@kevisual/ai 0.0.6 → 0.0.7

package/src/provider/utils/token.ts

@@ -0,0 +1,34 @@
+import { encoding_for_model, get_encoding } from 'tiktoken';
+
+
+const MODEL_TO_ENCODING = {
+  'gpt-4': 'cl100k_base',
+  'gpt-4-turbo': 'cl100k_base',
+  'gpt-3.5-turbo': 'cl100k_base',
+  'text-embedding-ada-002': 'cl100k_base',
+  'text-davinci-002': 'p50k_base',
+  'text-davinci-003': 'p50k_base',
+} as const;
+
+export function numTokensFromString(text: string, model: keyof typeof MODEL_TO_ENCODING = 'gpt-3.5-turbo'): number {
+  try {
+    // 对于特定模型使用专门的编码器
+    const encoder = encoding_for_model(model);
+    const tokens = encoder.encode(text);
+    const tokenCount = tokens.length;
+    encoder.free(); // 释放编码器
+    return tokenCount;
+  } catch (error) {
+    try {
+      // 如果模型特定的编码器失败，尝试使用基础编码器
+      const encoder = get_encoding(MODEL_TO_ENCODING[model]);
+      const tokens = encoder.encode(text);
+      const tokenCount = tokens.length;
+      encoder.free(); // 释放编码器
+      return tokenCount;
+    } catch (error) {
+      // 如果编码失败，使用一个粗略的估计：平均每个字符0.25个token
+      return Math.ceil(text.length * 0.25);
+    }
+  }
+}

package/src/test/chunks/01-get.ts

@@ -0,0 +1,65 @@
+import { getChunks } from '../../provider/utils/chunk.ts';
+
+const str = 'Hello world this is a test 你好沙盒 very big';
+
+
+const str2 = `不能直接使用 tiktoken（OpenAI的分词器）来计算 Qwen 模型的 Token 数量，因为两者的分词规则（Tokenization）和词表（Vocabulary）完全不同。
+
+为什么不能混用？
+词表不同
+
+tiktoken 是 OpenAI 为 GPT 系列设计的（如 gpt-3.5-turbo, gpt-4），其词表针对英语和代码优化。
+
+Qwen 使用独立训练的 BPE 词表，对中文、多语言的支持更友好，分词粒度可能不同。
+
+分词结果差异大
+同一段文本，tiktoken 和 Qwen 的分词结果可能完全不同。例如：
+
+OpenAI (tiktoken): "你好" → ['你', '好']（2 Tokens）
+
+Qwen: "你好" → ['你好']（1 Token，如果词表中包含该组合）
+
+性能问题
+即使强制使用 tiktoken 计算 Qwen 的 Token，结果也不准确，可能导致：
+
+输入超出模型上下文限制（因统计偏差）。
+
+API 计费或本地推理时出现意外错误。
+
+正确方法：用 Qwen 的分词器
+通过 Hugging Face transformers 加载 Qwen 的原生分词器：
+
+python
+复制
+from transformers import AutoTokenizer
+
+# 加载 Qwen 的分词器（以 Qwen-7B 为例）
+tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen-7B", trust_remote_code=True)
+
+text = "你好，Qwen模型！"
+tokens = tokenizer.tokenize(text)  # 查看分词结果
+token_count = len(tokenizer.encode(text, add_special_tokens=False))
+
+print("分词结果:", tokens)
+print("Token数量:", token_count)
+常见问题
+为什么需要 trust_remote_code=True？
+Qwen 的分词器是自定义实现的（非 Hugging Face 原生），此参数允许从模型仓库加载运行代码。
+
+其他语言的 Token 计算？
+Qwen 对非英语（如中文、日文）的分词效率较高，但仍需用其原生分词器统计。
+
+与 tiktoken 的速度对比？
+tiktoken 是纯 Python 实现，速度较快；Qwen 的分词器基于 Hugging Face，可能稍慢但对齐模型需求。
+
+总结
+禁止混用：tiktoken ≠ Qwen 分词器。
+
+始终使用模型配套工具：Qwen 需通过 transformers 加载其官方分词器。
+
+中文场景特别注意：Qwen 对中文的分词更高效，直接使用可避免偏差。
+
+如果需要验证分词规则，可通过 tokenizer.vocab 查看词表内容（但注意词表通常较大）。`
+
+const chunks = getChunks(str2);
+console.log(chunks);

