Files
Sakurasan f9e9a1572f feat: 系统设置新增原始请求/响应记录开关(仅管理员)
- 系统配置键 log_raw_requests:开启后,仅管理员账号的每次请求
  在用量明细中保存客户端原始请求体与上游原始响应体
  (流式含全部 SSE 事件),用于排障
- UsageLog 新增 raw_request / raw_response 字段(type:text)
- AuthLLM 附带 user_role 供网关判断管理员
- gateway:10s TTL 缓存开关;streamResponse/bufferResponse
  支持累积上游原始响应;recordUsage 填充原始字段
- 前端 SystemConfig 新增开关(会显著增加存储的提示)
- 新增 doc/flow.md 网关调用流程示意图
2026-09-01 02:32:31 +08:00

277 lines
11 KiB
Markdown
Raw Permalink Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# 网关调用流程示意图
> 对应实现:`backend/router/setRouter.go`(路由注册)、`backend/middleware/auth_llm.go`(鉴权)、
> `backend/internal/proxy/{gateway.go,handlers.go,convert/*}`(网关与三协议互转)、
> `backend/internal/channel/{channel.go,health.go}`(渠道路由与健康)、
> `backend/internal/usage/recorder.go`(用量异步落库)。
## 0. 总览
```
┌────────────────────────────────────────────────┐
│ Gin Router (/v1) │
│ ┌──────────────────────────────────────────┐ │
客户端 ───────────▶│ │ middleware.AuthLLM (密钥鉴权, 401拦截) │ │
Bearer sk-ot-… │ └──────────────────────────────────────────┘ │
│ ┌───────┬────────┬─────────┬─────────┐ │
│ │ chat │messages│responses│ models │ │
│ │Handle │Handle │ Handle │ Handle │ │
│ │Chat │Messages│Responses│ Models │ │
│ └───┬───┴───┬────┴────┬────┴─────────┘ │
│ └───────┴────┬────┘ │
│ ParseRequest │
│ (model/stream) │
│ │ │
│ Dispatch ◀── Candidates/Pick │
│ (转换+故障转移+用量记录) │
└───────────────────┼──────────────────────────────┘
│
▼
上游 /v1/* (chat|messages|responses)
```
## 1. 请求入口与鉴权
```mermaid
sequenceDiagram
autonumber
participant C as 客户端
participant R as Gin /v1 路由
participant A as AuthLLM
participant DB as SQLite(APIKey)
participant H as HandleXxx
C->>R: POST /v1/chat/completions 等
Note over R: /v1 组挂 middleware.AuthLLM
R->>A: 进入中间件
A->>A: 提取 Bearer token(兼容无前缀直传)
alt 未携带 Authorization
A-->>C: 401 「未提供认证信息」
else token 长度 < 12 或 prefix 不匹配
A-->>C: 401 「无效的API密钥」
else prefix 命中
A->>DB: SELECT * WHERE key_prefix=? AND status=active
A->>A: sha256(token) == KeyHash ?
