多轮图片记忆:
- 此前历史轮次的图片不回传 LLM(attachments 仅存压缩 preview,历史组装
时被丢弃)— 跨轮对话中模型对图片内容"失忆"
- 修复:SessionSummaryService.buildHistoryMessages 从持久化的 attachments
恢复 images(type=image 的 preview base64),历史图片随上下文回传
- Token 控制:最多注入最近 10 张(MAX_HISTORY_IMAGES,从最新消息向前
收集)— 每张 1024px 压缩图约数百至千余 token,无上限会吃满上下文
- 摘要区间(summarizedUntilRowid 之前)的图片不恢复,符合滚动摘要语义
多模态总开关 llm.multimodalEnabled(默认关闭):
- 新配置项:CONFIG_DEFAULTS 种子 + 设置弹框 LLM 配置 Switch +
首次引导向导 LLM 步骤 Switch(含说明文案)
- 上传入口双重判断:总开关 × 模型能力 — 未开启时即使模型支持多模态
也不能上传图片(ChatInput 的选择/拖拽/粘贴统一拦截,Toast 区分
"开关未开启"与"当前模型不支持"两种原因)
- 保存成功后同步 Agent Store 立即生效;App 启动时随 setProvider 加载
DeepSeek vision 模型支持:
- 新增 deepseek-v4-flash-vision-exp(OpenAI image_url content parts 格式,
128K 上下文 / 8K 输出)
- adapter 按 isVisionModel() 判断:vision 模型将带 images 的消息转换为
[{type:'text'},{type:'image_url'}] parts;非 vision 模型保持 images
静默丢弃(防 API 400)
测试(236 → 243 用例):
- 多轮图片记忆 ×3(session-summary.test.ts):历史 attachments 恢复
images / 上限 10 张从最新向前 / 摘要区间图片不恢复
- DeepSeek vision 请求格式 ×4(deepseek-vision.test.ts,契约级 mock
fetch 断言请求体):image_url parts 转换 / 非 vision 模型丢弃 /
max_tokens 钳制 8192 / 无图不转换
- 测试顺序修正:多轮图片用例置于 describe 末尾(插入新行消耗全局自增
rowid,插在中间会破坏既有用例对 rowid 数值的断言)
文档: README 同步(DeepSeek 模型表 + vision 多模态列、llm.multimodalEnabled
配置项、多轮图片记忆特性行、243 用例数)
验证: lint 0 / typecheck 双工程 0 / test:electron 243 全过 / build 成功
341 lines
12 KiB
TypeScript
341 lines
12 KiB
TypeScript
/**
|
||
* DeepSeek Provider Adapter
|
||
*
|
||
* 基于 OpenAI 兼容 API。支持 Tool Calling、Thinking 模式、流式输出。
|
||
* 模型: deepseek-v4-flash / deepseek-v4-pro(1M 上下文,384K 最大输出)
|
||
*
|
||
* 独立继承 BaseAdapter,通过 shared/openai-format 和 shared/sse-stream 复用
|
||
* OpenAI 兼容格式构建和 SSE 流式解析逻辑。不与其他 Provider Adapter 耦合。
|
||
*
|
||
* @see apis/deepseek-api-docs-20260518.html
|
||
*/
|
||
|
||
import log from 'electron-log';
|
||
import { BaseAdapter } from './base-adapter';
|
||
import type { MetonaRequest, MetonaResponse, MetonaStreamEvent } from '../types';
|
||
import { MetonaFinishReason } from '../types';
|
||
import type { MetonaModelInfo } from '../types/metona-adapter';
|
||
import { buildOpenAICompatibleMessages, buildOpenAICompatibleTools } from './shared/openai-format';
|
||
import { parseSSEStream, parseOpenAICompatibleResponse } from './shared/sse-stream';
|
||
|
||
export class DeepSeekAdapter extends BaseAdapter {
|
||
// H-2 修复: provider → providerId(规范要求)
|
||
override readonly providerId: string = 'deepseek';
|
||
readonly supportedModels = [
|
||
'deepseek-v4-pro',
|
||
'deepseek-v4-flash',
|
||
'deepseek-v4-flash-vision-exp',
|
||
];
|
||
readonly supportsToolCalling = true;
|
||
readonly supportsThinking = true;
|
||
|
||
// H-2 修复: DeepSeek 模型元信息(1M 上下文,384K 最大输出)
|
||
private static readonly MODEL_INFO: Record<string, MetonaModelInfo> = {
|
||
'deepseek-v4-pro': {
|
||
id: 'deepseek-v4-pro',
|
||
name: 'DeepSeek V4 Pro',
|
||
contextWindow: 1_000_000,
|
||
maxOutputTokens: 384_000,
|
||
supportsToolCalling: true,
|
||
supportsThinking: true,
|
||
description: 'DeepSeek 旗舰模型,1M 上下文,支持深度推理与工具调用',
|
||
},
|
||
'deepseek-v4-flash': {
|
||
id: 'deepseek-v4-flash',
|
||
name: 'DeepSeek V4 Flash',
|
||
contextWindow: 1_000_000,
