到阿里云百炼），旧查询找不到→INSERT→撞唯一键崩（实测）。
|
||
recs = await sor.sqlExe(
|
||
"SELECT id, vendor_id FROM llm_model WHERE name=${m}$ OR vendor_model_id=${m}$",
|
||
{"m": vmid})
|
||
await sor.sqlExe("COMMIT", {})
|
||
if recs:
|
||
mid = getattr(recs[0], "id", "")
|
||
cur_vendor = getattr(recs[0], "vendor_id", "") or ""
|
||
if cur_vendor and cur_vendor != vendor_id:
|
||
# 供应商归属不一致：治理决策，不静默迁移——只更新元数据并如实报告
|
||
conflicts.append(
|
||
"%s: 模型已存在且挂供应商 id=%s，与本次规格供应商（%s）不一致——"
|
||
"未改供应商归属，仅更新描述/同步模式；如需迁移请明确指示"
|
||
% (vmid, cur_vendor, vendor_name))
|
||
await sor.sqlExe(
|
||
"UPDATE llm_model SET description=${d}$, "
|
||
"sync_mode=${sm}$, query_profile_ids=${q}$, updated_at=NOW() "
|
||
"WHERE id=${i}$",
|
||
{"d": desc, "sm": sync_mode, "q": query_ids_json, "i": mid})
|
||
await sor.sqlExe("COMMIT", {})
|
||
updated.append(vmid)
|
||
continue
|
||
mid = getID()
|
||
await sor.C("llm_model", {
|
||
"id": mid, "vendor_id": vendor_id, "account_id": "",
|
||
"name": vmid, "vendor_model_id": vmid, "capability": cap,
|
||
"sync_mode": sync_mode, "profile_id": profile_ids.get(cap, ""),
|
||
"query_profile_ids": query_ids_json,
|
||
"ppid": "",
|
||
"default_params": "{}", "status": "active",
|
||
"description": desc, "org_id": "0"})
|
||
created.append(vmid)
|
||
|
||
rt_note = ""
|
||
if protocol not in RUNTIME_PROTOCOLS:
|
||
rt_note = ("⚠️ 协议「%s」的适配模板已存档，但当前运行时调用链仅支持 %s——"
|
||
"该供应商模型暂不可被产线直接调用" % (protocol, "/".join(RUNTIME_PROTOCOLS)))
|
||
audit_note = await _audit_media_convention(sor, list(profile_ids.values()))
|
||
return json.dumps({
|
||
"vendor": vendor_action, "profiles": profile_ids,
|
||
"models_created": created, "models_updated": updated, "models_skipped": skipped,
|
||
"vendor_conflicts": conflicts,
|
||
"template_errors": tpl_errors,
|
||
"template_notes": tpl_notes,
|
||
"runtime_note": rt_note,
|
||
"media_audit": audit_note,
|
||
}, ensure_ascii=False)
|
||
|
||
|
||
# 生成类能力：出参是文件（图/视频/音频/3D），response 必须 downloadfile2url 落地
|
||
_MEDIA_CAPS = ('t2i', 'i2v', 't2v', 't2a', 'tts', 'i2i', 'v2v', '3d')
|
||
|
||
|
||
async def _audit_media_convention(sor, profile_ids):
|
||
"""强制检查（2026-09-05 用户规则：每个 llm 配置都要检查媒体转换约定）：
|
||
|
||
生成类能力的适配模板——
|
||
request_template 含上传媒体参数时，必须用 {{b64media2url(request, xxx_file)}}
|
||
转本地公网 URL 再传上游；
|
||
response_template 必须用 {{downloadfile2url(request, <产物url>)}}
|
||
把生成物落地为本地持久 URL（上游 URL 有效期短，视频仅 24 小时）。
|
||
返回检查结论（无问题为空串），随 apply 结果返回给内部 agent 转告。
|
||
"""
|
||
if not profile_ids:
|
||
return ""
|
||
issues = []
|
||
for pid in profile_ids:
|
||
recs = await sor.sqlExe(
|
||
"SELECT name, capability, request_template, response_template "
|
||
"FROM llm_api_profile WHERE id=${i}$", {"i": pid})
|
||
await sor.sqlExe("COMMIT", {})
|
||
if not recs:
|
||
continue
|
||
r = recs[0]
|
||
cap = getattr(r, 'capability', '') or ''
|
||
name = getattr(r, 'name', '') or pid
|
||
if cap not in _MEDIA_CAPS:
|
||
continue # 非生成类（t2t/embedding/rerank）无生成物
|
||
rt = getattr(r, 'request_template', '') or ''
|
||
st = getattr(r, 'response_template', '') or ''
|
||
if '__note__' in st and 'downloadfile2url' not in st:
|
||
issues.append("「%s」response 模板是骨架——生成物必须用 "
|
||
"downloadfile2url(request, <产物url>) 落地后再返回" % name)
|
||
elif st and 'downloadfile2url' not in st:
|
||
issues.append("「%s」response 模板缺 downloadfile2url——生成类能力"
|
||
"（%s）的产物 URL 必须落地为本地持久 URL" % (name, cap))
|
||
# 上传参数线索：模板里出现 image/video/audio 文件参数但没有 b64media2url
|
||
has_upload_ref = any(k in rt for k in ('image_file', 'video_file', 'audio_file',
|
||
'first_frame', 'image_url', 'media'))
|
||
if has_upload_ref and 'b64media2url' not in rt:
|
||
issues.append("「%s」request 模板含媒体上传参数但缺 b64media2url——"
|
||
"上传文件必须转本地公网 URL 再传上游" % name)
|
||
if issues:
|
||
return "⚠️ 媒体转换约定检查未通过：" + "；".join(issues)
|
||
return ""
|
||
|
||
|
||
# 能力 → 统一出参键（同类能力对外契约一致：视频出 video、图出 image）
|
||
_MEDIA_OUT_KEY = {
|
||
't2v': 'video', 'i2v': 'video', 'v2v': 'video',
|
||
't2i': 'image', 'i2i': 'image',
|
||
'tts': 'audio', 't2a': 'audio',
|
||
'3d': 'glb',
|
||
}
|
||
# 能力 → 上游产物字段名兜底（文档没给响应示例时用供应商惯例）
|
||
_MEDIA_OUT_FIELD = {
|
||
'video': 'video_url', 'image': 'image_url', 'audio': 'audio_url', 'glb': 'model_url',
|
||
}
|
||
# 文档示例里表示「上传媒体」的结构线索
|
||
_MEDIA_TYPES = ('first_frame', 'last_frame', 'ref_image', 'ref_images', 'image', 'img')
|
||
_MEDIA_KEY_HINTS = ('img_url', 'image_url', 'first_frame_url', 'last_frame_url',
|
||
'ref_image_url', 'video_url', 'audio_url')
|
||
# 需要「输入媒体」的能力（这些能力里出现 URL 叶子才判定为上传媒体）
|
||
_MEDIA_INPUT_CAPS = ('i2v', 'i2i', 'v2v', 'i2t', '2i2v')
|
||
|
||
|
||
