content/ChatGPT使用指南——相似匹配.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "id": "a4980428",
   "metadata": {
    "toc": true
   },
   "source": [
    "<h1>Table of Contents<span class=\"tocSkip\"></span></h1>\n",
    "<div class=\"toc\"><ul class=\"toc-item\"><li><span><a href=\"#ChatGPT使用指南——相似匹配\" data-toc-modified-id=\"ChatGPT使用指南——相似匹配-1\">ChatGPT使用指南——相似匹配</a></span><ul class=\"toc-item\"><li><span><a href=\"#1.-何为Embedding\" data-toc-modified-id=\"1.-何为Embedding-1.1\">1. 何为Embedding</a></span></li><li><span><a href=\"#2.-相关API\" data-toc-modified-id=\"2.-相关API-1.2\">2. 相关API</a></span><ul class=\"toc-item\"><li><span><a href=\"#2.1-LMAS-Embedding-API\" data-toc-modified-id=\"2.1-LMAS-Embedding-API-1.2.1\">2.1 LMAS Embedding API</a></span></li><li><span><a href=\"#2.2-ChatGPT-Style\" data-toc-modified-id=\"2.2-ChatGPT-Style-1.2.2\">2.2 ChatGPT Style</a></span></li></ul></li><li><span><a href=\"#3.-Embedding应用\" data-toc-modified-id=\"3.-Embedding应用-1.3\">3. Embedding应用</a></span><ul class=\"toc-item\"><li><span><a href=\"#3.1-QA\" data-toc-modified-id=\"3.1-QA-1.3.1\">3.1 QA</a></span></li><li><span><a href=\"#3.2-聚类\" data-toc-modified-id=\"3.2-聚类-1.3.2\">3.2 聚类</a></span></li><li><span><a href=\"#3.3-推荐\" data-toc-modified-id=\"3.3-推荐-1.3.3\">3.3 推荐</a></span></li></ul></li><li><span><a href=\"#相关文献\" data-toc-modified-id=\"相关文献-1.4\">相关文献</a></span></li></ul></li></ul></div>"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "2ca90e2d",
   "metadata": {},
   "source": [
    "# ChatGPT使用指南——相似匹配"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "c08b87df",
   "metadata": {},
   "source": [
    "## 1. 何为Embedding"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "0403a6f6",
   "metadata": {},
   "source": [
    "&emsp;&emsp;对于自然语言，因为它的输入是一段文本，在中文里就是一个一个字，或一个一个词，行业内把这个字或词叫Token。如果要使用模型，拿到一段文本的第一件事就是把它Token化，当然，可以按字、也可以按词，或按你想要的其他方式，比如每两个字一组（Bi-Gram）。举个例子：\n",
    "\n",
    "- 给定文本：我们相信AI可以让世界变得更美好。\n",
    "- 按字Token化：我/们/相/信/A/I/可/以/让/世/界/变/得/更/美/好/。\n",
    "- 按词Token化：我们/相信/AI/可以/让/世界/变得/更/美好/。\n",
    "- 按Bi-Gram Token化：我们/们相/相信/信A/AI/I可/可以/以让/让世/世界/界变/变得/得更/更美/美好/好。"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "cb576e6d",
   "metadata": {},
   "source": [
    "&emsp;&emsp;那自然就有一个新的问题：我们应该怎么选择Token化方式？其实每种不同的方法都有自己的优点和不足，在大模型之前，按词的方式比较常见。但在有了大模型之后，基本都是按字来了，不用再纠结这个点了。\n",
    "\n",
    "&emsp;&emsp;Token化后，第二件事就是要怎么表示这些Token，我们知道计算机只能处理数字，所以要想办法把这些Token给变成计算机「认识」的数字才行。读者不妨思考一下如果要你来做这件事会怎么做。"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "8eeab223",
   "metadata": {},
   "source": [
    "&emsp;&emsp;其实很简单很直观，把所有字作为一个字典，序号就代表它自己。我们还是以上面的句子为例，假设词表就包含上面那些字，那么词表就可以用一个txt文件存储，内容如下：\n",
    "\n",
    "```text\n",
    "我\n",
    "们\n",
    "相\n",
    "信\n",
    "A\n",
    "I\n",
    "可\n",
    "以  \n",
    "让  \n",
    "世  \n",
    "界  \n",
    "变  \n",
    "得  \n",
    "更  \n",
    "美  \n",
    "好  \n",
    "。 \n",
    "```"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "6ba7ad32",
   "metadata": {},
   "source": [
    "&emsp;&emsp;一行一个字，每个字作为一个Token，此时，0=我，1=们，……，以此类推。拿中文来说，这个词表可能只要几千行，即使包含各种特殊符号、生僻字，也就2万个多点，我们假设词表大小为N。"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "96d6352c",
   "metadata": {},
   "source": [
    "&emsp;&emsp;接下来我们考虑如何用这些数字来表示一段文本。最简单的方法就是用它的ID直接串起来，这样也不是不行，但这种表示方法的特征是一维的，也就是说只能表示一个特征。这种方法不太符合实际情况，效果也不理想。所以，研究人员就想到另一种表示方法：One-Hot编码。其实，将文本变为数字表示的过程本质上就是一种编码过程。One-Hot的意思是，对每一个字都有N（词表大小）个特征，除了该字的ID位置值为1，其余都为0。我们依然用上面的例子来说明，把整个词表表示为下面的形式：\n",
    "\n",
    "\n",
    "> 我 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0  \n",
    "们 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0  \n",
    "相 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0  \n",
    "信 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0  \n",
    "……下面省略"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "4f9c02a1",
   "metadata": {},
   "source": [
    "&emsp;&emsp;此时，对每一个Token（字），它的表示就变成了一个一维向量，比如「我」：`[1,0,...0]`，这个向量的长度就是词表的大小N，它被称为「我」这个字的One-Hot表示。"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "4846a023",
   "metadata": {},
   "source": [