package/src/test/encrypt/index.ts

@@ -0,0 +1,9 @@
+import { encryptAES, decryptAES } from '../..//provider/utils/parse-config.ts';
+
+const plainx = process.env.API_KEY;
+const decryptKey = process.env.DECRYPT_KEY;
+const encrypt = encryptAES(plainx, decryptKey);
+console.log('encrypt', encrypt);
+
+const decrypt = decryptAES(encrypt, decryptKey);
+console.log(decrypt);

package/src/test/func-call/curl.sh

@@ -0,0 +1,35 @@
+curl --request POST \
+  --url https://api.siliconflow.cn/v1/chat/completions \
+  --header 'Authorization: Bearer sk-qbiigkzoaamuqxtwlgkugodncebkfbosemadfubjrseobpvx' \
+  --header 'Content-Type: application/json' \
+  --data '{
+  "model": "Qwen/Qwen3-14B",
+  "messages": [
+    {
+      "role": "user",
+      "content": "计算a+b的值"
+    }
+  ],
+  "stream": false,
+  "max_tokens": 512,
+  "stop": null,
+  "temperature": 0.7,
+  "top_p": 0.7,
+  "top_k": 50,
+  "frequency_penalty": 0.5,
+  "n": 1,
+  "response_format": {
+    "type": "text"
+  },
+  "tools": [
+    {
+      "type": "function",
+      "function": {
+        "description": "计算a，b，c算法的值，a=1，b=2，c=3",
+        "name": "compouted",
+        "parameters": {},
+        "strict": false
+      }
+    }
+  ]
+}'

package/src/test/func-call/demo.ts

@@ -0,0 +1,116 @@
+import { SiliconFlow } from '../..//provider/chat-adapter/siliconflow.ts';
+import { Ollama } from '../..//provider/chat-adapter/ollama.ts';
+import dotenv from 'dotenv';
+
+dotenv.config();
+const siliconflow = new SiliconFlow({
+  apiKey: process.env.SILICONFLOW_API_KEY,
+  model: 'Qwen/Qwen3-14B',
+});
+const ollama = new Ollama({
+  model: 'qwen3:32b',
+  apiKey: process.env.OLLAMA_API_KEY,
+  baseURL: process.env.OLLAMA_BASE_URL,
+});
+const main = async () => {
+  const usage = await siliconflow.getUsageInfo();
+  console.log(usage);
+};
+// 1. 定义工具函数
+const availableFunctions: Record<string, (args: any) => Promise<any>> = {
+  get_time: async (args: { location: string }) => {
+    // 模拟API调用
+    console.log('time', args);
+    return {
+      time: '2022-03-22 12:00:00',
+    };
+  },
+  get_location: async (args: { symbol: string }) => {
+    // 模拟API调用
+    console.log('location', args);
+    return {
+      city: 'Beijing',
+    };
+  },
+};
+
+// main();
+const funcCall = async (model = siliconflow) => {
+  const tools = [
+    {
+      type: 'function',
+      function: {
+        name: 'get_time',
+        description: '获取当前时间',
+        parameters: {
+          type: 'object',
+          properties: {
+            place: {
+              type: 'string',
+              description: '位置',
+            },
+          },
+          required: ['place'],
+        },
+      },
+    },
+    {
+      type: 'function',
+      function: {
+        name: 'get_location',
+        description: '获取当前位置',
+        // parameters: {},
+        parameters: {},
+        strict: false,
+      },
+    },
+  ];
+  const messages: any[] = [{ role: 'user', content: '获取当前位置的当前时间' }];
+  const res = await model.chat(messages, {
+    tools: tools as any,
+  });
+  console.log(res.choices[0]);
+  const assistantMessage = res.choices[0].message;
+  const finish_reason = res.choices[0].finish_reason;
+  messages.push(assistantMessage);
+  let toolCalls = assistantMessage.tool_calls;
+  console.log("toolCalls", JSON.stringify(toolCalls));
+  let maxRetries = 3;
+  while (toolCalls && toolCalls.length > 0) {
+    // 处理每个函数调用
+    for (const toolCall of toolCalls) {
+      const functionName = toolCall.function.name;
+      const functionArgs = JSON.parse(toolCall.function.arguments);
+      // 调用本地函数
+      const functionResponse = await availableFunctions[functionName](functionArgs);
+      // 将结果添加到消息历史
+      messages.push({
+        role: 'tool',
+        name: functionName,
+        content: JSON.stringify(functionResponse),
+        tool_call_id: toolCall.id,
+      });
+    }
+
+    // 第二次调用 - 将函数结果发送给模型获取最终回复
+    const secondResponse = await model.chat(messages, {
+      tools: tools as any,
+    });
+
+    const finalMessage = secondResponse.choices[0].message;
+    messages.push(finalMessage);
+    const _toolCalls = finalMessage.tool_calls;
+    console.log("toolCalls", JSON.stringify(toolCalls) ,finalMessage.role);
+    toolCalls = _toolCalls ? _toolCalls : [];
+    maxRetries--;
+    if (maxRetries <= 0) {
+      break;
+    }
+    
+    console.log('tool calls', toolCalls);
+  }
+
+  console.log(messages);
+};
+
+funcCall(ollama as any);

