init: api-gateway v2.4.0(Agnes/b.ai/火山方舟 三家供应商网关)

This commit is contained in:
api-gateway committed 2026-09-15 17:05:46 +08:00
commit f136c9ee81
91 files changed
+24653

No files matched your search

+208
View File
@@ -0,0 +1,208 @@
"""请求 / 响应数据模型。"""
from __future__ import annotations
from typing import Any
from pydantic import BaseModel, ConfigDict, Field
# 不允许被客户端 `extra` / 未知字段覆盖的"核心键"。
# 这些键承载网关已解析的路由结果与协议语义,一旦让 extra 覆盖就会出现:
# - `extra={"model": "..."}` → 劫持路由,用网关的供应商 Key 调任意(更贵/未登记)模型;
# - `extra={"stream": ...}` → 与网关自身的解析结果矛盾,非流式客户端收到 SSE(502)或反之;
# - `extra={"messages": [...]}` → 绕过入参校验。
# 第六轮安全审计实测确认了上述越权路径,故在**组装时**统一剔除。
PROTECTED_PAYLOAD_KEYS: frozenset[str] = frozenset({
"model", "messages", "stream", "tools", "tool_choice",
"temperature", "top_p", "max_tokens", "stop", "n",
"presence_penalty", "frequency_penalty", "seed", "user",
"response_format", "logprobs", "top_logprobs",
"parallel_tool_calls", "stream_options",
"prompt", "input", "image", "images", "seconds", "mode",
"aspect_ratio", "size", "extra", "extra_body",
})
# 视频路径的保护集**刻意更窄**:first_frame / last_frame / images / audios / seed
# 是模式专用参数,必须允许透传;只挡住会改变"调哪个模型/多长/什么模式"的键。
PROTECTED_VIDEO_KEYS: frozenset[str] = frozenset({
"model", "prompt", "seconds", "mode", "size", "aspect_ratio", "n", "extra", "extra_body",
})
# 图像路径同理:只挡"会改变调哪个模型/什么尺寸/几张/输入图"的键。
# 注意**不能**沿用 PROTECTED_PAYLOAD_KEYS —— 那会把 response_format / user 一起删掉。
PROTECTED_IMAGE_KEYS: frozenset[str] = frozenset({
"model", "prompt", "size", "n", "image", "images", "extra", "extra_body",
})
# 适配器 `_build_payload` **自己写进去**的键(chat)。
# 只有这些才需要防覆盖 —— 其余键(n / seed / user / response_format /
# presence_penalty / logprobs …)恰恰是 ChatRequest._PASSTHROUGH 特意塞进 extra
# 要透传给上游的,第七轮回归审计实测:若用宽集合过滤,这些参数会被**静默丢弃**
# (等于废掉 _PASSTHROUGH 机制),属本轮引入的兼容性倒退。
CHAT_CORE_KEYS: frozenset[str] = frozenset({
"model", "messages", "stream", "tools", "tool_choice",
"temperature", "top_p", "max_tokens", "stop",
})
def strip_protected(extra: dict[str, Any] | None,
keys: frozenset[str] = PROTECTED_PAYLOAD_KEYS) -> dict[str, Any]:
"""剔除受保护的核心键,其余原样透传(适配器侧的最后一道防线)。"""
return {k: v for k, v in (extra or {}).items() if k not in keys}
class Message(BaseModel):
# 兼容性修复(第六轮): 此前只有 role/content,导致 function calling 的多轮对话
# **直接坏掉** —— assistant 消息上的 tool_calls、tool 消息上的 tool_call_id 被静默丢弃,
# 上游收到的是 `{"role":"assistant","content":null}` 这种缺胳膊少腿的历史。
# 现在显式声明这几个字段,并用 extra="allow" 兜住其余规范字段(refusal / audio 等)。
model_config = ConfigDict(extra="allow")
role: str
content: Any = None # string / array(多模态)/ null(带 tool_calls 的 assistant 消息)
name: str | None = None
tool_calls: list[dict[str, Any]] | None = None
tool_call_id: str | None = None
def to_upstream(self) -> dict[str, Any]:
"""转发给上游用:丢掉 None 字段,其余(含未知字段)原样带上。"""
return self.model_dump(exclude_none=True)
class ToolFunction(BaseModel):
name: str
description: str | None = None
parameters: dict[str, Any] | None = None
class ToolDef(BaseModel):
type: str = "function"
function: ToolFunction
class ChatRequest(BaseModel):
# 兼容性修复:OpenAI 官方 SDK 会带不少可选参数(n / presence_penalty / response_format …),
# 之前未声明的字段被 pydantic 静默丢弃,导致"网关号称 OpenAI 全兼容但参数悄悄丢失"。
# 现在:常用标准参数显式声明,其余未知字段用 extra="allow" 捕获后一并透传上游。
model_config = ConfigDict(extra="allow")
model: str
messages: list[Message] = Field(min_length=1) # 空 messages 上游会报错,这里先给 400
temperature: float | None = None