    "&emsp;&emsp;对于一段文本，我们一般会将每个Token的表示结合起来，结合方式可以采用求和或平均。这样，对于任意长度的任意文本，我们都能将其表示为固定大小的向量，非常方便进行各种矩阵或张量（三维以上的数组）计算，这对深度学习至关重要。\n",
    "\n",
    "&emsp;&emsp;举个例子，比如有这么一句话：让世界更美好。现在我们使用刚刚的方法将其表示为一个向量，采用平均的方式。"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "5f645fd0",
   "metadata": {},
   "source": [
    "&emsp;&emsp;首先，列出每个字的向量：\n",
    "\n",
    "> 让 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0  \n",
    "世 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0  \n",
    "界 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0  \n",
    "更 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0  \n",
    "美 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0  \n",
    "好 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0  \n",
    "。 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1  "
   ]
  },
  {
   "cell_type": "markdown",
   "id": "ddad8a7a",
   "metadata": {},
   "source": [
    "&emsp;&emsp;然后针对每一列取平均，结果为：`0 0 0 0 0 0 0 1/7 1/7 1/7 0 0 1/7 1/7 1/7 1/7`。不难发现，对任两句话，只要其中包含的字不完全一样，最终得到的向量表示也不会完全一样。注①"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "fc51f689",
   "metadata": {},
   "source": [
    "&emsp;&emsp;当然，在实际使用时，往往不会这么简单的用1/0来表示，因为每个字在句子中的作用是不一样的，所以一般会给不同的Token赋予不同的权重。最常见的方法是使用在句子中出现的频率，那些高频的（但不是「的」「更」这样的虚词）被认为是重要的。更多可以参考【相关文献1】。"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "d0be8361",
   "metadata": {},
   "source": [
    "&emsp;&emsp;这种方法不错，在深度学习之前很长一段时间里都是这样的，不过它有两个很大的问题：\n",
    "\n",
    "1. 数据维度太高：太高的维度会导致向量在空间中聚集在一个非常狭窄的角落，模型难以训练。\n",
    "2. 数据稀疏，向量之间缺乏语义上的交互（语义鸿沟）：比如「我爱吃苹果」和「我爱用苹果」，前者是水果，后者是手机，怎么判断出来的呢？根据上下文。但由于现在这种表示方式，导致上下文之间是孤立的，所以模型学不到这个知识点。还有类似「我喜欢你」和「你喜欢我」这样会得到同样的表示，但其实是不同的意思。注①\n",
    "\n",
    "\n",
    "更多可以参考【相关文献 1】。"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "87446397",
   "metadata": {},
   "source": [
    "&emsp;&emsp;终于轮到我们的主角Embedding登场了，它的主要思想是这样的：\n",
    "\n",
    "- 把特征固定在某一个维度D，比如256、300、768等等，这个不重要，总之不再是词表那么大的数字。这就避免了维度过高的问题。\n",
    "- 利用自然语言文本的上下文关系学习一个稠密表示。也就是说，每个Token的表示不再是预先算好的了，而是在过程中学习到的，元素也不再是很多个0，而是每个位置都有一个小数，这D个小数构成了一个Token表示。至于D个特征到底是什么，不知道，也不重要。我们只需要知道这D个小数就表示这个Token。\n",
    "\n",
    "&emsp;&emsp;还是继续以前面的例子来说明，这时候词表的表示变成下面这样了：\n",
    "\n",
    "\n",
    "> 我 0.xxx0, 0.yyy0, 0.zzz0, ... D个小数  \n",
    "们 0.xxx1, 0.yyy1, 0.zzz1, ... D个小数  \n",
    "相 0.xxx2, 0.yyy2, 0.zzz2, ... D个小数  \n",
    "信 0.xxx3, 0.yyy3, 0.zzz3, ... D个小数  \n",
    "……下面省略  \n",
    "\n",
    "&emsp;&emsp;这些小数怎么来的？简单，随机来的。就像下面这样："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "id": "4ef7ffe6",
   "metadata": {},
   "outputs": [],
   "source": [
    "import numpy as np"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "id": "62817ba2",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(16, 256)"
      ]
     },
     "execution_count": 16,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "rng = np.random.default_rng(42)\n",
    "# 词表大小N=16，维度D=256\n",
    "table = rng.uniform(size=(16, 256))\n",
    "table.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "id": "0fd77b1e",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([[0.77395605, 0.43887844, 0.85859792, ..., 0.24783956, 0.23666236,\n",
       "        0.74601428],\n",
       "       [0.81656876, 0.10527808, 0.06655886, ..., 0.11585672, 0.07205915,\n",
       "        0.84199321],\n",
       "       [0.05556792, 0.28061144, 0.33413004, ..., 0.00925978, 0.18832197,\n",
       "        0.03128351],\n",
       "       ...,\n",
       "       [0.50647331, 0.22303613, 0.94414565, ..., 0.79202324, 0.40169878,\n",
       "        0.72247782],\n",
       "       [0.9151384 , 0.80071297, 0.39044651, ..., 0.03994193, 0.79502741,\n",
       "        0.28297954],\n",
       "       [0.68255979, 0.64272531, 0.65262805, ..., 0.18645529, 0.21927175,\n",
       "        0.32320729]])"
      ]
     },
     "execution_count": 17,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "table"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "13b8f0cf",
   "metadata": {},
   "source": [
    "&emsp;&emsp;在模型训练过程中，会根据不同的上下文不断地更新这个参数，最后模型训练完后得到的这个矩阵就是Token的表示。我们完全可以把它当成一个黑盒子，输入一个X，根据标签Y不断更新参数，最终就得到一组参数，这些参数的名字就叫「模型」。\n",