alt 哈希不一致
A-->>C: 401 「无效的API密钥」
else 校验通过
A->>H: c.Set(api_key, user_id) → 放行
end
end
```
> 关键点:`key_prefix` 取 `sk-ot-` 后前 12 位(`api.go` 创建密钥时 `keyValue[:12]`),
> `auth_llm.go` 用同一常量 `keyPrefixLen=12` 切片,避免越界 panic。
## 2. 三协议主调用流程
```mermaid
sequenceDiagram
autonumber
participant C as 客户端
participant H as HandleChat/Messages/Responses
participant P as ParseRequest
participant G as Dispatch
participant S as ChannelService
participant U as usage.Recorder
participant UP as 上游(OpenRouter等)
C->>H: 请求体 (model, stream, messages/input…)
H->>P: ParseRequest(protocol)
P->>P: 读 body → 解析 model / stream
P-->>H: Request{Model, Stream, Protocol, Body, …}
H->>G: Dispatch(req)
G->>S: Candidates(req.Model)
S-->>G: []Candidate{Channel, Binding?}
G->>S: FilterHealthy(cands)
G->>S: Pick(cands) → 加权随机选定一个候选
Note over G,S: 绑定优先(携带 upstream_model 映射);<br/>无绑定回退到权重最低的健康备用渠道
loop 故障转移(候选耗尽前)
G->>G: conversionTarget(ch, proto) → 渠道首选协议
alt 客户端协议 ≠ 渠道协议
G->>G: ConvertRequest(body, from, to) 转换请求体
Note over G: 走 convert 包(chat/messages/responses 互转)
end
alt 有 Binding.UpstreamModel
G->>G: rewriteModel(body, upstreamModel) 别名映射
end
G->>UP: POST {base}/v1/{path} (按渠道协议拼 URL/头)
alt 连接失败 或 429/5xx
S->>S: RecordFailure(ch) → 连续2次 degraded 熔断
G->>U: Record(error 事件, error_code)
Note over G: continue → 换下一个候选渠道
else 4xx
G-->>C: 透传上游错误体 (不重试)
else 2xx
S->>S: RecordSuccess(ch)
alt stream=true
G->>G: streamResponse → 逐块转发 + 累计 usage
else
G->>G: bufferResponse → 整体转发 + 提取 usage
end
G->>G: recordUsage (按模型定价计算 cost)
G->>U: Record(成功事件, tokens, cost)
G-->>C: 响应
end
end
Note over G,C: 全部候选失败 → 502/503
```
> **故障转移规则**(对齐参考实现 `doProxy`):
> - 连接错误、429、5xx → 可重试,换下一个候选;
> - 4xx(如 400 参数错误)→ 透传上游错误体,不重试;
> - 无可用渠道(全部不健康/无绑定且无备用)→ 502/503 + `error_code=no_channel`。
## 3. 路由选择细节
```mermaid
flowchart TD
A[客户端 model 名] --> B{存在启用模型行?}
B -- 是 --> C{有绑定且渠道健康?}
C -- 是 --> D[候选 = 绑定该模型的渠道<br/>排序 priority ASC, weight DESC, id ASC]
C -- 否 --> E
B -- 否 --> E[候选 = 全部启用渠道<br/>取权重最低的健康备用渠道]
D --> F[FilterHealthy 内存熔断过滤]
E --> F
F --> G[Pick 加权随机选中一个]
G --> H[Dispatch 开始尝试]
H --> I{尝试成功?}
I -- 失败可重试 --> J[RecordFailure + 换下一个]
J --> H
I -- 成功 --> K[RecordSuccess + 响应 + 记账]
J -. 全部耗尽 .-> L[502/503]
```
## 4. 跨协议转换(client ↔ 渠道原生协议)
```mermaid
flowchart LR
subgraph 客户端协议
CHAT[/"chat<br/>chat/completions"/]
MSG[/"messages<br/>(Anthropic)"/]
RESP[/"responses<br/>(OpenAI)"/]
end
subgraph 中间模型
MID["Chat 形状<br/>(标准中间模型)"]
end
subgraph 渠道协议
UCHAT[/"chat"/]
UMSG[/"messages"/]
URESP[/"responses"/]
end
CHAT -->|直通| UCHAT
MSG -->|messagesToChat| MID -->|chatToMessages| UMSG
MSG -->|messagesToChat| MID -->|chatToResponses| URESP
RESP -->|responsesToChat| MID -->|chatToResponses| URESP
RESP -->|responsesToChat| MID -->|chatToMessages| UMSG
```
> 转换入口:`convert.ConvertRequest`(请求体)、`convert.ConvertResponse`(非流式响应)、
> `convert.NewStreamTransformer`(流式 SSE 逐行转换)。跨两跳时经 Chat 中转(如
