- 神识 Go:store 双实现(PG 自动建表+60s 就绪重试 / 内存回退)承载 usage_logs·byok_keys·quota_balances;BYOK 密钥库(AES-GCM 加密、API 脱敏、撤回即失效、主密钥未配置 fail-closed);配额三段式(预扣 max_tokens 钳位 1M→结算实际 token→全败幂等退款,HJ-2004);计量归属(IR user_id→usage_logs,/me/usage 用户维度聚合+配额总览,/admin/usage/sample 对账抽样);Anthropic/Gemini 出站适配器(SSE 解析、system 提升、用量映射);stub 端点补名修复计量空 endpoint - IR 契约:user_id/byok_key_id 双端同步(Go types.go + Py types.py) - 灵台:runs.owner_id(迁移 005,server_default 兼容存量)经 kernel 全链透传为 IR user_id;shenshi_client 扩展用户面方法(X-Hunjing-User-Id 代理,Echo 桩带内存实现);/me/byok-keys 登记·列表·撤回与 /me/usage 端点(4 用例) - 前端神识阁(/models 重建):偏好(生成参数表单)/自带密钥(登记加密保存、脱敏列表、撤回二次确认)/用量(聚合表+配额进度条)三区;me API 库+mock+5 用例;侧栏恢复入口 - 配置与部署:.env.example 增 SHENSHI_DATABASE_URL/SHENSHI_BYOK_MASTER_KEY/SHENSHI_USER_BUDGET_TOKENS;compose 神识容器接自有库(init-shenshi.sql 预建 hunjing_shenshi,与灵台不共库共表) - 证据:Go 34 用例(BYOK 独占/配额拒绝/结算退款/归属聚合)、后端 532、前端 210、BFF 4 全绿;跨进程 E2E 四场景(灵台 HttpShenshiClient→真 Go 神识:登记脱敏/BYOK 独占不可达整体失败/用量归属聚合/撤回即失效)通过;rewrite-plan M4 标记完成、能力矩阵神识行升已验证
192 lines
7.8 KiB
Python
192 lines
7.8 KiB
Python
"""神识客户端:灵台唯一的模型出口边界。
|
||
|
||
两种实现可互换(Protocol):
|
||
- ``HttpShenshiClient``:经 HTTP+SSE 调用 Go 神识服务(SHENSHI_BASE_URL 配置时启用);
|
||
- ``EchoShenshi``:进程内回声桩,逐字回流输入,供本地开发与测试在无神识时跑通端到端。
|
||
|
||
领域模块禁止绕过本边界 import Provider SDK(docs/architecture/service-boundaries.md §4)。
|
||
"""
|
||
|
||
from __future__ import annotations
|
||
|
||
import json
|
||
from collections.abc import AsyncIterator
|
||
from typing import Any, Protocol
|
||
|
||
import httpx
|
||
|
||
from app.shenshi_client.types import ShenshiChatRequest, ShenshiDelta, ShenshiError
|
||
|
||
|
||
class ShenshiClient(Protocol):
|
||
"""神识数据面契约:流式对话补全 + 用户面(BYOK/用量,神识阁数据源)。"""
|
||
|
||
def stream_chat(self, request: ShenshiChatRequest) -> AsyncIterator[ShenshiDelta]: ...
|
||
|
||
async def get_my_usage(self, user_id: str) -> dict[str, Any]: ...
|
||
|
||
async def list_byok_keys(self, user_id: str) -> list[dict[str, Any]]: ...
|
||
|
||
async def register_byok_key(
|
||
self, user_id: str, *, name: str, type: str, base_url: str, api_key: str, model: str
|
||
) -> dict[str, Any]: ...
|
||
|
||
async def revoke_byok_key(self, user_id: str, key_id: str) -> None: ...
|
||
|
||
|
||
class EchoShenshi:
|
||
"""进程内回声桩:把最后一条用户消息按固定粒度切分回流。
|
||
|
||
数据面只用于骨架验证(登录 → 占位 Run → SSE 端到端);用户面为最小内存实现,
|
||
供神识阁在无神识部署时演示与测试桩。
|
||
"""
|
||
|
||
chunk_size = 8
|
||
|
||
def __init__(self) -> None:
|
||
self._byok: dict[str, list[dict[str, Any]]] = {}
|
||
self._usage: dict[str, list[dict[str, Any]]] = {}
|
||
|
||
async def stream_chat(self, request: ShenshiChatRequest) -> AsyncIterator[ShenshiDelta]:
|
||
prompt = next((m.content for m in reversed(request.messages) if m.role == "user"), "")
|
||
text = f"【回声桩】{prompt}"
|
||
for i in range(0, len(text), self.chunk_size):
|
||
yield ShenshiDelta(delta=text[i : i + self.chunk_size])
|
||
yield ShenshiDelta(
|
||
finish_reason="stop",
|
||
usage_prompt_tokens=max(len(prompt) // 4, 1),
|
||
usage_completion_tokens=max(len(text) // 4, 1),
|
||
)
|
||
if request.user_id:
|
||
self._usage.setdefault(request.user_id, []).append(
|
||
{
|
||
"user_id": request.user_id,
|
||
"model": request.model or "",
|
||
"endpoint": "echo",
|
||
"requests": 1,
|
||
"ok": 1,
|
||
"prompt_tokens": max(len(prompt) // 4, 1),
|
||
"completion_tokens": max(len(text) // 4, 1),
|
||
}
|
||
)
|
||
|
||
async def get_my_usage(self, user_id: str) -> dict[str, Any]:
|
||
return {"rows": self._usage.get(user_id, [])}
|
||
|
||
async def list_byok_keys(self, user_id: str) -> list[dict[str, Any]]:
|
||
return self._byok.get(user_id, [])
|
||
|
||
async def register_byok_key(
|
||
self, user_id: str, *, name: str, type: str, base_url: str, api_key: str, model: str
|
||
) -> dict[str, Any]:
|
||
entry = {
|
||
"id": f"byok_echo_{len(self._byok.get(user_id, [])) + 1}",
|
||
"name": name,
|
||
"type": type,
|
||
"base_url": base_url,
|
||
"model": model,
|
||
"api_key_mask": api_key[:3] + "****",