def _media_param_name(typeval, keyname):
|
||
"""按文档媒体结构推断运行时统一媒体参数名（xxx_file 契约）。"""
|
||
t = ('%s %s' % (typeval or '', keyname or '')).lower()
|
||
if 'video' in t:
|
||
return 'video_file'
|
||
if 'audio' in t:
|
||
return 'audio_file'
|
||
return 'image_file'
|
||
|
||
|
||
def _is_url_example(v):
|
||
return isinstance(v, str) and v.strip().lower().startswith(('http://', 'https://'))
|
||
|
||
|
||
def _example_to_nested(fields):
|
||
"""request_fields 点号路径 → 嵌套 dict（文档无请求示例时的兜底结构来源）。"""
|
||
root = {}
|
||
for f in fields or []:
|
||
parts = [p for p in str(f).split('.') if p]
|
||
if not parts:
|
||
continue
|
||
cur = root
|
||
for p in parts[:-1]:
|
||
nxt = cur.get(p)
|
||
if not isinstance(nxt, dict):
|
||
nxt = {}
|
||
cur[p] = nxt
|
||
cur = nxt
|
||
cur[parts[-1]] = ''
|
||
return root
|
||
|
||
|
||
def _tpl_from_example(example, capability):
|
||
"""文档请求示例 → Jinja2 请求体模板（结构逐字保留，值换成运行时变量）。
|
||
|
||
规则（2026-09-05，替代旧的硬编码骨架）：
|
||
model 键 → {{ model|tojson }}
|
||
prompt/text 键 → {{ prompt|tojson }}
|
||
上传媒体（示例里是 URL 叶子，且结构/键名带媒体线索或能力属输入媒体类）
|
||
→ {{ b64media2url(request, params.<media>_file)|tojson }}
|
||
其他标量 → {{ params.<key>|default(<文档示例值>)|tojson }}
|
||
**示例里没有上传媒体就不生成媒体参数**——纯文生视频(t2v)曾因硬编码
|
||
params.xxx_file 骨架在运行时崩（'dict object' has no attribute 'xxx_file'）。
|
||
|
||
返回 (模板串, media_params, biz_params, used_example)。
|
||
"""
|
||
tokens = {}
|
||
media_params, biz_params = [], []
|
||
used_example = isinstance(example, dict) and bool(example)
|
||
tree = example if used_example else _example_to_nested(None)
|
||
|
||
def _put(expr):
|
||
k = '__TPL%d__' % len(tokens)
|
||
tokens[k] = expr
|
||
return k
|
||
|
||
def walk(node, parent_key='', sib_type=None):
|
||
if isinstance(node, list):
|
||
return [walk(v, parent_key, sib_type) for v in node]
|
||
if not isinstance(node, dict):
|
||
return node
|
||
tval = node.get('type') if isinstance(node.get('type'), str) else sib_type
|
||
out = {}
|
||
for k, v in node.items():
|
||
if isinstance(v, (dict, list)):
|
||
out[k] = walk(v, k, tval)
|
||
continue
|
||
kl = str(k).lower()
|
||
if kl == 'model' and isinstance(v, str):
|
||
out[k] = _put('{{ model|tojson }}')
|
||
continue
|
||
if kl in ('prompt', 'text') and isinstance(v, str):
|
||
out[k] = _put('{{ prompt|tojson }}')
|
||
continue
|
||
is_media = _is_url_example(v) and (
|
||
str(tval or '').lower() in _MEDIA_TYPES
|
||
or kl in _MEDIA_KEY_HINTS
|
||
or capability in _MEDIA_INPUT_CAPS)
|
||
if is_media:
|
||
mp = _media_param_name(tval, k)
|
||
if mp not in media_params:
|
||
media_params.append(mp)
|
||
out[k] = _put('{{ b64media2url(request, params.%s)|tojson }}' % mp)
|
||
continue
|
||
biz_params.append({'name': k, 'example': v})
|
||
if v == '' or v is None:
|
||
out[k] = _put('{{ params.%s|tojson }}' % k) # 无示例值：调用必传
|
||
else:
|
||
out[k] = _put('{{ params.%s|default(%s)|tojson }}'
|
||
% (k, json.dumps(v, ensure_ascii=False)))
|
||
return out
|
||
|
||
body = walk(tree)
|
||
s = json.dumps(body, ensure_ascii=False)
|
||
for k, expr in tokens.items():
|
||
s = s.replace('"%s"' % k, expr)
|
||
return s, media_params, biz_params, used_example
|
||
|
||
|
||
def _collect_url_fields(node, prefix=''):
|
||
"""递归收集响应示例里的 URL 字段点号路径（如 output.video_url）。"""
|
||
hits = []
|
||
if isinstance(node, dict):
|
||
for k, v in node.items():
|
||
p = (prefix + '.' + str(k)) if prefix else str(k)
|
||
if isinstance(v, (dict, list)):
|
||
hits.extend(_collect_url_fields(v, p))
|
||
elif _is_url_example(v):
|
||
hits.append(p)
|
||
elif isinstance(node, list):
|
||
for i, v in enumerate(node[:3]):
|
||
hits.extend(_collect_url_fields(v, prefix))
|
||
return hits
|
||
|
||
|
||
def _resp_tpl_for(capability, spec):
|
||
"""生成类响应模板：产物 URL 经 downloadfile2url 落地（上游 URL 有效期短）。
|
||
|
||
产物字段路径优先取文档响应示例里的 URL 字段（递归，支持 output.video_url
|
||
这类嵌套——运行时 ns 有 output/usage/task_id），缺则用供应商惯例兜底，
|
||
并在 notes 里如实说明来源（不假装是文档确证）。
|
||
"""
|
||
outkey = _MEDIA_OUT_KEY.get(capability, 'video')
|
||
field = ''
|
||
rex = spec.get('response_example')
|
||
if isinstance(rex, dict):
|
||
hits = _collect_url_fields(rex)
|
||
# 优先 *_url 结尾且与产物类型匹配的字段
|
||
want = ('video', 'image', 'audio', 'glb', 'model')
|
||
for h in hits:
|
||
leaf = h.split('.')[-1].lower()
|
||
if leaf.endswith('_url') and any(w in leaf for w in want):
|
||
field = h
|
||
break
|
||
if not field and hits:
|
||
field = hits[0]
|
||
note = ''
|
||
if not field:
|
||
field = _MEDIA_OUT_FIELD.get(outkey, 'result_url')
|
||
note = ('产物字段名「%s」按供应商惯例兜底（文档未提供响应示例或示例中无 URL 字段），'
|
||
'如上游字段不同需修正' % field)