    "\n",
    "&emsp;&emsp;这种表示方法在深度学习早期（2013-2015年左右）比较流行，不过由于这个矩阵训练好后就固定不变了，这在有些时候就不合适。比如「你好坏」这句话在不同的情况下可能完全是不同的意思。"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "8f65406a",
   "metadata": {},
   "source": [
    "&emsp;&emsp;我们知道，句子才是语义的最小单位，因此相比Token，我们其实更加关注和需要句子的表示，我们期望可以根据不同上下文动态地获得句子表示。这中间当然经历了比较多的探索，一直到如今的大模型时代，对模型输入任意一句话，它都能给我们返回一个非常不错的表示，而且依然是固定长度的向量。\n",
    "\n",
    "&emsp;&emsp;如果对这方面感兴趣，可以进一步阅读【相关文献2】和【相关文献3】。"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "89325cdf",
   "metadata": {},
   "source": [
    "&emsp;&emsp;我们总结一下，Embedding本质就是一组稠密向量，用来表示一段文本（可以是字、词、句、段等），获取到这个表示后，我们就可以进一步做一些任务。大家不妨先思考一下，当给定任意句子并获得到它的固定长度的语义表示时，我们可以干什么？我们在下一节将先介绍一下OpenAI提供的接口，以及一些后面任务可能用到的概念。"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "9bfde057",
   "metadata": {},
   "source": [
    "## 2. 相关API"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "597c3fc4",
   "metadata": {},
   "source": [
    "### 2.1 LMAS Embedding API"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "id": "e6b93864",
   "metadata": {},
   "outputs": [],
   "source": [
    "import os\n",
    "OPENAI_API_KEY = os.environ.get(\"OPENAI_API_KEY\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "id": "1f093735",
   "metadata": {},
   "outputs": [],
   "source": [
    "import openai\n",
    "# OPENAI_API_KEY = \"填入专属的API key\"\n",
    "openai.api_key = OPENAI_API_KEY"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 31,
   "id": "530f4286",
   "metadata": {},
   "outputs": [],
   "source": [
    "text = \"我喜欢你\"\n",
    "model = \"text-embedding-ada-002\""
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 32,
   "id": "1b4f9c80",
   "metadata": {
    "scrolled": true
   },
   "outputs": [],
   "source": [
    "emb_req = openai.Embedding.create(input=[text], model=model)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 33,
   "id": "1da5ef4c",
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(1536, list)"
      ]
     },
     "execution_count": 33,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "emb = emb_req.data[0].embedding\n",
    "len(emb), type(emb)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "fdeb078c",
   "metadata": {},
   "source": [
    "&emsp;&emsp;与Embedding息息相关的一个概念是「相似度」，准确来说是「语义相似度」。在自然语言处理领域，我们一般使用**cosine相似度**作为语义相似度的度量，评估两个向量在语义空间上的分布情况。\n",
    "\n",
    "&emsp;&emsp;具体来说就是下面这个式子：\n",
    "\n",
    "$$\n",
    "\\text{cosine}(v,w) = \\frac {v·w}{|v||w|} = \\frac {\\displaystyle \\sum_{i=1}^N v_iw_i} {\\displaystyle \\sqrt{\\sum_{i=1}^N v_i^2} \\sqrt{\\sum_{i=1}^N w_i^2}}\n",
    "$$"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "454fb8b2",
   "metadata": {},
   "source": [
    "&emsp;&emsp;我们举个例子："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 27,
   "id": "f67ed38f",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "0.9925833339709301"
      ]
     },
     "execution_count": 27,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "import numpy as np\n",
    "a = [0.1, 0.2, 0.3]\n",
    "b = [0.2, 0.3, 0.4]\n",
    "cosine_ab = (0.1*0.2+0.2*0.3+0.3*0.4)/(np.sqrt(0.1**2+0.2**2+0.3**2) * np.sqrt(0.2**2+0.3**2+0.4**2))\n",
    "cosine_ab"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "7e7c7fb2",
   "metadata": {},
   "source": [
    "&emsp;&emsp;OpenAI官方提供了一个集成接口，使用起来更加简单（但其实你也可以自己写一个）："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "id": "adde7b98",
   "metadata": {},
   "outputs": [],
   "source": [
    "from openai.embeddings_utils import get_embedding, cosine_similarity"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 36,
   "id": "1815d2ca",
   "metadata": {},
   "outputs": [],
   "source": [
    "# 注意它默认的模型是text-similarity-davinci-001，我们也可以换成text-embedding-ada-002\n",
    "text1 = \"我喜欢你\"\n",
    "text2 = \"我钟意你\"\n",
    "text3 = \"我不喜欢你\"\n",
    "emb1 = get_embedding(text1)\n",
    "emb2 = get_embedding(text2)\n",
    "emb3 = get_embedding(text3)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 37,
   "id": "aeb63bcf",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(12288, list)"
      ]
     },