> `responses→messages` = `responsesToChatReq` + `chatToMessagesReq`)。
## 5. 流式 / 非流式响应与用量提取
```mermaid
sequenceDiagram
autonumber
participant G as Dispatch
participant S as streamResponse
participant B as bufferResponse
participant ACC as StreamUsageAccum
participant W as 客户端 Writer
participant UP as 上游
alt stream=true
G->>S: streamResponse(resp, clientProto, upstreamProto)
S->>UP: 按 \n\n 读块 (bufio)
loop 每个 SSE 块
S->>ACC: sseDataPayloads(chunk) → Feed(data, upstreamProto)
Note over ACC: 逐协议累计 usage 字段
alt 跨协议
S->>S: NewStreamTransformer(upstream→client).line(chunk)
end
S->>W: 写块 + Flush
alt 遇到流结束标记
Note over S: chat: data:[DONE]<br/>messages: message_stop<br/>responses: response.completed
S-->>G: 返回累计 TokenUsage
end
end
else stream=false
G->>B: bufferResponse(resp, clientProto, upstreamProto)
B->>B: io.ReadAll
B->>B: ExtractUsageJSON(body, upstreamProto)
alt 跨协议
B->>B: ConvertResponse(body, upstream→client)
else 直通
B->>B: CleanJSON(body) 去空白/SSE注释前缀
end
B-->>G: 返回 TokenUsage
end
G->>G: recordUsage(req, cand, ch, ev, tok)
Note over G: 定价 cost = (非缓存输入×输入价 + 缓存读×缓存价<br/>+ 缓存写×输出价 + 输出×输出价) / 1e6
G->>G: usageRec.Record(Event)
```
## 6. 用量异步落库
```mermaid
sequenceDiagram
autonumber
participant G as Gateway
participant R as usage.Recorder
participant U as UsageDAO
participant D as DailyUsageDAO
participant DB as SQLite
G->>R: Record(Event) 每次请求(成功/错误/取消)
Note over R: 缓冲 channel (10000), 每 5s 或满 100 条 flush
R->>U: BatchCreate(UsageLog[])
R->>D: UpsertDailyUsage(UsageDaily) 按(user_id,model_id,date)增量累加
U->>DB: INSERT usage_logs
D->>DB: ON CONFLICT 累加 requests/input/output/cache/cost
```
> `usage_dailies` 用 `gorm.Expr("requests + ?")` 增量累加而非覆盖,保证多次 flush 不互相清零。
## 7. 渠道健康与熔断
```mermaid
flowchart TD
A[请求失败] --> B[RecordFailure: consecutive++]
B --> C{consecutive >= 2?}
C -- 是 --> D[status=degraded + 5min cooldown]
C -- 否 --> E[仅计数]
D --> F{后续请求 Candidates}
F --> G{FilterHealthy 该渠道}
G -- degraded/cooldown 未过期 --> H[排除, 走其他渠道/备用]
G -- healthy --> I[参与选择]
D -. 冷却过期 .-> J[复位 healthy]
J --> I
K[健康检查周期探测成功] --> L[RecordSuccess: 复位 healthy]
```
> 说明:`health.go` 的 `StartPeriodicCheck`(默认 5min)会探测各渠道 `/models`,
> 成功调 `RecordSuccess` 复位;失败调 `RecordFailure` 进入熔断计数。
## 8. 关键代码锚点
| 环节 | 位置 |
|---|---|
| /v1 路由注册 + AuthLLM | `router/setRouter.go:146` |
| 密钥鉴权 | `middleware/auth_llm.go` |
| 请求解析 | `proxy/gateway.go:104 ParseRequest` |
| 主调度 + 故障转移 | `proxy/gateway.go:157 Dispatch` |
| 流式转发 + 结束检测 | `proxy/gateway.go:388 streamResponse` |
| 非流式转发 | `proxy/gateway.go:496 bufferResponse` |
| 用量记账 | `proxy/gateway.go:302 recordUsage` |
| 候选构建 | `channel/channel.go:51 Candidates` |
| 内存健康过滤 | `channel/channel.go:252 FilterHealthy` |
| 加权选择 | `channel/channel.go:222 Pick` |
| 失败熔断 | `channel/channel.go:139 RecordFailure` |
| 三协议互转 | `proxy/convert/{convert.go,json_chat.go,json_responses.go,stream_transform.go}` |
| 用量异步落库 | `usage/recorder.go:114 flush` |