- 客户端解析上游 usage(含 DeepSeek 缓存命中/未命中),跨重试累加 - 新增计价表(三档单价 + 峰谷时段),费用在结算时冻结 - 新增用量上报与本地 outbox 重试;4xx 判定为确定性拒绝,不无限重试 - 面板补充 Token/费用/调用次数与计价快照 - 顺带修复面板内联脚本转义错误(模板字面量吃掉 \n 与 \',面板一直空白)
179 lines
5.6 KiB
TypeScript
179 lines
5.6 KiB
TypeScript
/**
|
||
* LLM 客户端 — DeepSeek OpenAI 兼容接口
|
||
*
|
||
* 设计要点:
|
||
* - Key 从环境变量读取,永不写入代码或日志
|
||
* - 未配置 Key 时 isAvailable() 为 false,调用方返回 503,由 Java 端降级
|
||
* - 超时保护,避免长时间挂起
|
||
*/
|
||
|
||
import type { ReasoningEffort } from "./thinking.js";
|
||
import { normalizeUsage, type TokenUsage, type UsageSource } from "./usage.js";
|
||
|
||
export interface ChatMessage {
|
||
role: "system" | "user" | "assistant";
|
||
content: string;
|
||
}
|
||
|
||
/**
|
||
* 一次成功调用的返回值。
|
||
*
|
||
* 除了正文,还带回上游上报的 token 用量:调用方必须把它记下来,
|
||
* 否则重试与失败尝试产生的消耗就再也拿不到了(上游不会补发)。
|
||
*/
|
||
export interface ChatResult {
|
||
content: string;
|
||
usage: TokenUsage;
|
||
usageSource: UsageSource;
|
||
/** 上游回显的模型名,可能与本项目配置的模型名不同 */
|
||
model: string;
|
||
finishReason: string | null;
|
||
}
|
||
|
||
/** 单次调用的采样参数:温度与输出上限按任务类型定,思考强度按 prompt 长度定 */
|
||
export interface ChatOptions {
|
||
temperature?: number;
|
||
maxTokens?: number;
|
||
/** 思考强度;不传时不带 `reasoning_effort`,由供应商按自己的默认值处理 */
|
||
reasoningEffort?: ReasoningEffort;
|
||
}
|
||
|
||
export interface LlmConfig {
|
||
apiUrl: string;
|
||
apiKey: string;
|
||
model: string;
|
||
timeoutMs: number;
|
||
}
|
||
|
||
export function loadConfig(): LlmConfig {
|
||
return {
|
||
apiUrl: process.env.LLM_API_URL ?? "https://api.deepseek.com/v1/chat/completions",
|
||
apiKey: process.env.LLM_API_KEY ?? "",
|
||
model: process.env.LLM_MODEL ?? "deepseek-flash",
|
||
timeoutMs: Number(process.env.LLM_TIMEOUT_MS ?? 120000),
|
||
};
|
||
}
|
||
|
||
export function isAvailable(config: LlmConfig): boolean {
|
||
return config.apiKey.length > 0;
|
||
}
|
||
|
||
export class LlmError extends Error {
|
||
constructor(
|
||
message: string,
|
||
public readonly statusCode: number,
|
||
/**
|
||
* 服务端给出的建议重试等待(毫秒)。
|
||
*
|
||
* 来自 429/503 的 `Retry-After` 头;有值时退避重试应优先采用它,
|
||
* 而不是用自己的指数退避——服务商限流时自己的退避往往偏短。
|
||
*/
|
||
public readonly retryAfterMs?: number,
|
||
/**
|
||
* 这次失败的调用是否已经产生了 token 消耗。
|
||
*
|
||
* 典型场景是上游返回 200 但内容为空:请求已经计费,只是没法用。
|
||
* 有值时调用方必须把它计入用量,否则这部分消耗会凭空消失。
|
||
*/
|
||
public readonly usage?: TokenUsage,
|
||
) {
|
||
super(message);
|
||
this.name = "LlmError";
|
||
}
|
||
}
|
||
|
||
/**
|
||
* 解析 `Retry-After` 头。
|
||
*
|
||
* 规范允许秒数或 HTTP 日期两种形式,两种都支持;解析不出来时返回 undefined。
|
||
*/
|
||
export function parseRetryAfter(value: string | null, now: number = Date.now()): number | undefined {
|
||
if (!value) return undefined;