     "execution_count": 37,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "len(emb1), type(emb1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 38,
   "id": "bb4dbbfc",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "0.9246855139297101"
      ]
     },
     "execution_count": 38,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "cosine_similarity(emb1, emb2)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 39,
   "id": "c76e7bcf",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "0.8578009661644189"
      ]
     },
     "execution_count": 39,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "cosine_similarity(emb1, emb3)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "id": "92801b94",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "0.8205299527695261"
      ]
     },
     "execution_count": 20,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "cosine_similarity(emb2, emb3)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "id": "40bf0057",
   "metadata": {},
   "outputs": [],
   "source": [
    "text1 = \"我喜欢你\"\n",
    "text2 = \"我钟意你\"\n",
    "text3 = \"我不喜欢你\"\n",
    "emb1 = get_embedding(text1, \"text-embedding-ada-002\")\n",
    "emb2 = get_embedding(text2, \"text-embedding-ada-002\")\n",
    "emb3 = get_embedding(text3, \"text-embedding-ada-002\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "id": "f02ae887",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "0.8931105629213952"
      ]
     },
     "execution_count": 22,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "cosine_similarity(emb1, emb2)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "id": "9481b45a",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "0.9262074073566393"
      ]
     },
     "execution_count": 23,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "cosine_similarity(emb1, emb3)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "id": "200317d4",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "0.845821877417193"
      ]
     },
     "execution_count": 24,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "cosine_similarity(emb2, emb3)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "f0c8dcd9",
   "metadata": {},
   "source": [
    "&emsp;&emsp;text-embedding-ada-002模型在这个例子上表现不太令人满意。更多模型可以在这里查看：[New and improved embedding model](https://openai.com/blog/new-and-improved-embedding-model)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "6991b67b",
   "metadata": {},
   "source": [
    "### 2.2 ChatGPT Style"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "b5dd5c74",
   "metadata": {},
   "source": [
    "&emsp;&emsp;接下来我们用万能的ChatGPT尝试一下，注意它不会给你返回Embedding，它是尝试直接告诉你答案！"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 29,
   "id": "86f116d3",
   "metadata": {},
   "outputs": [],
   "source": [
    "content = \"请告诉我下面三句话的相似程度：\\n1. 我喜欢你。\\n2. 我钟意你。\\n3.我不喜欢你。\\n\""
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 30,
   "id": "29ad8225",
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "'\\n\\n1和2相似，都表达了对某人的好感或喜欢之情。而3则与前两句截然相反，表示对某人的反感或不喜欢。'"
      ]
     },
     "execution_count": 30,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "response = openai.ChatCompletion.create(\n",
    "    model=\"gpt-3.5-turbo\", \n",
    "    messages=[{\"role\": \"user\", \"content\": content}]\n",
    ")\n",
    "\n",
    "response.get(\"choices\")[0].get(\"message\").get(\"content\")"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "4ed55c55",
   "metadata": {},
   "source": [
    "牛逼，不过这个格式不太好，我们调整一下："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 31,
   "id": "4b4ec156",
   "metadata": {},
   "outputs": [],
   "source": [
    "content += '第一句话用a表示，第二句话用b表示，第三句话用c表示，请以json格式输出两两相似度，类似下面这样：\\n{\"ab\": a和b的相似度}'"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 32,
   "id": "b7bdde56",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "'\\n\\n{\"ab\": 0.8, \"ac\": -1, \"bc\": 0.7}\\n\\n解释：a和b的相似度为0.8，因为两句话表达了相同的情感；a和c的相似度为-1，因为两句话表达了相反的情感；b和c的相似度为0.7，因为两句话都是表达情感，但一个是积极情感，一个是消极情感，相似度略低。'"
      ]
     },
