@nickyzj2023/ai 1.4.4 → 1.6.0

package/dist/{src-CuZ-6Jye.mjs → src-CKKPKPMG.mjs}

@@ -1,6 +1,5 @@
-import { createXMLText, extractErrorMessage, fetcher, isObject, logger, omit, parseSSE, pick, to } from "@nickyzj2023/utils";
-import { Client } from "@modelcontextprotocol/sdk/client";
-import { StreamableHTTPClientTransport } from "@modelcontextprotocol/sdk/client/streamableHttp.js";
+import { i as estimateTokens, n as defineTool, r as estimateTextTokens } from "./helper-juLVGxlQ.mjs";
+import { createXMLText, extractErrorMessage, fetcher, logger, parseSSE, pick, to } from "@nickyzj2023/utils";
 //#region src/llm.ts
 /**
 * 分离ToolDefinition中的valid/invalid字段，前者可以传给模型，后者用于本地运算
@@ -22,7 +21,7 @@ const extractReasoning = (msgLike) => {
 /**
 * 流式请求模型，转发
 */
-async function* stream(model, messages, tools = []) {
+async function* stream(model, messages, tools = [], options) {
 	const validTools = tools.map((tool) => detachToolArguments(tool)[0]);
 	const api = fetcher(model.baseUrl, {
 		headers: { Authorization: `Bearer ${model.apiKey}` },
@@ -32,9 +31,11 @@ async function* stream(model, messages, tools = []) {
 		stream: true,
 		model: model.model,
 		messages,
-		tools: validTools
-	}));
+		tools: validTools,
+		...options?.body
+	}, options));
 	if (error) {
+		if (options?.signal?.aborted) return;
 		yield {
 			type: "error",
 			message: error.message
@@ -44,39 +45,48 @@ async function* stream(model, messages, tools = []) {
 	const toolCallBuffers = /* @__PURE__ */ new Map();
 	let usage;
 	let finishReason = null;
-	for await (const chunk of parseSSE(response)) {
-		if (typeof chunk === "string") continue;
-		if (chunk.usage) usage = chunk.usage;
-		const choice = chunk.choices?.[0];
-		if (!choice) continue;
-		const { delta } = choice;
-		const { content: contentDelta, tool_calls: toolCalls } = delta;
-		const reasoning = extractReasoning(delta);
-		if (reasoning) yield {
-			type: "reasoning_delta",
-			delta: reasoning
-		};
-		if (contentDelta) yield {
-			type: "content_delta",
-			delta: contentDelta.toString()
-		};
-		if (toolCalls) for (const call of toolCalls) {
-			const { index = 0, type = "function", id, function: fn, ...extra } = call;
-			const existing = toolCallBuffers.getOrInsert(index, {
-				id: "",
-				type,
-				function: {
-					name: "",
-					arguments: ""
-				}
-			});
-			if (id) existing.id = id;
-			if (fn?.name) existing.function.name += fn.name;
-			if (fn?.arguments) existing.function.arguments += fn.arguments;
-			if (extra) Object.assign(existing, extra);
-			toolCallBuffers.set(index, existing);
+	try {
+		for await (const chunk of parseSSE(response)) {
+			if (typeof chunk === "string") continue;
+			if (chunk.usage) usage = chunk.usage;
+			const choice = chunk.choices?.[0];
+			if (!choice) continue;
+			const { delta } = choice;
+			const { content: contentDelta, tool_calls: toolCalls } = delta;
+			const reasoning = extractReasoning(delta);
+			if (reasoning) yield {
+				type: "reasoning_delta",
+				delta: reasoning
+			};
+			if (contentDelta) yield {
+				type: "content_delta",
+				delta: contentDelta.toString()
+			};
+			if (toolCalls) for (const call of toolCalls) {
+				const { index = 0, type = "function", id, function: fn, ...extra } = call;
+				const existing = toolCallBuffers.getOrInsert(index, {