|
||
}
|
||
self._byok.setdefault(user_id, []).append(entry)
|
||
return entry
|
||
|
||
async def revoke_byok_key(self, user_id: str, key_id: str) -> None:
|
||
self._byok[user_id] = [k for k in self._byok.get(user_id, []) if k["id"] != key_id]
|
||
|
||
|
||
class HttpShenshiClient:
|
||
"""HTTP+SSE 神识客户端。
|
||
|
||
认证:服务令牌经 ``Authorization: Bearer`` 透传(管理面 RBAC 不在此层)。
|
||
流解析:``data:`` 帧为 ShenshiDelta JSON,``[DONE]`` 结束;非 2xx 映射 ShenshiError。
|
||
"""
|
||
|
||
def __init__(self, base_url: str, service_token: str, timeout: float = 120.0) -> None:
|
||
self._base_url = base_url.rstrip("/")
|
||
self._headers = {"Authorization": f"Bearer {service_token}"} if service_token else {}
|
||
self._timeout = timeout
|
||
|
||
def _user_headers(self, user_id: str) -> dict[str, str]:
|
||
return {**self._headers, "X-Hunjing-User-Id": user_id}
|
||
|
||
async def _request_json(
|
||
self, method: str, path: str, *, user_id: str, json_body: dict[str, Any] | None = None
|
||
) -> Any:
|
||
async with httpx.AsyncClient(trust_env=False, timeout=self._timeout) as client:
|
||
response = await client.request(
|
||
method,
|
||
f"{self._base_url}{path}",
|
||
json=json_body,
|
||
headers=self._user_headers(user_id),
|
||
)
|
||
if response.status_code >= 400:
|
||
code = f"HJ-2{max(response.status_code - 400, 0):03d}"
|
||
try:
|
||
detail = response.json()
|
||
code = detail.get("error", {}).get("code", code)
|
||
message = detail.get("error", {}).get("message", response.text)
|
||
except ValueError:
|
||
message = response.text
|
||
raise ShenshiError(code, message)
|
||
if response.status_code == 204 or not response.content:
|
||
return None
|
||
return response.json()
|
||
|
||
async def get_my_usage(self, user_id: str) -> dict[str, Any]:
|
||
data = await self._request_json("GET", "/me/usage", user_id=user_id)
|
||
return data or {}
|
||
|
||
async def list_byok_keys(self, user_id: str) -> list[dict[str, Any]]:
|
||
data = await self._request_json("GET", "/me/byok-keys", user_id=user_id)
|
||
return (data or {}).get("keys", [])
|
||
|
||
async def register_byok_key(
|
||
self, user_id: str, *, name: str, type: str, base_url: str, api_key: str, model: str
|
||
) -> dict[str, Any]:
|
||
data = await self._request_json(
|
||
"POST",
|
||
"/me/byok-keys",
|
||
user_id=user_id,
|
||
json_body={"name": name, "type": type, "base_url": base_url, "api_key": api_key, "model": model},
|
||
)
|
||
return data or {}
|
||
|
||
async def revoke_byok_key(self, user_id: str, key_id: str) -> None:
|
||
await self._request_json("DELETE", f"/me/byok-keys/{key_id}", user_id=user_id)
|
||
|
||
async def stream_chat(self, request: ShenshiChatRequest) -> AsyncIterator[ShenshiDelta]:
|
||
payload = request.model_dump(mode="json")
|
||
assert payload.get("stream") is True # 数据面当前只有流式形态
|
||
async with httpx.AsyncClient(trust_env=False, timeout=self._timeout) as client:
|
||
async with client.stream(
|
||
"POST",
|
||
f"{self._base_url}/v1/chat/completions",
|
||
json=payload,
|
||
headers=self._headers,
|
||
) as response:
|
||
if response.status_code >= 400:
|
||
body = (await response.aread()).decode("utf-8", errors="replace")
|
||
code = f"HJ-2{max(response.status_code - 400, 0):03d}"
|
||
try:
|
||
detail = json.loads(body)
|
||
code = detail.get("error", {}).get("code", code)
|
||
body = detail.get("error", {}).get("message", body)
|
||
except json.JSONDecodeError:
|
||
pass
|
||
raise ShenshiError(code, body)
|
||
async for line in response.aiter_lines():
|
||
if not line.startswith("data:"):
|
||
continue
|
||
data = line[len("data:") :].strip()
|
||
if data == "[DONE]":
|
||
return
|
||
if data:
|
||
yield ShenshiDelta.model_validate_json(data)
|
||
|
||
|
||
def build_shenshi_client(base_url: str, service_token: str) -> ShenshiClient:
|
||
"""按配置装配客户端:未配置神识地址时降级为回声桩(M0 骨架形态)。"""
|
||
if base_url:
|
||
return HttpShenshiClient(base_url, service_token)
|
||
return EchoShenshi()
|