|
||
resp = json.dumps({
|
||
'status': 'SUCCEEDED',
|
||
outkey: '{{ downloadfile2url(request, %s) }}' % field,
|
||
'usage': '{{ json.dumps(usage) }}',
|
||
'task_id': '{{ task_id }}',
|
||
}, ensure_ascii=False)
|
||
return resp, note
|
||
|
||
|
||
def _headers_from_spec(spec):
|
||
"""请求头：认证头统一 api_key 变量，其余按文档示例逐字带上
|
||
（如 DashScope 异步必需的 X-DashScope-Async: enable）。"""
|
||
headers = {'Authorization': 'Bea' + 'rer {{api_key}}',
|
||
'Content-Type': 'application/json'}
|
||
rh = spec.get('request_headers')
|
||
if isinstance(rh, dict):
|
||
for k, v in rh.items():
|
||
if str(k).lower() in ('authorization', 'content-type'):
|
||
continue
|
||
if isinstance(v, str) and v.strip():
|
||
headers[k] = v.strip()
|
||
# 协议兜底：dashscope_async 提交必须带异步开关头（文档示例遗漏也不至于提交即失败）
|
||
if str(spec.get('protocol') or '').strip() == 'dashscope_async':
|
||
headers.setdefault('X-DashScope-Async', 'enable')
|
||
return headers
|
||
|
||
|
||
def _gen_templates(protocol: str, capability: str, spec: dict) -> dict:
|
||
"""按文档示例动态生成适配模板（path/headers/request/response/param_schema）。
|
||
|
||
返回 {path, headers, req, resp, param_schema, media_params, biz_params, notes}。
|
||
"""
|
||
chat_path = (spec.get("chat_path") or "/chat/completions").strip()
|
||
headers = _headers_from_spec(spec)
|
||
notes = []
|
||
if protocol == "openai_compat":
|
||
req = json.dumps({
|
||
"model": "{{model}}", "messages": "{{messages}}",
|
||
"temperature": "{{temperature}}", "stream": False,
|
||
}, ensure_ascii=False)
|
||
resp = json.dumps({
|
||
"content": "choices[0].message.content",
|
||
"usage": {"prompt_tokens": "usage.prompt_tokens",
|
||
"completion_tokens": "usage.completion_tokens"},
|
||
}, ensure_ascii=False)
|
||
return {"path": chat_path, "headers": headers, "req": req, "resp": resp,
|
||
"param_schema": [{"name": "prompt", "label": "提示词",
|
||
"uitype": "textarea", "required": True}],
|
||
"media_params": [], "biz_params": [], "notes": notes}
|
||
|
||
example = spec.get("request_example")
|
||
if not (isinstance(example, dict) and example):
|
||
example = _example_to_nested(spec.get("request_fields"))
|
||
notes.append("文档未提供完整请求示例，模板按 request_fields 字段路径生成，"
|
||
"业务参数无默认值（调用时必须显式传入），建议人工核对")
|
||
req, media_params, biz_params, _used = _tpl_from_example(example, capability)
|
||
if capability in _MEDIA_CAPS:
|
||
resp, rnote = _resp_tpl_for(capability, spec)
|
||
if rnote:
|
||
notes.append(rnote)
|
||
if not media_params and capability in _MEDIA_INPUT_CAPS:
|
||
notes.append("能力 %s 通常需要输入媒体，但文档示例未见上传字段——"
|
||
"模板未生成媒体参数，请核对文档" % capability)
|
||
else:
|
||
resp = json.dumps({"content": "{{ text }}"}, ensure_ascii=False)
|
||
schema = [{"name": "prompt", "label": "提示词", "uitype": "textarea",
|
||
"required": capability not in ('embedding', 'rerank')}]
|
||
for mp in media_params:
|
||
schema.append({"name": mp, "label": "上传媒体（%s）" % mp,
|
||
"uitype": "file", "required": True})
|
||
for b in biz_params:
|
||
schema.append({"name": b['name'], "label": b['name'],
|
||
"uitype": "number" if isinstance(b['example'], (int, float)) else "text",
|
||
"required": False, "default": b['example']})
|
||
return {"path": chat_path, "headers": headers, "req": req, "resp": resp,
|
||
"param_schema": schema, "media_params": media_params,
|
||
"biz_params": biz_params, "notes": notes}
|
||
|
||
|
||
# 旧骨架模板标记：含这些标记的 profile 视为不可用，apply 时按文档自愈重建
|
||
_SKELETON_MARKERS = ('__from_doc__', '__note__', 'xxx_file')
|
||
|
||
|
||
def _dry_render_check(req_tpl, capability, biz_params, media_params):
|
||
"""落库前干跑渲染（StrictUndefined）：模板引用了运行时拿不到的变量当场报错。
|
||
|
||
运行时命名空间见 pipeline-llm inference._build_async_body：
|
||
model / prompt / messages / params(业务参数) / api_key / org_id
|
||
+ request / json / b64media2url / downloadfile2url。
|
||
旧缺陷正是漏了这步：模板硬编码 params.xxx_file，直到 test_model_call 才崩。
|
||
"""
|
||
try:
|
||
from jinja2 import Environment, StrictUndefined
|
||
except Exception as e:
|
||
return "干跑校验跳过（jinja2 不可用：%s）" % str(e)[:60]
|
||
|
||
def _stub(request, value, *a, **k):
|
||
return str(value) # StrictUndefined 传入未定义值时在此抛错
|
||
|
||
params = {}
|
||
for b in biz_params or []:
|
||
# 干跑模拟「业务参数全部提供」场景：无示例值的参数给占位串
|
||
# （模板 {{ params.x|tojson }} 无默认——运行时必须显式传，属预期契约）
|
||
params[b['name']] = b['example'] if b.get('example') not in ('', None) else 'dry'
|
||
for mp in media_params or []:
|
||
params[mp] = 'https://dry-run.invalid/sample.bin'
|
||
ns = {'model': 'dry-run-model', 'prompt': '干跑校验', 'messages': [],
|
||
'params': params, 'api_key': 'sk-dry-run', 'org_id': '0',
|
||
'request': None, 'json': json,