+					id: "",
+					type,
+					function: {
+						name: "",
+						arguments: ""
+					}
+				});
+				if (id) existing.id = id;
+				if (fn?.name) existing.function.name += fn.name;
+				if (fn?.arguments) existing.function.arguments += fn.arguments;
+				if (extra) Object.assign(existing, extra);
+				toolCallBuffers.set(index, existing);
+			}
+			if (choice.finish_reason) finishReason = choice.finish_reason;
 		}
-		if (choice.finish_reason) finishReason = choice.finish_reason;
+	} catch (e) {
+		if (options?.signal?.aborted) return;
+		yield {
+			type: "error",
+			message: extractErrorMessage(e)
+		};
+		return;
 	}
 	for (const [, call] of toolCallBuffers) yield {
 		type: "tool_call",
@@ -96,27 +106,48 @@ async function* runAgent(model, messages, tools) {
 	const toolMap = new Map(tools.map((tool) => [tool.function.name, tool]));
 	while (true) {
 		let content = "";
+		let reasoning = "";
 		const toolCalls = [];
-		for await (const e of stream(model, messages, tools)) switch (e.type) {
-			case "reasoning_delta":
-				yield e;
-				break;
-			case "content_delta":
-				content += e.delta;
-				yield e;
-				break;
-			case "tool_call":
-				toolCalls.push({
-					id: e.id,
-					type: "function",
-					function: {
-						name: e.name,
-						arguments: e.args
+		const promptMessage = messages.slice(-1);
+		let startTime = 0;
+		for await (const e of stream(model, messages, tools)) {
+			startTime ||= Date.now();
+			switch (e.type) {
+				case "reasoning_delta":
+					reasoning += e.delta;
+					yield e;
+					break;
+				case "content_delta":
+					content += e.delta;
+					yield e;
+					break;
+				case "tool_call":
+					toolCalls.push({
+						id: e.id,
+						type: "function",
+						function: {
+							name: e.name,
+							arguments: e.args
+						}
+					});
+					yield e;
+					break;
+				case "error":
+					yield e;
+					break;
+				case "done":
+					if (!e.usage) {
+						const promptTokens = estimateTokens(promptMessage);
+						const completionTokens = estimateTextTokens(content + reasoning);
+						e.usage = {
+							prompt_tokens: promptTokens,
+							completion_tokens: completionTokens,
+							total_tokens: estimateTokens(messages) + promptTokens + completionTokens
+						};
 					}
-				});
-				yield e;
-				break;
-			case "done": yield e;
+					e.usage.speed = Math.floor(e.usage.completion_tokens / ((Date.now() - startTime) / 1e3));
+					yield e;
+			}
 		}
 		const message = {
 			role: "assistant",
@@ -149,70 +180,6 @@ async function* runAgent(model, messages, tools) {
 	}
 }
 //#endregion
-//#region src/utils/helper.ts
-/**
-* 辅助定义一个POST /chat/completions支持的model参数
-* @remarks 只有baseUrl字段是必须的
-*/
-const defineModel = (config) => ({
-	modalities: ["text"],
-	context: 131072,
-	...config
-});
-/**
-* 辅助定义一个POST /chat/completions支持的tool对象
-* @param execute 实际执行工具的函数
-*/
-const defineTool = (name, description, properties, execute) => {
-	const _required = [];
-	return {
-		type: "function",
-		function: {
-			name,
-			description,
-			parameters: {
-				type: "object",
-				properties: Object.entries(properties).reduce((result, [key, property]) => {
-					if ("required" in property) {
-						_required.push(key);
-						delete property.required;
-					}
-					result[key] = property;
-					return result;