     "execution_count": 32,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "response = openai.ChatCompletion.create(\n",
    "    model=\"gpt-3.5-turbo\", \n",
    "    messages=[{\"role\": \"user\", \"content\": content}]\n",
    ")\n",
    "\n",
    "response.get(\"choices\")[0].get(\"message\").get(\"content\")"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "612b96fa",
   "metadata": {},
   "source": [
    "牛逼++！"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "f2f3b058",
   "metadata": {},
   "source": [
    "## 3. Embedding应用"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "d15f87f9",
   "metadata": {},
   "source": [
    "&emsp;&emsp;有读者可能会疑惑，既然有了这么牛逼Plus的ChatGPT了，为什么还要介绍Embedding这种看起来好像是「低级货」的技术呢？这里主要有两个原因：\n",
    "\n",
    "1. 有些问题使用Embedding解决（或其他非ChatGPT的方式）会更加合理。通俗来说就是「杀鸡焉用牛刀」。\n",
    "2. ChatGPT性能方面不是特别友好，毕竟是一个Token一个Token吐出来的。"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "776fe3df",
   "metadata": {},
   "source": [
    "&emsp;&emsp;关于第一点，我们要额外多说几句。选择技术方案就跟选对象一样，合适最重要。只要你的问题（需求）没变，能解决的技术就是好技术。比如你的任务就是一个二分类，明明一个很简单的模型就能解决，就没必要非得上个很复杂的。除非ChatGPT这样的LLM已经大众到一定阶段（任何人都能够非常流畅、自由地使用），我们从统一性角度考虑。\n",
    "\n",
    "&emsp;&emsp;言归正传，使用Embedding的应用大多跟语义相关，我们这里介绍与此相关的几个经典任务以及衍生的应用。"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "79f3bd15",
   "metadata": {},
   "source": [
    "### 3.1 QA"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "5fbdf2c9",
   "metadata": {},
   "source": [
    "&emsp;&emsp;QA是问答的意思，Q表示Question，A表示Answer，QA是NLP非常基础和常用的任务。简单来说，就是当用户提出一个问题时，我们能从已有的问题库中找到一个最相似的，并把它的答案返回给用户。这里有两个关键点：\n",
    "\n",
    "1. 事先需要有一个QA库。\n",
    "2. 用户提问时，系统要能够在QA库中找到一个最相似的。"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "8a534745",
   "metadata": {},
   "source": [
    "&emsp;&emsp;ChatGPT（或生成方式）做这类任务相对有点麻烦，尤其是当：\n",
    "\n",
    "- QA库非常庞大时\n",
    "- 给用户的答案是固定的，不允许自由发挥时\n",
    "\n",
    "&emsp;&emsp;生成方式做起来是事倍功半。但是Embedding确实天然的非常适合，因为该任务的核心就是在一堆文本中找出给定文本最相似的。简单来说，其实就是个相似度计算问题。"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "211cc967",
   "metadata": {},
   "source": [
    "&emsp;&emsp;我们使用Kaggle提供的Quora数据集：[FAQ Kaggle dataset! | Data Science and Machine Learning](https://www.kaggle.com/general/183270)，先把它给读进来。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 41,
   "id": "85786f71",
   "metadata": {},
   "outputs": [],
   "source": [
    "import pandas as pd"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 42,
   "id": "775a8f48",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(1166, 4)"
      ]
     },
     "execution_count": 42,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "df = pd.read_csv(\"dataset/Kaggle related questions on Qoura - Questions.csv\")\n",
    "df.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 43,
   "id": "b7ed7f42",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>Questions</th>\n",
       "      <th>Followers</th>\n",
       "      <th>Answered</th>\n",
       "      <th>Link</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>How do I start participating in Kaggle competi...</td>\n",
       "      <td>1200</td>\n",
       "      <td>1</td>\n",
       "      <td>/How-do-I-start-participating-in-Kaggle-compet...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>Is Kaggle dead?</td>\n",
       "      <td>181</td>\n",
       "      <td>1</td>\n",
       "      <td>/Is-Kaggle-dead</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>How should a beginner get started on Kaggle?</td>\n",
       "      <td>388</td>\n",
       "      <td>1</td>\n",
       "      <td>/How-should-a-beginner-get-started-on-Kaggle</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3</th>\n",
       "      <td>What are some alternatives to Kaggle?</td>\n",
       "      <td>201</td>\n",
       "      <td>1</td>\n",
       "      <td>/What-are-some-alternatives-to-Kaggle</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>4</th>\n",
       "      <td>What Kaggle competitions should a beginner sta...</td>\n",
       "      <td>273</td>\n",
       "      <td>1</td>\n",