|
||
'b64media2url': _stub, 'downloadfile2url': _stub}
|
||
try:
|
||
env = Environment(undefined=StrictUndefined)
|
||
out = env.from_string(req_tpl).render(**ns)
|
||
json.loads(out)
|
||
except Exception as e:
|
||
return ("请求模板干跑渲染失败（落库前拦截）：%s: %s——模板引用了运行时不存在的"
|
||
"变量（业务参数可用：%s）" % (type(e).__name__, str(e)[:160],
|
||
sorted(params.keys()) or '无'))
|
||
return ""
|
||
|
||
|
||
async def _ensure_step_profile(sor, protocol, cap, vmid, step, spec):
|
||
"""异步后续步骤的适配模板（query/download）。
|
||
|
||
2026-09-05 改造（用户：同类任务查询接口供应商级相同，可复用）：
|
||
幂等键=名称「{protocol}-query-{path}」（去掉模型维度）——同供应商同协议
|
||
下多个模型共享同一份查询模板；旧的按 {vmid}-{purpose} 命名继续兼容复用。
|
||
模板按文档真实生成：运行时查询只用 path/headers/method，task_id 经
|
||
path 占位（如 /tasks/{{task_id}}）；不再产 __from_doc__ 骨架。
|
||
返回 profile id。
|
||
"""
|
||
purpose = (step.get("purpose") or "query").strip() or "query"
|
||
path = (step.get("path") or "").strip()
|
||
method = (step.get("method") or ("GET" if purpose == "query" else "POST")).strip().upper()
|
||
shared_name = "%s-%s-%s" % (protocol, purpose, path or "default")
|
||
legacy_name = "%s-%s" % (vmid, purpose)
|
||
recs = await sor.sqlExe(
|
||
"SELECT id FROM llm_api_profile WHERE name=${n}$ AND status='active' LIMIT 1",
|
||
{"n": shared_name})
|
||
await sor.sqlExe("COMMIT", {})
|
||
if recs:
|
||
return getattr(recs[0], "id", "")
|
||
recs = await sor.sqlExe(
|
||
"SELECT id FROM llm_api_profile WHERE name=${n}$ AND status='active' LIMIT 1",
|
||
{"n": legacy_name})
|
||
await sor.sqlExe("COMMIT", {})
|
||
if recs:
|
||
return getattr(recs[0], "id", "")
|
||
headers = _headers_from_spec(spec)
|
||
if method == "GET":
|
||
req = json.dumps({"method": method}, ensure_ascii=False)
|
||
if "{{task_id}}" not in path:
|
||
# path 没带任务号占位时按 DashScope 惯例拼 /tasks/{id}；notes 如实标注
|
||
path = (path.rstrip("/") + "/{{task_id}}") if path else "/tasks/{{task_id}}"
|
||
else:
|
||
req = json.dumps({"method": method,
|
||
"data": {"task_id": "{{task_id}}"}}, ensure_ascii=False)
|
||
# 查询步骤的响应渲染在运行时由提交模板的 response_template 接管
|
||
# （inference 轮询直接调 _render_response(submit_profile,...)），这里存档即可
|
||
resp = json.dumps({
|
||
"__note__": "查询步骤模板：运行时响应解析走模型提交模板的 response_template",
|
||
"response_format": step.get("response_format", "") or "",
|
||
}, ensure_ascii=False)
|
||
pid = getID()
|
||
await sor.C("llm_api_profile", {
|
||
"id": pid, "name": shared_name, "protocol": protocol, "capability": cap,
|
||
"path": path,
|
||
"headers": json.dumps(headers, ensure_ascii=False),
|
||
"request_template": req, "response_template": resp,
|
||
"param_schema": "", "status": "active"})
|
||
return pid
|
||
|
||
|
||
# ────────────────────── 工具 5：定价自动导入 ──────────────────────
|
||
|
||
_FACTOR_LABELS = {
|
||
'duration': '时长', 'flat': '按次', 'prompt_tokens': '输入tokens',
|
||
'completion_tokens': '输出tokens',
|
||
}
|
||
|
||
|
||
def _build_pricing_yaml(currency_items, dimensions_all, factor):
|
||
"""生成定价 YAML（2026-09-05 用户定夺的新模式）：
|
||
|
||
- 不放 model 过滤：一个定价方案只服务一个模型；定价相同的多个模型共享
|
||
同一 ppid（不是往定价里加 model 维度）
|
||
- 不用 filters 子结构：维度直接平铺在定价项（引擎对非保留键做 AND 匹配）
|
||
- fields 需定义每个维度（role: filter），否则引擎报「在fields中没有定义」
|
||
"""
|
||
fields = {
|
||
'price_factors': {'type': 'string', 'role': 'factor', 'label': '计价因子'},
|
||
'unit_prices': {'type': 'float', 'role': 'factor', 'label': '单位定价'},
|
||
'unit': {'type': 'string', 'role': 'factor', 'label': '计价单位'},
|
||
}
|
||
if factor not in fields:
|
||
fields[factor] = {'type': 'float', 'role': 'factor',
|
||
'label': _FACTOR_LABELS.get(factor, factor)}
|
||
for dim in sorted(dimensions_all):
|
||
fields[dim] = {'type': 'string', 'role': 'filter', 'label': dim}
|
||
pricings = []
|
||
for it in currency_items:
|
||
item = {'price_factors': factor, 'unit_prices': it['unit_price'],
|
||
'unit': it['unit']}
|
||
for dk, dv in sorted((it.get('dimensions') or {}).items()):
|
||
item[dk] = dv
|
||
pricings.append(item)
|
||
units = sorted(set(it['unit'] for it in currency_items))
|
||
unit_values = {u: 1 for u in units}
|
||
if '百万' in unit_values:
|
||
unit_values['百万'] = 1000000
|
||
if '千' in unit_values:
|
||
unit_values['千'] = 1000
|
||
doc = {'unit_values': unit_values, 'fields': fields, 'pricings': pricings}
|
||
return yaml.dump(doc, allow_unicode=True, sort_keys=False)
|
||
|
||
|
||
async def _h_apply_model_pricing(sor, params, ctx):
|
||
"""定价自动导入：按提取规格建定价方案并挂模型 ppid（幂等）。
|
||
|
||
幂等键：模型已有 ppid 且方案描述含同一出处 URL → 更新时序（拉链）；
|
||
否则新建方案。价格只来自文档原文（提取层强制 doc_quote）。
|
||
|
||
规格来源（2026-09-05 防编造改造）：use_last_extract=true → 取本会话
|
||
extract 锚定规格（含 pricing/doc_url），禁止 LLM 复制转述大 JSON。