top_p: float | None = None
max_tokens: int | None = None
# strict: `"stream": "yes"` 这类拼错不应被宽松强转成 True(会拿到 SSE 而非 JSON)
stream: bool = Field(default=False, strict=True)
stop: str | list[str] | None = None
tools: list[ToolDef] | None = None
tool_choice: str | dict | None = None
# ── OpenAI 标准可选参数(透传上游) ──
n: int | None = None
presence_penalty: float | None = None
frequency_penalty: float | None = None
seed: int | None = None
user: str | None = None
response_format: dict[str, Any] | None = None
logprobs: bool | None = None
top_logprobs: int | None = None
parallel_tool_calls: bool | None = None
stream_options: dict[str, Any] | None = None
# 透传供应商特有参数
extra: dict[str, Any] = Field(default_factory=dict)
# 需要透传的标准参数白名单
_PASSTHROUGH = (
"n", "presence_penalty", "frequency_penalty", "seed", "user",
"response_format", "logprobs", "top_logprobs",
"parallel_tool_calls", "stream_options",
)
def upstream_extra(self) -> dict[str, Any]:
"""组装要透传给上游的额外参数。
优先级(后者覆盖前者):显式声明的标准参数 < 未知透传字段(model_extra) < `extra`。
但**核心键一律剔除**(见 PROTECTED_PAYLOAD_KEYS):`extra` 是"供应商特有参数"的
逃生舱,不是用来覆盖 model/messages/stream 的 —— 否则等同把路由与协议语义交给客户端。
"""
out: dict[str, Any] = {}
for k in self._PASSTHROUGH:
v = getattr(self, k, None)
if v is not None:
out[k] = v
merged: dict[str, Any] = {}
merged.update(self.model_extra or {})
merged.update(self.extra or {})
for k, v in merged.items():
if k in PROTECTED_PAYLOAD_KEYS:
continue
out[k] = v
return out
def unknown_fields(self) -> list[str]:
"""未被显式声明、靠 extra="allow" 兜住的字段名(用于日志提示,避免悄悄透传)。"""
return sorted(k for k in (self.model_extra or {}) if k not in PROTECTED_PAYLOAD_KEYS)
# ── 图像生成 ─────────────────────────────────────────
class ImageRequest(BaseModel):
# 第六轮: extra="allow" —— quality/style/response_format/user 等标准参数此前被
# pydantic 静默丢弃(客户端以为生效了)。现在统一收进 model_extra 再透传。
model_config = ConfigDict(extra="allow")
model: str
prompt: str
size: str | None = None
ratio: str | None = None
image: list[str] | None = None # 图生图 / 多图输入 URL 或 DataURI
n: int = 1
return_base64: bool = False
extra: dict[str, Any] = Field(default_factory=dict)
# ── 视频生成(异步任务) ─────────────────────────────
class VideoCreateRequest(BaseModel):
model: str
prompt: str
seconds: str = "5"
mode: str = "text" # text | keyframe | reference
size: str = "720P"
aspect_ratio: str = "16:9"
n: int = 1 # 第十三轮: 上游只支持 1,>1 在网关入站就回 400
seed: int | None = None
# 模式专用
first_frame: str | None = None
last_frame: str | None = None
images: list[str] | None = None
audios: list[str] | None = None
# 第十三轮: 声明出来只为**明确拒绝**(Agnes Video 2.5 Flash 不支持参考视频)。
# 此前该字段未声明 → pydantic 静默丢弃,用户以为"传了参考视频",实际被悄悄忽略。
videos: list[Any] | None = None
extra: dict[str, Any] = Field(default_factory=dict)
# ── 视频等待模式 / Webhook ───────────────────────────
class VideoCreateRequestExtended(VideoCreateRequest):
"""支持 wait=true 和 callback_url 的视频创建请求。"""
wait: bool = False
wait_timeout: int = 120 # 等待完成的最大秒数
callback_url: str | None = None # 任务完成/失败时 POST 通知
# ── Batch 调用 ────────────────────────────────────────
class BatchItem(BaseModel):
"""Batch 中单个调用项。"""
messages: list[Message]
temperature: float | None = None
top_p: float | None = None
max_tokens: int | None = None
label: str = "" # 自定义标签(结果对应)
class BatchRequest(BaseModel):
"""一次 POST 并发调多个 prompt(同模型)。"""
model: str
items: list[BatchItem]
max_concurrency: int = 8 # 最大并发数
extra: dict[str, Any] = Field(default_factory=dict)