package/src/test/model-scope/index.ts

@@ -0,0 +1,26 @@
+import { ModelScope } from '../..//provider/chat-adapter/model-scope.ts';
+import { log } from '../..//logger/index.ts';
+import util from 'util';
+import { config } from 'dotenv';
+config();
+
+const chat = new ModelScope({
+  apiKey: process.env.MODEL_SCOPE_API_KEY,
+  model: 'Qwen/Qwen2.5-Coder-32B-Instruct',
+});
+
+// chat.chat([{ role: 'user', content: 'Hello, world! 1 + 1 equals ?' }]);
+const chatMessage = [{ role: 'user', content: 'Hello, world! 1 + 1 equals ?' }];
+
+const main = async () => {
+  const res = await chat.test();
+  log.info('test', res);
+};
+
+main();
+const mainChat = async () => {
+  const res = await chat.chat(chatMessage as any);
+  log.info('chat', res);
+};
+
+// mainChat();

package/src/test/ollama-knowledge.ts

@@ -0,0 +1,37 @@
+import { Knowledge } from '../../../../src/provider/knowledge/knowledge.ts';
+import fs from 'fs';
+import dotenv from 'dotenv';
+
+dotenv.config();
+const knowledge = new Knowledge({
+  embeddingModel: 'bge-m3:latest',
+  baseURL: 'https://ollama.xiongxiao.me/v1',
+  model: 'qwq:latest',
+  apiKey: process.env.OLLAMA_API_KEY,
+});
+
+const main = async () => {
+  const res = await knowledge.generateEmbeddingCore('Hello world this is a test 你好沙盒 very big');
+  fs.writeFileSync('docs/embedding.json', JSON.stringify(res, null, 2));
+  console.log(res);
+};
+
+main();
+
+const main2 = async () => {
+  const text1 = 'Hello, world! this is a test';
+  const text2 = 'Hello, world! this is a test 2';
+  const text3 = 'Hello, world! this is a test 3';
+  const text4 = 'Hello, world! this is a test 4';
+  const text5 = 'Hello, world! this is a test 5';
+  const text6 = 'Hello, world! this is a test 6';
+  const text7 = 'Hello, world! this is a test 7';
+  const text8 = 'Hello, world! this is a test 8';
+  const text9 = 'Hello, world! this is a test 9';
+  const text10 = 'Hello, world! this is a test 10';
+  const res = await knowledge.generateEmbeddingCore([text1, text2, text3, text4, text5, text6, text7, text8, text9, text10]);
+  fs.writeFileSync('docs/embedding2.json', JSON.stringify(res, null, 2));
+  console.log(res);
+};
+
+// main2();

