189 lines
7.1 KiB
TypeScript
189 lines
7.1 KiB
TypeScript
/**
|
||
* OpenAI Provider Adapter（P3）
|
||
*
|
||
* OpenAI Chat Completions API（/v1/chat/completions），支持 Tool Calling、
|
||
* 流式输出、多模态图片、o 系列推理模型的 reasoning_effort 参数。
|
||
*
|
||
* v0.6.4 P3-1: 继承 OpenAICompatibleAdapter —— 传输/组装/回退链收敛到共享基类，
|
||
* 本文件只保留 OpenAI 差异点：o 系列/gpt-5 的字段名路由与 reasoning_effort、
|
||
* 推理模型拒图的前置拦截（升级为 ModelCapabilityError）、动态 /models 列表、
|
||
* 非推理模型 temperature 控制。
|
||
*
|
||
* @see https://platform.openai.com/docs/api-reference/chat
|
||
*/
|
||
|
||
import log from 'electron-log';
|
||
import type { MetonaRequest } from '../types';
|
||
import type { MetonaModelInfo } from '../types/metona-adapter';
|
||
import { buildOpenAICompatibleMessages, buildOpenAICompatibleTools } from './shared/openai-format';
|
||
import { ModelCapabilityError, OpenAICompatibleAdapter } from './shared/openai-compatible-base';
|
||
|
||
export class OpenAIAdapter extends OpenAICompatibleAdapter {
|
||
override readonly providerId: string = 'openai';
|
||
readonly supportedModels = ['gpt-4o', 'gpt-4o-mini', 'gpt-4.1', 'o3-mini'];
|
||
readonly supportsToolCalling = true;
|
||
readonly supportsThinking = true;
|
||
|
||
// v0.8.1: 仅承载展示与能力声明 —— 窗口/输出上限数值已按硬性契约删除，
|
||
// 唯一合法来源是设置面板 llm.contextWindow / llm.maxTokens
|
||
private static readonly MODEL_INFO: Record<string, MetonaModelInfo> = {
|
||
'gpt-4o': {
|
||
id: 'gpt-4o',
|
||
name: 'GPT-4o',
|
||
supportsToolCalling: true,
|
||
supportsThinking: false,
|
||
description: 'OpenAI 旗舰多模态模型',
|
||
},
|
||
'gpt-4o-mini': {
|
||
id: 'gpt-4o-mini',
|
||
name: 'GPT-4o mini',
|
||
supportsToolCalling: true,
|
||
supportsThinking: false,
|
||
description: 'OpenAI 高性价比模型',
|
||
},
|
||
'gpt-4.1': {
|
||
id: 'gpt-4.1',
|
||
name: 'GPT-4.1',
|
||
supportsToolCalling: true,
|
||
supportsThinking: false,
|
||
description: 'OpenAI 长上下文模型',
|
||
},
|
||
'o3-mini': {
|
||
id: 'o3-mini',
|
||
name: 'o3-mini',
|
||
supportsToolCalling: true,
|
||
supportsThinking: true,
|
||
description: 'OpenAI 推理模型，支持 reasoning_effort',
|
||
},
|
||
};
|
||
|
||
// ===== 共享基类差异声明 =====
|
||
|
||
protected override chatCompletionsUrl(): string {
|
||
return `${this.config.baseURL}/chat/completions`;
|
||
}
|
||
|
||
protected override sendTimeoutMs(): number {
|
||
return 120_000;
|
||
}
|
||
|
||
protected override modelInfoTable(): Record<string, MetonaModelInfo> {
|
||
return OpenAIAdapter.MODEL_INFO;
|
||
}
|
||
|
||
protected override providerLabel(): string {
|
||
return 'OpenAI';
|
||
}
|
||
|
||
// ===== GET /v1/models =====
|
||
|
||
override async listModels(): Promise<MetonaModelInfo[]> {
|
||
try {
|
||
const response = await fetch(`${this.config.baseURL}/models`, {
|
||
headers: { Authorization: `Bearer ${this.config.apiKey}` },
|
||
signal: AbortSignal.timeout(10_000),
|
||
});
|
||
if (response.ok) {
|
||
const data = (await response.json()) as { data?: Array<{ id: string }> };
|
||
if (data.data?.length) {
|
||
return data.data.map((m) => OpenAIAdapter.MODEL_INFO[m.id] ?? { id: m.id });
|
||
}
|
||
}
|
||
} catch {
|
||
// API 不可用时降级
|
||
}
|
||
return this.supportedModels.map((id) => OpenAIAdapter.MODEL_INFO[id] ?? { id });
|