-				}, {}),
-				required: _required
-			}
-		},
-		execute
-	};
-};
-/**
-* 根据上下文里的中/英文/多模态消息，估算出可能消耗的token
-* - 单词 ≈ 1.5token
-* - 标点/空白等非词字符每 4 个 ≈ 1token
-* - 图片/音频/视频/文件 ≈ 4096token（不好估算，取个较大的值）
-*/
-const estimateTokens = (messages) => {
-	if (!messages?.length) return 0;
-	const segmenter = new Intl.Segmenter([], { granularity: "word" });
-	const estimateTextTokens = (text) => {
-		let words = 0;
-		let others = 0;
-		for (const seg of segmenter.segment(text)) if (seg.isWordLike) words++;
-		else others++;
-		return Math.ceil(words * 1.5 + others / 4);
-	};
-	return messages.reduce((acc, message) => {
-		const { content, tool_calls, ...metadata } = message;
-		if (typeof content === "string") acc += estimateTextTokens(content);
-		else for (const part of content) if (part.type === "text") acc += estimateTextTokens(part.text);
-		else acc += 4096;
-		if (tool_calls) acc += estimateTextTokens(JSON.stringify(tool_calls));
-		acc += estimateTextTokens(JSON.stringify(metadata));
-		return acc;
-	}, 0);
-};
-//#endregion
 //#region src/tools/get-time.ts
 var get_time_default = defineTool("get_time", "查询指定时区的当前时间", { timezone: {
 	type: "string",
@@ -231,60 +198,6 @@ var get_weather_default = defineTool("get_weather", "查询指定城市的天气
 	return fetcher("https://wttr.in", { params: { format: "j1" } }).get(`/${city}`);
 });
 //#endregion
-//#region src/tools/mcp.ts
-/** 全局单例MCP加载器 */
-let router = null;
-var MCPRouter = class {
-	entries = /* @__PURE__ */ new Map();
-	/** 注册一个新的MCP客户端 */
-	async addClient(name, url, options) {
-		if (this.entries.has(name)) return;
-		const transport = new StreamableHTTPClientTransport(new URL(url), { requestInit: { headers: options?.headers } });
-		const client = new Client({
-			name,
-			version: "1.0.0"
-		});
-		await client.connect(transport);
-		const { tools } = await client.listTools();
-		const normalizedTools = tools.filter((tool) => !options?.ignoredToolNames?.includes(tool.name)).map((tool) => {
-			const _properties = { ...tool.inputSchema.properties ?? {} };
-			tool.inputSchema.required?.forEach((key) => {
-				if (isObject(_properties[key])) _properties[key] = {
-					..._properties[key],
-					required: true
-				};
-			});
-			return defineTool(tool.name, tool.description ?? "", _properties, (args) => client.callTool({
-				name: tool.name,
-				arguments: args
-			}));
-		});
-		this.entries.set(name, {
-			client,
-			tools: normalizedTools
-		});
-		return client;
-	}
-	/** 返回OpenAI API兼容的tools数组 */
-	async getTools() {
-		return [...this.entries.values()].flatMap((e) => e.tools);
-	}
-};
-/**
-* 把传入的MCPServer列表转换成OpenAI API兼容的tools数组
-*/
-const loadMCPTools = async (mcpServers = {}) => {
-	router ||= new MCPRouter();
-	await Promise.allSettled(Object.entries(mcpServers).map(async ([name, server]) => {
-		try {
-			await router?.addClient(name, server.url, omit(server, ["type", "url"]));
-		} catch (e) {
-			logger(`MCP服务器${name}加载失败：${extractErrorMessage(e)}`);
-		}
-	}));
-	return router.getTools();
-};
-//#endregion
 //#region src/utils/compact/helper.ts
 /**
 * 校验assistant(tool_calls)消息
@@ -296,12 +209,11 @@ const isToolCalls = (message) => {
 * 校验多模态消息
 */
 const isMediaMessage = (message) => {
-	const MEDIA_TYPES = [
-		"image_url",
-		"input_audio",
-		"video_url"
-	];