package/src/test/ollama.ts

@@ -0,0 +1,86 @@
+import { Ollama } from '../../../../src/provider/chat-adapter/ollama.ts';
+import util from 'util';
+const chat = new Ollama({
+  baseURL: 'https://ollama.xiongxiao.me/v1',
+  apiKey: 'xiongxiao2233',
+  model: 'qwq:latest',
+});
+
+// chat.chat([{ role: 'user', content: 'Hello, world!' }]);
+
+const main = async () => {
+  const res = await chat.test();
+  console.log(util.inspect(res, { depth: null, colors: true }));
+};
+
+// main();
+
+const getJson = async () => {
+  const res = await chat.chat(
+    [
+      { role: 'system', content: '把发送的数据，返回给我对应的json，只处理完发送的数据。如果发送了多个，给我一个数组' },
+      // { role: 'user', content: '{"name":"John","age":30}' },
+      { role: 'user', content: 'name: 张三' },
+      { role: 'user', content: 'name: 李四, age: 18' },
+    ],
+    {
+      response_format: {
+        type: 'json_schema',
+        json_schema: {
+          name: 'user',
+          description: '用户信息',
+          schema: {
+            type: 'object',
+            // properties: {
+            //   name: { type: 'string' },
+            //   // age: { type: 'number' },
+            // },
+            // // required: ['name', 'age'],
+            // required: ['name'],
+            properties: {
+              name: { type: 'string' },
+              age: { type: 'number' },
+            },
+            required: ['name', 'age'],
+          },
+        },
+      },
+      n: 10,
+    },
+  );
+  console.log(util.inspect(res, { depth: null, colors: true }));
+};
+
+// getJson();
+
+const createChat1 = async () => {
+  const res = await chat.chat(
+    [
+      { role: 'user', content: 'a=1, b=2, c=3' },
+      { role: 'user', content: 'a+b+c=?' },
+      { role: 'assistant', content: '给定的值为 \\( a = 1 \\), \\( b = 2 \\), \\( c = 3 \\)。\n' + '\n' + '因此，\\( a + b + c = 1 + 2 + 3 = 6 \\)。' },
+      { role: 'user', content: 'a+b+c+4=?' },
+    ],
+    {
+      model: 'qwen2.5:7b',
+    },
+  );
+  console.log(util.inspect(res, { depth: null, colors: true }));
+};
+
+//  createChat1();
+
+const getTags = async () => {
+  const res = await chat.listModels();
+  console.log(util.inspect(res, { depth: null, colors: true }));
+};
+
+// getTags();
+
+const getRunModels = async () => {
+  const res = await chat.listRunModels();
+  console.log('current', new Date().toISOString());
+  console.log(util.inspect(res, { depth: null, colors: true }));
+};
+
+// getRunModels();

package/src/test/provider/index.ts

@@ -0,0 +1,7 @@
+import { ProviderManager } from '../..//provider/index.ts';
+import { config } from 'dotenv';
+config();
+const providerConfig = { provider: 'ModelScope', model: 'Qwen/Qwen2.5-Coder-32B-Instruct', apiKey: process.env.MODEL_SCOPE_API_KEY };
+const provider = await ProviderManager.createProvider(providerConfig);
+const result = await provider.chat([{ role: 'user', content: '你好' }]);
+console.log(result);

package/src/test/siliconflow/common.ts

@@ -0,0 +1,15 @@
+import { SiliconFlow } from '../../../src/provider/chat-adapter/siliconflow.ts';
+import { SiliconFlowKnowledge } from '../../provider/knowledge-adapter/siliconflow.ts';
+import dotenv from 'dotenv';
+
+dotenv.config();
+export const siliconflow = new SiliconFlow({
+  apiKey: process.env.SILICONFLOW_API_KEY,
+  model: 'Qwen/Qwen2-7B-Instruct',
+});
+
+export const knowledge = new SiliconFlowKnowledge({
+  apiKey: process.env.SILICONFLOW_API_KEY,
+  model: 'Qwen/Qwen2-7B-Instruct',
+  embeddingModel: 'Pro/BAAI/bge-m3',
+});

package/src/test/siliconflow/get.ts

@@ -0,0 +1,22 @@
+import { SiliconFlow } from '../..//provider/chat-adapter/siliconflow.ts';
+import dotenv from 'dotenv';
+
+dotenv.config();
+const siliconflow = new SiliconFlow({
+  apiKey: process.env.SILICONFLOW_API_KEY,
+  model: 'Qwen/Qwen2-7B-Instruct',
+});
+
+
+const main = async () => {
+  const usage = await siliconflow.getUsageInfo();
+  console.log(usage);
+};
+
+main();
+const mainChat = async () => {
+  const res = await siliconflow.chat([{ role: 'user', content: 'Hello, world! 1 + 1 equals ?' }]);
+  console.log(res);
+};
+
+// mainChat();

package/src/test/siliconflow/knowledge/create.ts

@@ -0,0 +1,18 @@
+import { knowledge } from '../common.ts';
+import fs from 'node:fs';
+import path from 'node:path';
+import { fileURLToPath } from 'url';
+import { dirname } from 'path';
+
+const __filename = fileURLToPath(import.meta.url);
+const __dirname = dirname(__filename);
+// 包含： 9184 个汉字 953 个标点(全角) 2493 个字母 52 个数字
+const content = fs.readFileSync(path.join(__dirname, 'qwen.md'), 'utf-8');
+const text = 'Hello, world';
+const main = async () => {
+  const res = await knowledge.generateEmbeddingCore([content, content]);
+  console.log(res);
+  // 8000 tokens 大概1w个字 2万个字符
+  console.log('speak', knowledge.getChatUsage());
+};
+main();