|
||
maxOutputTokens: 384_000,
|
||
supportsToolCalling: true,
|
||
supportsThinking: true,
|
||
description: 'DeepSeek 快速版,1M 上下文,低延迟推理',
|
||
},
|
||
// v0.5.4: DeepSeek 多模态实验模型(OpenAI image_url content parts 格式)
|
||
'deepseek-v4-flash-vision-exp': {
|
||
id: 'deepseek-v4-flash-vision-exp',
|
||
name: 'DeepSeek V4 Flash Vision (Exp)',
|
||
contextWindow: 128_000,
|
||
maxOutputTokens: 8_192,
|
||
supportsToolCalling: true,
|
||
supportsThinking: false,
|
||
description: 'DeepSeek 多模态实验模型,支持图片输入(image_url content parts)',
|
||
},
|
||
};
|
||
|
||
/**
|
||
* v0.5.4: 当前模型是否支持多模态图片输入
|
||
*
|
||
* DeepSeek 仅 vision 系列模型支持图片(命名含 'vision');
|
||
* 非 vision 模型收到 images 时静默丢弃(避免 API 400)。
|
||
* 前端上传入口由 llm.multimodalEnabled 配置总开关控制,此处是 adapter 侧的模型级防线。
|
||
*/
|
||
private isVisionModel(): boolean {
|
||
return this.config.defaultModel.includes('vision');
|
||
}
|
||
|
||
// ===== POST /chat/completions (非流式) =====
|
||
|
||
// H-2 修复: chat → send(规范要求)
|
||
async send(request: MetonaRequest): Promise<MetonaResponse> {
|
||
const body = this.toNativeRequest(request, false);
|
||
|
||
// #24 修复: 使用 fetchWithTimeout 替代 getFetchSignal + fetch,确保 timer 清理
|
||
const response = await this.fetchWithTimeout(
|
||
`${this.config.baseURL}/chat/completions`,
|
||
{
|
||
method: 'POST',
|
||
headers: {
|
||
'Content-Type': 'application/json',
|
||
Authorization: `Bearer ${this.config.apiKey}`,
|
||
...this.config.headers,
|
||
},
|
||
body: JSON.stringify(body),
|
||
},
|
||
this.config.timeoutMs ?? 120_000,
|
||
);
|
||
|
||
if (!response.ok) {
|
||
await this.throwHttpError(response, 'DeepSeek API error');
|
||
}
|
||
|
||
const data = (await response.json()) as Record<string, unknown>;
|
||
const parsed = parseOpenAICompatibleResponse(data);
|
||
|
||
return {
|
||
meta: {
|
||
requestId: request.meta.requestId,
|
||
provider: this.providerId,
|
||
model: (data.model as string) ?? this.config.defaultModel,
|
||
latencyMs: 0,
|
||
timestamp: Date.now(),
|
||
},
|
||
content: parsed.content,
|
||
reasoningContent: parsed.reasoningContent,
|
||
toolCalls: parsed.toolCalls,
|
||
usage: parsed.usage,
|
||
finishReason: parsed.finishReason as MetonaFinishReason,
|
||
};
|
||
}
|
||
|
||
// ===== POST /chat/completions (流式) =====
|
||
|
||
// H-2 修复: chatStream → sendStream(规范要求)
|
||
async *sendStream(request: MetonaRequest): AsyncIterable<MetonaStreamEvent> {
|
||
const body = this.toNativeRequest(request, true);
|
||
|
||
// #24 修复: 使用 fetchWithTimeout 替代 getFetchSignal + fetch,确保 timer 清理
|
||
const response = await this.fetchWithTimeout(
|
||
`${this.config.baseURL}/chat/completions`,
|
||
{
|
||
method: 'POST',
|
||
headers: {
|
||
'Content-Type': 'application/json',
|
||
Authorization: `Bearer ${this.config.apiKey}`,
|
||
...this.config.headers,
|
||
},
|
||
body: JSON.stringify(body),
|
||
},
|
||
this.config.timeoutMs ?? 300_000,
|
||
);
|
||
|
||
if (!response.ok || !response.body) {
|
||
await this.throwHttpError(response, 'DeepSeek stream error');
|
||
}
|
||
|
||
yield* parseSSEStream(
|
||
// 非空断言:上方 if 已确保 response.body 不为 null
|
||