-	return message && Array.isArray(message.content) && message.content.some((part) => MEDIA_TYPES.includes(part.type));
+	if (message && Array.isArray(message.content)) for (const part of message.content) {
+		if (part.type === "image_url") return "image";
+		if (part.type === "input_audio") return "audio";
+		if (part.type === "video_url") return "video";
+	}
 };
 /**
 * 查找assistant(tool_calls) + tool配对组范围
@@ -330,7 +242,8 @@ const defaultReplacerOfToolResultContent = async (content, options) => {
 		content: "请用一句话简述上面这条消息（直击要点，不要添加任何前言后语）"
 	}];
 	let simplifiedContent = "";
-	for await (const e of runAgent(model, messages, [])) if (e.type === "content_delta") simplifiedContent += e.delta;
+	if (!model) simplifiedContent = "工具结果已被消费";
+	else for await (const e of runAgent(model, messages, [])) if (e.type === "content_delta") simplifiedContent += e.delta;
 	else if (e.type === "error") throw new Error(e.message);
 	return simplifiedContent;
 };
@@ -352,7 +265,7 @@ const compactToolResults = async (compressible, options) => {
 };
 /** 默认的压缩多模态消息策略：让大模型精简消息内容 */
 const defaultReplacerOfMediaContent = async (content, options) => {
-	const { model } = options ?? {};
+	const { model, modality = "image" } = options ?? {};
 	const messages = [{
 		role: "user",
 		content
@@ -361,7 +274,8 @@ const defaultReplacerOfMediaContent = async (content, options) => {
 		content: "请用一句话简述上面这条消息（直击要点，不要添加任何前言后语）"
 	}];
 	let simplifiedContent = "";
-	for await (const e of runAgent(model, messages, [])) if (e.type === "content_delta") simplifiedContent += e.delta;
+	if (!model?.modalities?.includes(modality)) simplifiedContent = "媒体资源已被消费";
+	else for await (const e of runAgent(model, messages, [])) if (e.type === "content_delta") simplifiedContent += e.delta;
 	else if (e.type === "error") throw new Error(e.message);
 	return simplifiedContent;
 };
@@ -373,21 +287,28 @@ const compactMediaMessages = async (compressible, options) => {
 	if (!replacer && !model) return 0;
 	const _replacer = replacer || defaultReplacerOfMediaContent;
 	let count = 0;
-	for (const message of compressible) if (isMediaMessage(message)) {
-		const compacted = await _replacer(message.content, { model });
-		message.content = createXMLText("media", compacted);
-		count++;
+	for (const message of compressible) {
+		const mediaModality = isMediaMessage(message);
+		if (mediaModality) {
+			const compacted = await _replacer(message.content, {
+				model,
+				modality: mediaModality
+			});
+			message.content = createXMLText("media", compacted);
+			count++;
+		}
 	}
 	if (count > 0) logger(`压缩了${count}条多模态消息`);
 	return count;
 };
 const summarizeMessages = async (compressible, options) => {
-	const { model, systemPrompt } = options ?? {};
+	const { model, systemPrompt, beforeSummarize } = options ?? {};
 	const summarizable = compressible.slice(1);
 	if (summarizable.length === 0) {
 		logger("消息太少，无需总结");
 		return 0;
 	}
+	beforeSummarize?.(summarizable);
 	const count = summarizable.length;
 	summarizable.push({
 		role: "system",
@@ -459,10 +380,11 @@ const compact = Object.assign(async (messages, model, options) => {
 		model
 	}) > 0;
 	if (tokens > context * ratioToSummarize) {
-		const { systemPrompt = "你现在的任务是总结历史消息" } = summarizeOptions ?? {};
+		const { systemPrompt = "你现在的任务是总结历史消息", beforeSummarize } = summarizeOptions ?? {};