       "      <td>/What-Kaggle-competitions-should-a-beginner-st...</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "                                           Questions  Followers  Answered  \\\n",
       "0  How do I start participating in Kaggle competi...       1200         1   \n",
       "1                                    Is Kaggle dead?        181         1   \n",
       "2       How should a beginner get started on Kaggle?        388         1   \n",
       "3              What are some alternatives to Kaggle?        201         1   \n",
       "4  What Kaggle competitions should a beginner sta...        273         1   \n",
       "\n",
       "                                                Link  \n",
       "0  /How-do-I-start-participating-in-Kaggle-compet...  \n",
       "1                                    /Is-Kaggle-dead  \n",
       "2       /How-should-a-beginner-get-started-on-Kaggle  \n",
       "3              /What-are-some-alternatives-to-Kaggle  \n",
       "4  /What-Kaggle-competitions-should-a-beginner-st...  "
      ]
     },
     "execution_count": 43,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "df.head()"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "8317a26b",
   "metadata": {},
   "source": [
    "&emsp;&emsp;这里，我们就把Link当做答案构造数据对。基本的流程如下：\n",
    "\n",
    "- 对每个Question计算Embedding\n",
    "- 存储Embedding，同时存储每个Question对应的答案\n",
    "- 从存储的地方检索最相似的Question"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "1bcd7255",
   "metadata": {},
   "source": [
    "&emsp;&emsp;第一步我们将借助OpenAI的Embedding接口，但是后两步得看实际情况了。如果Question的数量比较少，比如只有几万条甚至几千条，那我们可以把计算好的Embedding直接存储成文件，每次服务启动时直接加载到内存或缓存里就好了。使用时，挨个计算输入问题和存储的所有问题的相似度，然后给出最相似的问题的答案。"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "53d61b30",
   "metadata": {},
   "source": [
    "&emsp;&emsp;为了快速演示，我们只取前5个句子为例："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "id": "8689f565",
   "metadata": {},
   "outputs": [],
   "source": [
    "from openai.embeddings_utils import get_embedding, cosine_similarity\n",
    "import openai\n",
    "import numpy as np\n",
    "OPENAI_API_KEY = \"填入专属的API key\"\n",
    "openai.api_key = OPENAI_API_KEY"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "id": "f76ef71d",
   "metadata": {},
   "outputs": [],
   "source": [
    "vec_base = []\n",
    "for v in df.head().itertuples():\n",
    "    emb = get_embedding(v.Questions)\n",
    "    im = {\n",
    "        \"question\": v.Questions,\n",
    "        \"embedding\": emb,\n",
    "        \"answer\": v.Link\n",
    "    }\n",
    "    vec_base.append(im)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "4d6d260b",
   "metadata": {},
   "source": [
    "&emsp;&emsp;然后给定输入，比如：\"is kaggle alive?\"，我们先获取它的Embedding，然后逐个遍历`vec_base`计算相似度，并取最高的作为响应。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 45,
   "id": "a9144856",
   "metadata": {},
   "outputs": [],
   "source": [
    "query = \"is kaggle alive?\"\n",
    "q_emb = get_embedding(query)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "id": "b7570046",
   "metadata": {},
   "outputs": [],
   "source": [
    "sims = [cosine_similarity(q_emb, v[\"embedding\"]) for v in vec_base]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "id": "8e6dcb55",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "[0.665769204766594,\n",
       " 0.8711775410642538,\n",
       " 0.7489853201153621,\n",
       " 0.7384357684745508,\n",
       " 0.7287129153982224]"
      ]
     },
     "execution_count": 20,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "sims"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "c0d712c9",
   "metadata": {},
   "source": [
    "&emsp;&emsp;我们返回第二个即可："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "id": "ca1e0cc2",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "('Is Kaggle dead?', '/Is-Kaggle-dead')"
      ]
     },
     "execution_count": 23,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "vec_base[1][\"question\"], vec_base[1][\"answer\"]"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "0878f9b7",
   "metadata": {},
   "source": [
    "&emsp;&emsp;当然，在实际中，我们不建议使用循环，大家可以使用NumPy进行批量计算。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 27,