|
||
"""
|
||
err = await _require_owner(sor, ctx)
|
||
if err:
|
||
return err
|
||
use_last = str(params.get("use_last_extract") or "").lower() in ("true", "1", "yes")
|
||
if use_last:
|
||
spec = _spec_load(ctx)
|
||
if spec is None:
|
||
return ("本会话没有锚定的提取规格——先调 extract_llm_api_spec（成功后规格自动"
|
||
"锚定本会话，含 pricing 与 doc_url），再传 {\"use_last_extract\": true}。"
|
||
"会话级隔离：其他会话的规格本会话不可见。")
|
||
else:
|
||
try:
|
||
spec = json.loads(params.get("spec") or "{}")
|
||
except Exception:
|
||
return "spec 不是合法 JSON"
|
||
if not isinstance(spec, dict) or "pricing" not in spec:
|
||
return ("spec 结构不符——不要自己编写规格。正确做法：extract_llm_api_spec 后传 "
|
||
"{\"use_last_extract\": true}（pricing 与 doc_url 已锚定本会话）。")
|
||
pricing = spec.get("pricing") or {}
|
||
items = pricing.get("items") or []
|
||
if not items:
|
||
return "spec.pricing.items 为空——文档没提取到价格（禁止编造），如需定价请先核对文档"
|
||
doc_url = (spec.get("doc_url") or "").strip()
|
||
currency = (pricing.get("currency") or "CNY").strip() or "CNY"
|
||
# 校验：每条必须有 doc_quote（出处）；dimensions 禁止 model
|
||
for it in items:
|
||
if not (it.get("doc_quote") or "").strip():
|
||
return "定价条目缺 doc_quote（文档原文定价句）——定价只允许来自文档原文，拒绝落库"
|
||
if "model" in (it.get("dimensions") or {}):
|
||
return "定价 dimensions 含 model——一个定价方案只服务一个模型，禁止 model 维度；定价相同的模型共享同一 ppid"
|
||
# 按 vendor_model_id 分组（一个模型一个定价方案）
|
||
by_model = {}
|
||
for it in items:
|
||
vmid = (it.get("vendor_model_id") or "").strip()
|
||
if not vmid:
|
||
return "定价条目缺 vendor_model_id"
|
||
by_model.setdefault(vmid, []).append(it)
|
||
|
||
results = []
|
||
for vmid, mitems in by_model.items():
|
||
# 找模型
|
||
recs = await sor.sqlExe(
|
||
"SELECT id, name, ppid FROM llm_model WHERE vendor_model_id=${m}$ "
|
||
"AND status='active' LIMIT 1", {"m": vmid})
|
||
await sor.sqlExe("COMMIT", {})
|
||
if not recs:
|
||
results.append({"model": vmid, "ok": False,
|
||
"error": "模型未注册（先 apply_llm_config）"})
|
||
continue
|
||
model_id = getattr(recs[0], "id", "")
|
||
old_ppid = getattr(recs[0], "ppid", "") or ""
|
||
factors = sorted(set((it.get("factor") or "flat") for it in mitems))
|
||
if len(factors) != 1:
|
||
results.append({"model": vmid, "ok": False,
|
||
"error": "同一模型混用多个计价因子 %s——请拆分为多个定价方案" % factors})
|
||
continue
|
||
factor = factors[0]
|
||
dims_all = set()
|
||
for it in mitems:
|
||
dims_all.update((it.get("dimensions") or {}).keys())
|
||
yaml_str = _build_pricing_yaml(mitems, dims_all, factor)
|
||
# 引擎试算护栏：渲染一遍确认合法 YAML + fields 完整（不落库先验证）
|
||
try:
|
||
parsed = yaml.safe_load(yaml_str)
|
||
assert parsed.get('pricings') and parsed.get('fields')
|
||
except Exception as e:
|
||
results.append({"model": vmid, "ok": False,
|
||
"error": "定价 YAML 生成校验失败：%s" % str(e)[:150]})
|
||
continue
|
||
desc_quotes = "；".join((it.get("doc_quote") or "")[:80] for it in mitems)[:500]
|
||
pp_desc = "定价出处：%s | 文档原文：%s" % (doc_url or "(未提供)", desc_quotes)
|
||
now = time.strftime('%Y-%m-%d')
|
||
if old_ppid:
|
||
ppid = old_ppid
|
||
# 幂等核心（2026-09-05 用户纠正：时序应只有一条有效行，重复执行不得堆历史）：
|
||
# 先取当前生效行比对内容——
|
||
# 内容相同 → 跳过（不产生新时序行）
|
||
# 内容不同且生效行今天才启用 → 原地 UPDATE（同日替换无历史区间可保留，
|
||
# 拉链会产生 enabled==expired 的零宽死行，纯垃圾）
|
||
# 内容不同且生效行是历史日期 → 正常拉链（关旧行+插新行）
|
||
recs2 = await sor.sqlExe(
|
||
"SELECT id, pricing_data, enabled_date FROM pricing_program_timing "
|
||
"WHERE ppid=${p}$ AND expired_date='9999-12-31' "
|
||
"ORDER BY enabled_date DESC LIMIT 1", {"p": ppid})
|
||
await sor.sqlExe("COMMIT", {})
|
||
cur = recs2[0] if recs2 else None
|
||
same = False
|
||
if cur is not None:
|
||
try:
|
||
same = (yaml.safe_load(getattr(cur, 'pricing_data', '') or '{}')
|
||
== yaml.safe_load(yaml_str))
|
||
except Exception:
|
||
same = (getattr(cur, 'pricing_data', '') or '') == yaml_str
|
||
await sor.sqlExe(
|
||
"UPDATE pricing_program SET description=${d}$ WHERE id=${i}$",
|
||
{"d": pp_desc[:1000], "i": ppid})
|
||
await sor.sqlExe("COMMIT", {})
|
||
if same:
|
||
results.append({"model": vmid, "ok": True, "ppid": ppid,
|
||
"action": "无变化（幂等跳过，未产生新时序行）",
|
||
"items": len(mitems), "factor": factor,
|
||
"dimensions": sorted(dims_all)})
|
||
continue
|
||
if cur is not None and str(getattr(cur, 'enabled_date', ''))[:10] == now:
|
||
await sor.sqlExe(
|
||
"UPDATE pricing_program_timing SET pricing_data=${y}$, name=${n}$ "
|
||
"WHERE id=${i}$",
|
||
{"y": yaml_str, "n": "%s %s计价" % (vmid, factor),
|
||
"i": getattr(cur, 'id', '')})
|
||
await sor.sqlExe("COMMIT", {})
|
||
action = "原地更新（同日替换，不产生历史行）"
|
||
results.append({"model": vmid, "ok": True, "ppid": ppid,
|
||