 		const [error, count] = await to(summarizeMessages(compressible, {
 			model,
-			systemPrompt
+			systemPrompt,
+			beforeSummarize
 		}));
 		if (!error) result.hasSummarized = count > 0;
 		else result.hasDiscardMessages = discardMessagesUntil(compressible, (compressible) => {
@@ -484,4 +406,4 @@ const compact = Object.assign(async (messages, model, options) => {
 	discardMessagesUntil
 });
 //#endregion
-export { get_time_default as a, runAgent as c, get_weather_default as i, MCPRouter as n, defineModel as o, loadMCPTools as r, defineTool as s, compact as t };
+export { stream as a, runAgent as i, get_weather_default as n, get_time_default as r, compact as t };

package/dist/tools/mcp.d.mts

@@ -0,0 +1,55 @@
+import { p as ToolDefinition } from "../types--dXFoUow.mjs";
+import { Client } from "@modelcontextprotocol/sdk/client";
+//#region src/tools/mcp.d.ts
+type McpServer = {
+  type: "streamable_http" | "sse";
+  url: string;
+  headers?: Record<string, any>;
+  ignoredToolNames?: string[];
+};
+declare class MCPRouter {
+  private entries;
+  /** 注册一个新的MCP客户端 */
+  addClient(name: string, url: string, options?: Omit<McpServer, "type" | "url">): Promise<Client<{
+    method: string;
+    params?: {
+      [x: string]: unknown;
+      _meta?: {
+        [x: string]: unknown;
+        progressToken?: string | number | undefined;
+        "io.modelcontextprotocol/related-task"?: {
+          taskId: string;
+        } | undefined;
+      } | undefined;
+    } | undefined;
+  }, {
+    method: string;
+    params?: {
+      [x: string]: unknown;
+      _meta?: {
+        [x: string]: unknown;
+        progressToken?: string | number | undefined;
+        "io.modelcontextprotocol/related-task"?: {
+          taskId: string;
+        } | undefined;
+      } | undefined;
+    } | undefined;
+  }, {
+    [x: string]: unknown;
+    _meta?: {
+      [x: string]: unknown;
+      progressToken?: string | number | undefined;
+      "io.modelcontextprotocol/related-task"?: {
+        taskId: string;
+      } | undefined;
+    } | undefined;
+  }> | undefined>;
+  /** 返回OpenAI API兼容的tools数组 */
+  getTools(): Promise<ToolDefinition[]>;
+}
+/**
+ * 把传入的MCPServer列表转换成OpenAI API兼容的tools数组
+ */
+declare const loadMCPTools: (mcpServers?: Record<string, McpServer>) => Promise<ToolDefinition[]>;
+//#endregion
+export { MCPRouter, McpServer, loadMCPTools };

package/dist/tools/mcp.mjs

@@ -0,0 +1,2 @@
+import { n as loadMCPTools, t as MCPRouter } from "../mcp-DcntQg-a.mjs";
+export { MCPRouter, loadMCPTools };

package/dist/types--dXFoUow.d.mts

@@ -0,0 +1,144 @@
+//#region src/types.d.ts
+type Model = {
+  baseUrl: string;
+  /** 如果不传，会尝试使用GET {baseUrl}/models获取到的第一个模型 */
+  model?: string;
+  /** 使用本地llama.cpp等服务时可以不填 */
+  apiKey?: string;
+  /**
+   * 模型支持的消息输入类型
+   * @default ["text"]
+   * @remarks 会在每轮对话前校验消息类型，存在不支持的输入时报错
+   */
+  modalities?: Modality[];
+  /**
+   * 模型的最大上下文
+   * @default 131072
+   * @remarks 会在每轮对话结束后检查上下文阈值，自动压缩消息
+   */
+  context?: number;
+  /**
+   * 自定义请求体，会在每轮请求时带上
+   * @example chat_template_kwargs: { enable_thinking: false }
+   */
+  [key: string]: unknown;
+};
+/** 文/图/音/视频，暂未实现文件输入 */
+type Modality = "text" | "image" | "audio" | "video";
+type Message = {
+  role: "system" | "user" | "assistant" | "tool";