   "id": "4e97b201",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(5, 12288)"
      ]
     },
     "execution_count": 27,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "arr = np.array(\n",
    "    [v[\"embedding\"] for v in vec_base]\n",
    ")\n",
    "arr.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 47,
   "id": "bf47ca7c",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(1, 12288)"
      ]
     },
     "execution_count": 47,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "q_arr = np.expand_dims(q_emb, 0)\n",
    "q_arr.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 48,
   "id": "8090b806",
   "metadata": {},
   "outputs": [],
   "source": [
    "from sklearn.metrics.pairwise import cosine_similarity"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 50,
   "id": "06df6f50",
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([[0.6657692 ],\n",
       "       [0.87117754],\n",
       "       [0.74898532],\n",
       "       [0.73843577],\n",
       "       [0.72871292]])"
      ]
     },
     "execution_count": 50,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "cosine_similarity(arr, q_arr)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "a26fc82b",
   "metadata": {},
   "source": [
    "&emsp;&emsp;不过，当Question非常多，比如上百万甚至上亿时，这种方式就不合适了。一个是内存里可能放不下，另一个是算起来也很慢。这时候就必须借助一些专门用来做语义检索的工具了。"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "f608a1fd",
   "metadata": {},
   "source": [
    "&emsp;&emsp;比较常用的工具有：\n",
    "\n",
    "- [facebookresearch/faiss: A library for efficient similarity search and clustering of dense vectors.](https://github.com/facebookresearch/faiss)\n",
    "- [milvus-io/milvus: Vector database for scalable similarity search and AI applications.](https://github.com/milvus-io/milvus)\n",
    "- [Vector similarity | Redis](https://redis.io/docs/stack/search/reference/vectors/)\n",
    "\n",
    "此处，我们以Redis为例，其他工具用法类似。"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "f62d6616",
   "metadata": {},
   "source": [
    "&emsp;&emsp;首先，我们需要一个redis，建议使用docker直接运行：\n",
    "\n",
    "```shell\n",
    "docker run -p 6379:6379 -it redis/redis-stack:latest\n",
    "```\n",
    "\n",
    "执行后，docker会自动从hub把镜像拉到本地，默认是6379端口。"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "f582fb4f",
   "metadata": {},
   "source": [
    "&emsp;&emsp;然后安装redis-py，也就是Redis的Python客户端：\n",
    "\n",
    "```shell\n",
    "pip install redis\n",
    "```\n",
    "\n",
    "这样我们就可以用Python和Redis进行交互了。"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "471bbcef",
   "metadata": {},
   "source": [
    "&emsp;&emsp;先来个最简单的例子："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 49,
   "id": "1c69cfcd",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "True"
      ]
     },
     "execution_count": 49,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "import redis\n",
    "r = redis.Redis()\n",
    "r.set(\"key\", \"value\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "id": "79532818",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "b'value'"
      ]
     },
     "execution_count": 14,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "r.get(\"key\")"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "8103554d",
   "metadata": {},
   "source": [
    "&emsp;&emsp;如果大家使用过ElasticSearch，接下来的内容会非常容易理解。总的来说，和刚刚的步骤差不多，但是这里我们需要先建索引，然后生成Embedding并把它存储到Redis，再进行使用（从索引中搜索）。不过由于我们使用了工具，具体步骤会略微不同。"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "d5a6d6dd",
   "metadata": {},
   "source": [
    "&emsp;&emsp;索引的概念和数据库中的索引有点相似，就是要定义一组Schema，告诉Redis你的字段是什么，有哪些属性。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 50,
   "id": "afa5b12c",
   "metadata": {},
   "outputs": [],
   "source": [
    "VECTOR_DIM = 12288\n",
    "INDEX_NAME = \"faq\""
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "id": "e3bbdeaf",
   "metadata": {},
   "outputs": [],
   "source": [
    "from redis.commands.search.query import Query\n",
    "from redis.commands.search.field import TextField, VectorField\n",