"action": action, "items": len(mitems),
|
||
"factor": factor, "dimensions": sorted(dims_all)})
|
||
continue
|
||
# 历史生效行 → 拉链
|
||
await sor.sqlExe(
|
||
"UPDATE pricing_program_timing SET expired_date=${d}$ "
|
||
"WHERE ppid=${p}$ AND expired_date='9999-12-31'",
|
||
{"d": now, "p": ppid})
|
||
action = "更新（拉链：旧行截至今日，新时序生效 %s）" % now
|
||
else:
|
||
ppid = getID()
|
||
await sor.C("pricing_program", {
|
||
"id": ppid, "name": "%s 定价" % vmid, "ownerid": "0",
|
||
"providerid": "", "pricing_belong": "",
|
||
"description": pp_desc[:1000], "currency": currency})
|
||
action = "新建"
|
||
await sor.C("pricing_program_timing", {
|
||
"id": getID(), "ppid": ppid, "name": "%s %s计价" % (vmid, factor),
|
||
"pricing_data": yaml_str,
|
||
"enabled_date": now, "expired_date": "9999-12-31"})
|
||
await sor.sqlExe("COMMIT", {})
|
||
# 挂模型
|
||
await sor.sqlExe("UPDATE llm_model SET ppid=${p}$ WHERE id=${i}$",
|
||
{"p": ppid, "i": model_id})
|
||
await sor.sqlExe("COMMIT", {})
|
||
results.append({"model": vmid, "ok": True, "ppid": ppid,
|
||
"action": action, "items": len(mitems),
|
||
"factor": factor,
|
||
"dimensions": sorted(dims_all)})
|
||
ok_n = sum(1 for r in results if r.get("ok"))
|
||
return json.dumps({
|
||
"summary": "定价导入：%d/%d 个模型成功" % (ok_n, len(results)),
|
||
"results": results,
|
||
"convention": "一个定价方案只服务一个模型（无 model 维度/无 filters，维度平铺）；"
|
||
"定价完全相同的模型共享同一 ppid",
|
||
}, ensure_ascii=False)
|
||
|
||
|
||
# ────────────────────── 工具 6：模型测试（真实调用） ──────────────────────
|
||
|
||
async def _h_test_model_call(sor, params, ctx):
|
||
"""真实调用测试：走完整治理链（门禁→上游→结算），返回调用结果+流水ID。
|
||
|
||
参数：model_name 必填；prompt 文本提示（t2t）；业务参数放 params JSON
|
||
（如 image_file/resolution/duration）；timeout 等待秒数（异步模型默认 600）。
|
||
前置：供应商账号已配 api_key（无 key 会报可行动错误，不静默）。
|
||
"""
|
||
err = await _require_owner(sor, ctx)
|
||
if err:
|
||
return err
|
||
model_name = (params.get("model_name") or "").strip()
|
||
if not model_name:
|
||
return "缺 model_name"
|
||
prompt = (params.get("prompt") or "回复两个字：正常").strip()
|
||
try:
|
||
extra = json.loads(params.get("params") or "{}")
|
||
except Exception:
|
||
return "params 不是合法 JSON"
|
||
timeout_s = int(_f(params.get("timeout"), 0)) or 600
|
||
try:
|
||
from pipeline_llm.inference import chat_inference
|
||
except Exception as e:
|
||
return "推理模块未加载：%s" % str(e)[:100]
|
||
payload = {"model": model_name,
|
||
"messages": [{"role": "user", "content": prompt}]}
|
||
payload.update(extra)
|
||
if timeout_s:
|
||
payload["_timeout"] = timeout_s
|
||
task_ref = "agent-test:%s:%d" % (model_name[:20], int(time.time()))
|
||
t0 = time.time()
|
||
try:
|
||
data = await chat_inference(ctx.get("org_id", "") or "0",
|
||
ctx.get("user_id", ""), payload,
|
||
model_name=model_name, task_ref=task_ref)
|
||
except Exception as e:
|
||
return json.dumps({
|
||
"ok": False, "model": model_name, "task_ref": task_ref,
|
||
"elapsed_sec": round(time.time() - t0, 1),
|
||
"error": str(e)[:400],
|
||
"hint": "常见原因：账号无api_key/余额不足/端点超时/模板配置缺失——按错误消息处置",
|
||
}, ensure_ascii=False)
|
||
elapsed = round(time.time() - t0, 1)
|
||
content = ""
|
||
try:
|
||
content = data["choices"][0]["message"]["content"]
|
||
except Exception:
|
||
content = str(data)[:200]
|
||
out = {
|
||
"ok": True, "model": model_name, "task_ref": task_ref,
|
||
"elapsed_sec": elapsed,
|
||
"content_head": str(content)[:200],
|
||
"usage": data.get("usage") or {},
|
||
}
|
||
if data.get("media"):
|
||
out["media"] = data["media"] # 生成物本地持久 URL（已落地）
|
||
if data.get("task_id"):
|
||
out["task_id"] = data["task_id"]
|
||
return json.dumps(out, ensure_ascii=False)
|
||
|
||
|
||
# ────────────────────── 工具 7：记账检查 ──────────────────────
|
||
|
||
async def _h_check_model_accounting(sor, params, ctx):
|
||
"""记账正确性检查（三态：created/accounted/failed）。
|
||
|
||
按 task_ref（测试调用返回的）或 model_name 查最近流水：
|
||
1. 流水存在性 + accounting_status
|
||
2. usages 计价因子完整性
|
||
3. accounted → charge 金额与定价引擎重算对比（独立复算，不信记账侧）
|
||
4. failed → 给出 note 原因与处置建议
|
||
5. created → 出账循环 60 秒一轮，提示等待或查 worker 进程
|
||
"""
|
||
err = await _require_owner(sor, ctx)
|
||
if err:
|
||
return err
|
||
task_ref = (params.get("task_ref") or "").strip()
|
||
model_name = (params.get("model_name") or "").strip()
|
||
if not task_ref and not model_name:
|
||
return "需要 task_ref（测试调用返回值）或 model_name 之一"
|
||
if task_ref:
|
||
recs = await sor.sqlExe(
|
||
"SELECT id, model_id, status, accounting_status, charge, cost, usages, "
|
||
"ppid, note, created_at FROM llm_usage WHERE task_ref=${t}$ "
|
||
"ORDER BY created_at DESC LIMIT 3", {"t": task_ref})