||
}
|
||
|
||
// ========== 协议参数映射（OpenAI 差异点） ==========
|
||
|
||
protected override toNativeRequest(
|
||
request: MetonaRequest,
|
||
stream: boolean,
|
||
): Record<string, unknown> {
|
||
// 推理模型检测（o 系列使用新参数名）
|
||
// v0.8.2 P2-6: 正则补边界 — 旧 /^(o\d|gpt-5)/ 对命名变体存在两处漏判/误判：
|
||
// ① `chatgpt-5-*` 系列（gpt-5 非前缀）漏判；② 未来 `o10` 等多位数编号无影响
|
||
// 但 `o\d` 会把 `openai/...` 路径形态误判（当前配置层不透传路径，防御性加边界）。
|
||
// 现改为带词边界的枚举前缀匹配（o<digits> / gpt-5 / chatgpt-5）。
|
||
const model = this.config.defaultModel;
|
||
const isReasoningModel = /^(o\d+(?:[-.][\w.]+)?|gpt-5[\w.-]*|chatgpt-5[\w.-]*)/.test(model);
|
||
|
||
// 推理模型不支持图片输入 — 前置校验
|
||
// v0.6.4 升级: 原实现抛裸 Error 落入 UNKNOWN 错误码；现在抛 ModelCapabilityError
|
||
// （携带 status=400），引擎按"不可重试请求级错误"处理，UI 可区分能力限制与一般故障。
|
||
if (isReasoningModel) {
|
||
const hasImages = request.messages.some((m) => m.images?.length);
|
||
if (hasImages) {
|
||
throw new ModelCapabilityError(model, 'image inputs');
|
||
}
|
||
}
|
||
|
||
// v0.6.2: images 处理收敛至共享层（原索引对齐循环在孤立 tool 过滤后会错位）
|
||
const messages = buildOpenAICompatibleMessages(request, true);
|
||
const tools = buildOpenAICompatibleTools(request.tools);
|
||
|
||
const body: Record<string, unknown> = {
|
||
model,
|
||
messages,
|
||
stream,
|
||
};
|
||
|
||
// Token 上限参数：o 系列/gpt-5 使用 max_completion_tokens
|
||
// v0.8.1 硬性契约: 原样透传设置面板「最大输出上限」（llm.maxTokens），删除了
|
||
// 旧的按模型元信息钳制逻辑与未配置时的 32_768 兜底 —— 未配置时不下发该字段。
|
||
if (request.params.maxTokens != null) {
|
||
if (isReasoningModel) {
|
||
body.max_completion_tokens = request.params.maxTokens;
|
||
} else {
|
||
body.max_tokens = request.params.maxTokens;
|
||
}
|
||
}
|
||
|
||
if (stream) {
|
||
body.stream_options = { include_usage: true };
|
||
}
|
||
|
||
if (tools) {
|
||
body.tools = tools;
|
||
}
|
||
|
||
// Thinking 模式：推理模型映射 reasoning_effort；非推理模型忽略
|
||
if (request.params.thinkingEnabled && isReasoningModel) {
|
||
const effortMap: Record<string, string> = {
|
||
low: 'low',
|
||
medium: 'medium',
|
||
high: 'high',
|
||
max: 'high',
|
||
};
|
||
body.reasoning_effort = effortMap[request.params.thinkingEffort ?? 'high'] ?? 'high';
|
||
// v0.8.0 P0-3: 推理 token 计入 max_completion_tokens —— 用户配置的预算过小时告警
|
||
if (typeof request.params.maxTokens === 'number' && request.params.maxTokens < 8192) {
|
||
log.warn(
|
||
`[OpenAI] reasoning enabled with small output budget (${request.params.maxTokens} tokens per user config) — reasoning may consume the entire budget and truncate the answer`,
|
||
);
|
||
}
|
||
} else if (!isReasoningModel) {
|
||
// 非推理模型使用温度控制
|
||
body.temperature = request.params.temperature;
|
||
}
|
||
|
||
// 停止序列
|
||
// v0.8.2 P2-6 根治: 推理模型不支持 stop 参数 —— 旧实现"仍透传，400 属上游约束"
|
||
// 的注释性放弃改为前置拦截：推理模型请求丢弃 stop 并告警（与拒图的
|
||
// ModelCapabilityError 同思路，但 stop 是可选增强参数，静默丢弃 + 告警优于
|
||
// 让整个请求 400 失败）。
|
||
if (request.params.stopSequences?.length) {
|
||
if (isReasoningModel) {
|
||
log.warn(
|
||
`[OpenAI] stop sequences are not supported by reasoning model "${model}" — dropping ${request.params.stopSequences.length} stop sequence(s) for this request`,
|
||
);
|
||
} else {
|
||
body.stop = request.params.stopSequences;
|
||
}
|
||
}
|
||
|
||
return body;
|
||
}
|
||
}
|