|
||
const trimmed = value.trim();
|
||
if (trimmed.length === 0) return undefined;
|
||
|
||
// 纯数字:秒
|
||
if (/^\d+$/.test(trimmed)) {
|
||
return Number(trimmed) * 1000;
|
||
}
|
||
|
||
// HTTP 日期
|
||
const at = Date.parse(trimmed);
|
||
if (Number.isNaN(at)) return undefined;
|
||
return Math.max(0, at - now);
|
||
}
|
||
|
||
/**
|
||
* 非流式调用,返回正文与上游上报的 token 用量。
|
||
*/
|
||
export async function chat(
|
||
config: LlmConfig,
|
||
messages: ChatMessage[],
|
||
options?: ChatOptions,
|
||
): Promise<ChatResult> {
|
||
if (!isAvailable(config)) {
|
||
throw new LlmError("LLM API Key 未配置", 503);
|
||
}
|
||
|
||
const controller = new AbortController();
|
||
const timer = setTimeout(() => controller.abort(), config.timeoutMs);
|
||
|
||
try {
|
||
const response = await fetch(config.apiUrl, {
|
||
method: "POST",
|
||
headers: {
|
||
"Content-Type": "application/json",
|
||
Authorization: `Bearer ${config.apiKey}`,
|
||
},
|
||
body: JSON.stringify({
|
||
model: config.model,
|
||
messages,
|
||
temperature: options?.temperature ?? 0.3,
|
||
max_tokens: options?.maxTokens ?? 2048,
|
||
stream: false,
|
||
// 思考强度交给调用方按内容长度决定;不传表示由供应商默认(DeepSeek 为 high)
|
||
...(options?.reasoningEffort ? { reasoning_effort: options.reasoningEffort } : {}),
|
||
}),
|
||
signal: controller.signal,
|
||
});
|
||
|
||
if (!response.ok) {
|
||
const body = await response.text().catch(() => "");
|
||
const retryAfterMs = parseRetryAfter(response.headers.get("retry-after"));
|
||
// 保留上游真实状态码:重试策略靠它区分 429/5xx(可重试)与 4xx(不可重试)
|
||
throw new LlmError(
|
||
`LLM 请求失败 (${response.status}): ${body.slice(0, 200)}`,
|
||
response.status,
|
||
retryAfterMs,
|
||
);
|
||
}
|
||
|
||
const data = (await response.json()) as {
|
||
model?: string;
|
||
choices?: { message?: { content?: string }; finish_reason?: string }[];
|
||
usage?: unknown;
|
||
};
|
||
// 先把用量取出来:即使下面因为内容为空而抛错,这次请求也已经计费了
|
||
const snapshot = normalizeUsage(data.usage);
|
||
const content = data.choices?.[0]?.message?.content;
|
||
if (!content) {
|
||
throw new LlmError("LLM 返回内容为空", 502, undefined, snapshot.usage);
|
||
}
|
||
return {
|
||
content,
|
||
usage: snapshot.usage,
|
||
usageSource: snapshot.source,
|
||
model: data.model ?? config.model,
|
||
finishReason: data.choices?.[0]?.finish_reason ?? null,
|
||
};
|
||
} catch (err) {
|
||
if (err instanceof LlmError) throw err;
|
||
if (err instanceof Error && err.name === "AbortError") {
|
||
throw new LlmError("LLM 请求超时", 504);
|
||
}
|
||
throw new LlmError(`LLM 请求异常: ${err instanceof Error ? err.message : String(err)}`, 502);
|
||
} finally {
|
||
clearTimeout(timer);
|
||
}
|
||
}
|