|
||
else:
|
||
recs = await sor.sqlExe(
|
||
"SELECT u.id, u.model_id, u.status, u.accounting_status, u.charge, u.cost, "
|
||
"u.usages, u.ppid, u.note, u.created_at FROM llm_usage u "
|
||
"JOIN llm_model m ON m.id=u.model_id "
|
||
"WHERE m.name=${n}$ OR m.vendor_model_id=${n}$ "
|
||
"ORDER BY u.created_at DESC LIMIT 3", {"n": model_name})
|
||
await sor.sqlExe("COMMIT", {})
|
||
if not recs:
|
||
return "未找到流水（task_ref=%s model=%s）——调用可能没发生或没结算" % (
|
||
task_ref, model_name)
|
||
out_rows = []
|
||
for r in recs:
|
||
row = dict(r)
|
||
st = row.get("accounting_status", "")
|
||
item = {
|
||
"usage_id": row.get("id", ""),
|
||
"call_status": row.get("status", ""),
|
||
"accounting_status": st,
|
||
"charge": float(row.get("charge") or 0),
|
||
"usages": row.get("usages", "") or "",
|
||
"ppid": row.get("ppid", "") or "",
|
||
"created_at": str(row.get("created_at", "")),
|
||
}
|
||
if st == "failed":
|
||
item["verdict"] = "记账失败"
|
||
item["reason"] = (row.get("note") or "")[:200]
|
||
item["hint"] = ("常见处置：模型未挂ppid→apply_model_pricing；"
|
||
"模型未映射产品→产品管理导入产线模型；"
|
||
"定价无匹配档位→核对usages维度值与定价YAML是否一致")
|
||
elif st == "created":
|
||
item["verdict"] = "待记账（出账循环60秒一轮，稍后复查；持续不变查记账worker进程）"
|
||
elif st == "accounted":
|
||
# 独立复算：用 usages 因子过定价引擎，对比 charge
|
||
item["verdict"] = "已记账"
|
||
try:
|
||
usage_data = json.loads(row.get("usages") or "{}")
|
||
except Exception:
|
||
usage_data = {}
|
||
ppid = row.get("ppid") or ""
|
||
if ppid and usage_data:
|
||
try:
|
||
from ahserver.serverenv import ServerEnv
|
||
env = ServerEnv()
|
||
fn = getattr(env, "buffered_charging", None)
|
||
if fn:
|
||
prices = await fn(ppid, usage_data)
|
||
expect = round(sum(float(getattr(p, "amount", 0) or 0)
|
||
for p in (prices or [])), 6)
|
||
item["recomputed_amount"] = expect
|
||
item["amount_match"] = abs(expect - float(row.get("charge") or 0)) < 0.01
|
||
if not item["amount_match"]:
|
||
item["verdict"] = "已记账但金额不符（复算 %.4f ≠ charge %.4f）" % (
|
||
expect, float(row.get("charge") or 0))
|
||
except Exception as e:
|
||
item["recompute_error"] = str(e)[:150]
|
||
else:
|
||
item["verdict"] = "已记账（缺ppid或usages，无法独立复算）"
|
||
out_rows.append(item)
|
||
return json.dumps({"rows": out_rows}, ensure_ascii=False)
|
||
|
||
|
||
# ────────────────────── 工具 8：模块/应用信息查询 ──────────────────────
|
||
|
||
async def _h_platform_modules(sor, params, ctx):
|
||
"""列出平台已装载的业务模块（内部 agent 了解平台构成用）。"""
|
||
err = await _require_owner(sor, ctx)
|
||
if err:
|
||
return err
|
||
mods = []
|
||
for name, title in [
|
||
("pipeline_core", "产线核心（会话/技能/项目）"),
|
||
("pipeline_service", "执行引擎（任务链/角色/治理）"),
|
||
("pipeline-llm", "模型治理（供应商/账号/模型/定价/记账）"),
|
||
("pipeline-bidding", "投标产线"),
|
||
("pipeline-opportunity", "商机产线"),
|
||
("pipeline-sdlc", "开发产线前端"),
|
||
("pipeline-ops", "运维"),
|
||
("pipeline-dist", "分销"),
|
||
("pipeline-task", "任务中心"),
|
||
]:
|
||
try:
|
||
__import__(name.replace("-", "_"))
|
||
loaded = True
|
||
except Exception:
|
||
loaded = False
|
||
mods.append({"module": name, "title": title, "loaded": loaded})
|
||
return json.dumps(mods, ensure_ascii=False)
|
||
|
||
|
||
# ────────────────────── 工具定义 ──────────────────────
|
||
|
||
PLATFORM_TOOLS = [
|
||
ToolDefinition(
|
||
name="platform_llm_status",
|
||
description="查看模型治理状态：供应商/账号余额/模型定价/各状态分布。用户问「模型配置状态/账号余额」时调用。仅owner组织角色可用。",
|
||
parameters={},
|
||
category="platform",
|
||
),
|
||
ToolDefinition(
|
||
name="fetch_model_doc",
|
||
description="抓取大模型供应商的 API 文档页面（返回纯文本）。配置新模型前先用此抓取官方文档。仅owner组织角色可用。",
|
||
parameters={"url": "文档页面 URL（必须是公网 http/https）"},
|
||
category="platform",
|
||
),
|
||
ToolDefinition(
|
||
name="extract_llm_api_spec",
|
||
description="通读文档文本，LLM 提取 API 配置规格（端点/协议/请求响应格式/定价）。配合 fetch_model_doc 使用。仅owner组织角色可用。",
|
||
parameters={"doc_text": "fetch_model_doc 返回的文档文本"},
|
||
category="platform",
|
||
),
|
||
ToolDefinition(
|
||
name="apply_llm_config",
|
||
description="按提取的规格写入模型治理配置：供应商/端点/适配模板/模型。幂等。推荐 use_last_extract=true 取本会话锚定规格（禁止把规格复制进 spec——复制即编造）；需改动加 overrides（仅 vendor_name/base_url/protocol/doc_url/doc_notes）。仅owner组织角色可用。",
|
||
parameters={
|
||
"use_last_extract": "true=用本会话 extract_llm_api_spec 锚定的规格（推荐，会话级隔离）",
|
||
"overrides": "覆盖项 JSON（白名单：vendor_name/base_url/protocol/doc_url/doc_notes，如用户要求改供应商名）",
|
||
"spec": "（仅无锚定时的兜底）extract 返回的 JSON 规格原文",
|
||
},
|
||
category="platform",
|
||
# 交互原则（2026-09-05 用户定）：配置链工具不设确认门——缺信息才问，
|
||
# 否则一口气做完给测试结果，用户发现问题再按说明修改。
|
||
),
|
||
ToolDefinition(
|
||
name="apply_model_pricing",
|
||