+  reasoning?: string | null;
+  content: string | ContentPart[];
+  tool_calls?: ToolCall[];
+  tool_call_id?: string;
+  [key: string]: unknown;
+};
+type ContentPart = TextContent | ImageContent | AudioContent | VideoContent;
+type TextContent = {
+  type: "text";
+  text: string;
+};
+type ImageContent = {
+  type: "image_url";
+  image_url: {
+    url: string;
+  };
+};
+type AudioContent = {
+  type: "input_audio";
+  input_audio: {
+    /** 使用公网可访问的音频链接 */
+    url?: string;
+    /** 使用base64 */
+    data?: string;
+    format: string;
+  };
+};
+type VideoContent = {
+  type: "video_url";
+  video_url: {
+    url: string;
+  };
+};
+type ToolDefinition = {
+  type: "function";
+  function: {
+    name: string;
+    description: string;
+    parameters: {
+      type: "object";
+      properties: Record<string, {
+        type: string;
+        description?: string;
+        /** defineTool()提供的语法糖，等价于在parameters.required.push(当前property key) */
+        required?: boolean;
+      }>;
+      required?: string[];
+    };
+  };
+  /** 工具的实际执行函数 */
+  execute: (...args: any) => any;
+};
+type ToolCall = {
+  id: string;
+  type: "function";
+  function: {
+    name: string;
+    arguments: string;
+  };
+};
+/** stream()对外输出的事件 */
+type LLMEvent = {
+  type: "reasoning_delta";
+  delta: string;
+} | {
+  type: "content_delta";
+  delta: string;
+} | {
+  type: "tool_call";
+  id: string;
+  name: string;
+  args: any;
+} | {
+  type: "done";
+  finishReason: FinishReason;
+  usage?: Usage;
+} | {
+  type: "error";
+  message: string;
+};
+/** stream()内部要处理的事件流 */
+type ChatCompletionsChunk = {
+  id: string;
+  object: "chat.completion.chunk";
+  created: number;
+  model: string;
+  choices: {
+    index: number;
+    delta: Pick<Message, "reasoning" | "content"> & {
+      tool_calls?: ({
+        index: number;
+      } & Partial<ToolCall>)[];
+    };
+    finish_reason: FinishReason;
+  }[];
+  usage?: Usage;
+} | "[DONE]";
+/** 模型可能返回的finish_reason值：停止/工具调用/上下文超限 */
+type FinishReason = "stop" | "tool_calls" | "length" | null;
+type Usage = {
+  prompt_tokens: number;
+  completion_tokens: number;
+  total_tokens: number;
+  /** 每秒生成多少个token */
+  speed?: number;
+  [key: string]: any;
+};
+type AgentEvent = LLMEvent | {
+  type: "tool_result";
+  id: string;
+  name: string;
+  result: string;
+};
+//#endregion
+export { FinishReason as a, Message as c, TextContent as d, ToolCall as f, VideoContent as h, ContentPart as i, Modality as l, Usage as m, AudioContent as n, ImageContent as o, ToolDefinition as p, ChatCompletionsChunk as r, LLMEvent as s, AgentEvent as t, Model as u };

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@nickyzj2023/ai",
-  "version": "1.4.4",
+  "version": "1.6.0",
   "description": "我的“pi”，参考了pi-from-scratch",
   "type": "module",
   "main": "./dist/index.mjs",
@@ -9,6 +9,10 @@
     ".": {
       "types": "./dist/index.d.mts",
       "import": "./dist/index.mjs"
+    },
+    "./mcp": {
+      "types": "./dist/tools/mcp.d.mts",
+      "import": "./dist/tools/mcp.mjs"
     }
   },
   "files": [
@@ -27,7 +31,7 @@
   },
   "devDependencies": {
     "@biomejs/biome": "^2.5.12",
-    "@types/node": "^26.4.1",
+    "@types/node": "^26.5.0",
     "tsdown": "^0.22.14",
     "typescript": "^7.0.2"
   },