// TypeScript 无法通过 await Promise<never> 正确收窄,需显式断言
|
||
response.body!,
|
||
request.meta.requestId,
|
||
request.meta.sessionId,
|
||
request.meta.iteration,
|
||
);
|
||
}
|
||
|
||
// ===== GET /models =====
|
||
|
||
/**
|
||
* H-2 修复: 返回 MetonaModelInfo[](规范要求)
|
||
*
|
||
* 优先尝试从 API 获取实时模型列表,并合并本地 MODEL_INFO 元数据。
|
||
* API 不可用时回退到 supportedModels。
|
||
*/
|
||
async listModels(): Promise<MetonaModelInfo[]> {
|
||
try {
|
||
const response = await fetch(`${this.config.baseURL}/models`, {
|
||
headers: { Authorization: `Bearer ${this.config.apiKey}` },
|
||
signal: AbortSignal.timeout(10_000),
|
||
});
|
||
if (response.ok) {
|
||
const data = (await response.json()) as { data?: Array<{ id: string }> };
|
||
if (data.data?.length) {
|
||
// 合并 API 返回的模型 ID 与本地元数据
|
||
return data.data.map((m) => DeepSeekAdapter.MODEL_INFO[m.id] ?? { id: m.id });
|
||
}
|
||
}
|
||
} catch {
|
||
// API 不可用时降级
|
||
}
|
||
// 回退到 supportedModels(带本地元数据)
|
||
return this.supportedModels.map((id) => DeepSeekAdapter.MODEL_INFO[id] ?? { id });
|
||
}
|
||
|
||
/**
|
||
* H-2 修复: 获取上下文窗口大小(规范要求)
|
||
*
|
||
* v0.3.1: 优先使用配置注入的 contextWindow,回退到 MODEL_INFO 默认值。
|
||
* DeepSeek OpenAI 兼容 API 不支持 context_window 参数,此值仅用于
|
||
* Engine 压缩判断和前端 UI 显示。
|
||
*/
|
||
override getContextWindow(): number {
|
||
// v0.3.1: 优先使用配置注入的 contextWindow
|
||
if (typeof this.config.contextWindow === 'number' && this.config.contextWindow > 0) {
|
||
return this.config.contextWindow;
|
||
}
|
||
// 回退到 MODEL_INFO
|
||
const modelInfo = DeepSeekAdapter.MODEL_INFO[this.config.defaultModel];
|
||
return modelInfo?.contextWindow ?? 1_000_000;
|
||
}
|
||
|
||
// ===== GET /user/balance =====
|
||
|
||
/**
|
||
* 查询账户余额
|
||
*
|
||
* v0.5.2 修复: DeepSeek 官方 API 实际返回 `balance_infos` 数组格式:
|
||
* { "is_available": true, "balance_infos": [{ "currency": "CNY",
|
||
* "total_balance": "110.00", "granted_balance": "10.00", "topped_up_balance": "100.00" }] }
|
||
* 此前按扁平字段解析(data.total_balance)→ 永远取到 undefined → 恒显示 0。
|
||
* 现优先取 balance_infos[0],回退扁平格式(兼容网关/代理的简化响应)。
|
||
*
|
||
* URL 规范化: 余额端点为 {root}/user/balance(无 /v1 前缀)。用户配置的
|
||
* baseURL 可能带 /v1 或尾斜杠(chat 端点两种写法都合法),此处剥离后拼接。
|
||
*/
|
||
async getBalance(): Promise<{
|
||
currency: string;
|
||
totalBalance: string;
|
||
grantedBalance: string;
|
||
toppedUpBalance: string;
|
||
} | null> {
|
||
try {
|
||
// 规范化 baseURL:去尾斜杠、去尾 /v1(余额端点在根路径下)
|
||
const root = this.config.baseURL.replace(/\/+$/, '').replace(/\/v1$/, '');
|
||
const response = await fetch(`${root}/user/balance`, {
|
||
headers: { Authorization: `Bearer ${this.config.apiKey}` },
|
||
signal: AbortSignal.timeout(10_000),
|
||
});
|
||
if (!response.ok) return null;
|
||
const data = (await response.json()) as {
|
||
is_available?: boolean;
|
||
balance_infos?: Array<{
|
||
currency?: string;
|
||
total_balance?: string;
|
||
granted_balance?: string;
|
||
topped_up_balance?: string;
|
||
}>;
|
||
// 扁平格式字段(网关/代理兼容)
|
||
currency?: string;
|
||
total_balance?: string;
|
||
granted_balance?: string;
|
||
topped_up_balance?: string;
|
||
};
|
||
// 优先官方 balance_infos 数组,回退扁平格式
|
||