description="定价自动导入：建定价方案（pricing_program+时序YAML）并挂模型 ppid。幂等（内容无变化跳过；同日原地更新；历史行才拉链——时序表始终只一条有效行）。一个定价方案只服务一个模型；定价相同的模型共享同一 ppid。推荐 use_last_extract=true 取本会话锚定规格（含 pricing/doc_url），禁止复制转述。仅owner组织角色可用。",
|
||
parameters={
|
||
"use_last_extract": "true=用本会话锚定规格的 pricing+doc_url（推荐）",
|
||
"spec": "（仅无锚定时的兜底）extract 返回的 JSON 规格原文",
|
||
},
|
||
category="platform",
|
||
),
|
||
ToolDefinition(
|
||
name="test_model_call",
|
||
description="模型真实调用测试：走完整治理链（门禁→上游→结算）。返回 ok/content/usage/media（生成物本地URL）/task_ref。前置：供应商账号已配 api_key。异步模型（视频等）等待至任务完成（默认600秒）。仅owner组织角色可用。",
|
||
parameters={
|
||
"model_name": "模型名（llm_model.name 或 vendor_model_id）",
|
||
"prompt": "文本提示词（t2t 默认「回复两个字：正常」；生成类填生成描述）",
|
||
"params": "业务参数 JSON 字符串（生成类模型用，如 {\"image_file\": \"https://...\", \"resolution\": \"480P\", \"duration\": 5}）",
|
||
"timeout": "等待秒数（默认600，上限900）",
|
||
},
|
||
category="platform",
|
||
),
|
||
ToolDefinition(
|
||
name="check_model_accounting",
|
||
description="记账正确性检查（三态 created/accounted/failed）：查最近流水的记账状态/usages因子/charge金额，accounted 时用定价引擎独立复算金额对比。failed 给原因与处置建议。仅owner组织角色可用。",
|
||
parameters={
|
||
"task_ref": "test_model_call 返回的 task_ref（优先）",
|
||
"model_name": "或按模型名查最近流水",
|
||
},
|
||
category="platform",
|
||
),
|
||
ToolDefinition(
|
||
name="platform_modules",
|
||
description="列出平台已装载的业务模块清单。用户问「平台有哪些模块」时调用。仅owner组织角色可用。",
|
||
parameters={},
|
||
category="platform",
|
||
),
|
||
]
|
||
|
||
PLATFORM_PROMPT = """
|
||
你是产线平台的内部运维 agent，服务对象是 owner 组织的角色。
|
||
|
||
## 模型自动配置全链工作流（用户给文档 URL 要求配置模型时，一口气独立完成）
|
||
1. **先 `load_skill` 加载 `model-auto-config`**（含媒体转换铁律/能力分类先行/定价新模式），严格执行
|
||
2. `fetch_model_doc` 抓取官方文档（返回文本末尾带 [出处URL] 行）
|
||
3. `extract_llm_api_spec` 通读文档提取规格——**成功后规格自动锚定到本会话缓存**
|
||
（会话级隔离），返回里有 __anchored__ 提示
|
||
4. **不要停下来找用户确认**——直接进入下一步落库。只有文档里确实缺失、
|
||
且无法从上下文合理推断的信息（如 api_key、能力分类需新增）才向用户提问。
|
||
5. `apply_llm_config` 传 **{"use_last_extract": true}**（用户要求改名等只加
|
||
overrides 白名单覆盖项，如 {"overrides": {"vendor_name": "阿里云百炼"}}）。
|
||
**绝对禁止把规格 JSON 复制/转述进 spec 参数——转述必编造结构（历史教训：
|
||
三轮全编造 vendor{}/api_profile{} 等不存在的字段）。**
|
||
6. `apply_model_pricing` 同样传 **{"use_last_extract": true}**（定价+出处已锚定）
|
||
7. `test_model_call` 真实调用测试（前置：账号已配 api_key，没配则如实告知用户去补，
|
||
不要伪造结果）——生成类模型传 params（image_file/resolution/duration 等）
|
||
8. `check_model_accounting` 检查记账（传上一步的 task_ref）：
|
||
- accounted 且 amount_match=true → 全链通过，汇报金额
|
||
- created → 等 60~120 秒再查一次（出账循环 60 秒一轮）
|
||
- failed → 按 reason/hint 处置（未映射产品→告知用户在产品管理导入产线模型）
|
||
9. 汇报：配置/定价/测试/记账四段结果 + 出处 URL + 遗留事项（如模板需人工核对项）。
|
||
用户发现问题后按其说明修改，再重新走对应环节（全链幂等，重跑安全）。
|
||
|
||
## 交互原则（用户定，2026-09-05）
|
||
- **只有缺失信息需要用户补充时才交互**（api_key、业务参数、需新增能力分类等）
|
||
- 否则从抓取到记账检查**一口气做完**，最后给出完整测试结果
|
||
- **不要每步找用户确认**——落库/定价/测试都直接执行（全部幂等，改起来安全）
|
||
- 用户发现问题 → 按用户说明修改重跑，这才是正确的纠错循环
|
||
|
||
## 工具报错自救规则
|
||
- 报「spec 结构不符/缺少 xxx」→ **不要改写结构重试**，改用 use_last_extract=true；
|
||
本会话没锚定就重新走 fetch→extract（锚定是会话级的，跨会话不可见）
|
||
- 报「能力分类未登记」→ 如实转告用户需先加能力分类（这是人工评审门，不要绕过）
|
||
- 报「media_audit/runtime_note/需人工核对」警告 → 原样转告用户，不假装完成
|
||
|
||
## 硬规则
|
||
- **定价只允许来自文档原文**（doc_quote 强制），禁止编造/换算/推导
|
||
- **出处必留**：模型注册表描述字段与定价方案描述字段都必须保存文档 URL
|
||
- 能力分类不存在 → 先加分类（字典/种子/提示词/端点注释四处同步）再配模型，禁止塞近似分类
|
||
- 生成类模型媒体铁律：上行 b64media2url(request, xxx_file)、下行 downloadfile2url(request, url)；
|
||
apply 返回的 media_audit 警告必须如实转告
|
||
- 非标协议骨架模板需人工核对的项（__note__）必须如实转告用户，不假装完成
|
||
- 文档抓取失败/内容不足时如实说明，不要凭记忆编造 API 格式
|
||
- 测试没有 api_key 就停在配置阶段如实汇报，禁止跳过测试谎称完成
|
||
- 所有工具仅 owner 组织角色可用，权限报错时如实转告用户
|
||
|
||
## 平台知识
|
||
平台模块清单用 `platform_modules` 查询；技能库中「平台」相关技能（skills_library/all/ 与
|
||
pipelines/platform_general/）包含各模块的开发规范与踩坑记录，需要时用 load_skill 加载。
|
||
"""
|
||
|
||
PLATFORM_HANDLERS = {
|
||
"platform_llm_status": _h_platform_llm_status,
|
||
"fetch_model_doc": _h_fetch_model_doc,
|
||
"extract_llm_api_spec": _h_extract_llm_api_spec,
|
||
"apply_llm_config": _h_apply_llm_config,
|
||
"apply_model_pricing": _h_apply_model_pricing,
|
||
"test_model_call": _h_test_model_call,
|
||
"check_model_accounting": _h_check_model_accounting,
|
||
"platform_modules": _h_platform_modules,
|
||
}
|
||
|
||
|
||
def register_platform_ability():
|
||
"""注册平台内部 agent 能力包（幂等）。"""
|
||
ability = PipelineAbility(
|
||
pipeline_id=PLATFORM_PIPELINE_ID,
|
||
name="平台内部 agent",
|
||
tools=PLATFORM_TOOLS,
|
||
system_prompt=PLATFORM_PROMPT,
|
||
handlers=PLATFORM_HANDLERS,
|
||
roles=[],
|
||
menus=[
|
||
{"label": "📊 模型治理", "icon": "", "url": "/pipeline-llm",
|
||
"type": "tab"},
|
||
],
|
||
)
|
||
register_ability(ability)
|
||
return ability
|
||
|
||
|
||
register_platform_ability()
|