const info = data.balance_infos?.[0] ?? data;
|
||
return {
|
||
currency: info.currency ?? 'CNY',
|
||
totalBalance: info.total_balance ?? '0',
|
||
grantedBalance: info.granted_balance ?? '0',
|
||
toppedUpBalance: info.topped_up_balance ?? '0',
|
||
};
|
||
} catch {
|
||
return null;
|
||
}
|
||
}
|
||
|
||
// ========== 私有方法 ==========
|
||
|
||
/**
|
||
* 构建 DeepSeek 原生请求体
|
||
*
|
||
* DeepSeek 特有参数:
|
||
* - thinking: { type: "enabled" } — 启用思考模式
|
||
* - reasoning_effort — 思考强度映射
|
||
* - stream_options: { include_usage: true } — 流式返回 usage
|
||
*/
|
||
private toNativeRequest(request: MetonaRequest, stream: boolean): Record<string, unknown> {
|
||
const messages = buildOpenAICompatibleMessages(request);
|
||
const tools = buildOpenAICompatibleTools(request.tools);
|
||
|
||
// v0.5.3: max_tokens 按模型上限钳制 — 引擎默认 63488 超过部分模型上限时 API 直接 400
|
||
const modelInfo = DeepSeekAdapter.MODEL_INFO[this.config.defaultModel];
|
||
const maxOutput = modelInfo?.maxOutputTokens ?? 384_000;
|
||
const maxTokens = Math.min(request.params.maxTokens ?? maxOutput, maxOutput);
|
||
|
||
const body: Record<string, unknown> = {
|
||
model: this.config.defaultModel,
|
||
messages,
|
||
temperature: request.params.temperature,
|
||
max_tokens: maxTokens,
|
||
stream,
|
||
};
|
||
|
||
// v0.5.4: vision 模型的图片处理(OpenAI image_url content parts 格式)
|
||
// 非 vision 模型保持 images 静默丢弃(共享层行为,避免 API 400)
|
||
if (this.isVisionModel()) {
|
||
const nonSystemMsgs = request.messages.filter((m) => m.role !== 'system');
|
||
let imageCount = 0;
|
||
// messages[0] 是 system,非 system 消息从 messages[1] 开始(与 nonSystemMsgs 对齐)
|
||
for (let i = 1; i < messages.length; i++) {
|
||
const origMsg = nonSystemMsgs[i - 1];
|
||
if (!origMsg?.images?.length) continue;
|
||
|
||
imageCount += origMsg.images.length;
|
||
const contentParts: Array<Record<string, unknown>> = [];
|
||
if (origMsg.content) {
|
||
contentParts.push({ type: 'text', text: origMsg.content });
|
||
}
|
||
for (const img of origMsg.images) {
|
||
contentParts.push({
|
||
type: 'image_url',
|
||
image_url: { url: img.url },
|
||
});
|
||
}
|
||
messages[i].content = contentParts;
|
||
}
|
||
if (imageCount > 0) {
|
||
log.info(`[DeepSeek] Vision model processing ${imageCount} image(s)`);
|
||
}
|
||
}
|
||
|
||
if (stream) {
|
||
body.stream_options = { include_usage: true };
|
||
}
|
||
|
||
if (tools) {
|
||
body.tools = tools;
|
||
}
|
||
|
||
// Thinking 模式
|
||
// API 默认 thinking.type = "enabled",必须显式发送 disabled 才能关闭
|
||
if (request.params.thinkingEnabled === false) {
|
||
body.thinking = { type: 'disabled' };
|
||
} else if (request.params.thinkingEnabled) {
|
||
body.thinking = { type: 'enabled' };
|
||
const effortMap: Record<string, string> = {
|
||
low: 'high',
|
||
medium: 'high',
|
||
high: 'high',
|
||
max: 'max',
|
||
};
|
||
// DeepSeek API 仅支持 high / max 两档,low/medium 映射为 high
|
||
body.reasoning_effort = effortMap[request.params.thinkingEffort ?? 'high'] ?? 'high';
|
||
}
|
||
|
||
// 停止序列
|
||
if (request.params.stopSequences?.length) {
|
||
body.stop = request.params.stopSequences;
|
||
}
|
||
|
||
return body;
|
||
}
|
||
}
|