    "from redis.commands.search.indexDefinition import IndexDefinition"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "id": "cdf4ab06",
   "metadata": {},
   "outputs": [],
   "source": [
    "# 建好要存字段的索引，针对不同属性字段，使用不同Field\n",
    "question = TextField(name=\"question\")\n",
    "answer = TextField(name=\"answer\")\n",
    "embedding = VectorField(\n",
    "    name=\"embedding\", \n",
    "    algorithm=\"HNSW\", \n",
    "    attributes={\n",
    "        \"TYPE\": \"FLOAT32\",\n",
    "        \"DIM\": VECTOR_DIM,\n",
    "        \"DISTANCE_METRIC\": \"COSINE\"\n",
    "    }\n",
    ")\n",
    "schema = (question, embedding, answer)\n",
    "index = r.ft(INDEX_NAME)\n",
    "try:\n",
    "    info = index.info()\n",
    "except:\n",
    "    index.create_index(schema, definition=IndexDefinition(prefix=[INDEX_NAME + \"-\"]))"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "10793899",
   "metadata": {},
   "source": [
    ">Hierarchical Navigable Small Worlds "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 320,
   "id": "14b22b25",
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "b'OK'"
      ]
     },
     "execution_count": 320,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# 如果需要删除已有文档的话，可以使用下面的命令\n",
    "index.dropindex(delete_documents=True)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "8ef78580",
   "metadata": {},
   "source": [
    "&emsp;&emsp;接下来就是把数据存到Redis。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 322,
   "id": "71c0b706",
   "metadata": {},
   "outputs": [],
   "source": [
    "for v in df.head().itertuples():\n",
    "    emb = get_embedding(v.Questions)\n",
    "    # 注意，redis要存储bytes或string\n",
    "    emb = np.array(emb, dtype=np.float32).tobytes()\n",
    "    im = {\n",
    "        \"question\": v.Questions,\n",
    "        \"embedding\": emb,\n",
    "        \"answer\": v.Link\n",
    "    }\n",
    "    # 重点是这句\n",
    "    r.hset(name=f\"{INDEX_NAME}-{v.Index}\", mapping=im)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "13f954c8",
   "metadata": {},
   "source": [
    "&emsp;&emsp;然后我们就可以进行搜索查询了，这一步构造查询输入稍微有一点麻烦。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "f73cd66f",
   "metadata": {},
   "outputs": [],
   "source": [
    "# 构造查询输入\n",
    "query = \"kaggle alive?\"\n",
    "embed_query = get_embedding(query)\n",
    "params_dict = {\"query_embedding\": np.array(embed_query).astype(dtype=np.float32).tobytes()}"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "id": "8f64e696",
   "metadata": {
    "code_folding": []
   },
   "outputs": [],
   "source": [
    "k = 3\n",
    "# {some filter query}=>[ KNN {num|$num} @vector_field $query_vec]\n",
    "base_query = f\"* => [KNN {k} @embedding $query_embedding AS score]\"\n",
    "return_fields = [\"question\", \"answer\", \"score\"]\n",
    "query = (\n",
    "    Query(base_query)\n",
    "     .return_fields(*return_fields)\n",
    "     .sort_by(\"score\")\n",
    "     .paging(0, k)\n",
    "     .dialect(2)\n",
    ")"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "98ed9b6b",
   "metadata": {},
   "source": [
    "> KNN（K最近邻算法），简单来说就是对未知点，分别和已有的点算距离，挑距离最近的K个点。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 365,
   "id": "00caba57",
   "metadata": {
    "scrolled": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "faq-1, Is Kaggle dead?, /Is-Kaggle-dead (Score: 0.831)\n",
      "faq-2, How should a beginner get started on Kaggle?, /How-should-a-beginner-get-started-on-Kaggle (Score: 0.735)\n",
      "faq-3, What are some alternatives to Kaggle?, /What-are-some-alternatives-to-Kaggle (Score: 0.73)\n"
     ]
    }
   ],
   "source": [
    "# 查询\n",
    "res = index.search(query, params_dict)\n",
    "for i,doc in enumerate(res.docs):\n",
    "    similarity = 1 - float(doc.score)\n",
    "    print(f\"{doc.id}, {doc.question}, {doc.answer} (Similarity: {round(score ,3) })\")"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "8dc777b1",
   "metadata": {},
   "source": [
    "&emsp;&emsp;上面，我们通过几种不同的方法为大家介绍了如何使用Embedding进行QA任务。简单回顾一下，要做QA任务首先咱们得有一个QA库，这些QA就是我们的仓库，每当一个新的问题过来时，我们就用这个问题去和咱们仓库里的每一个Q去匹配，然后找到最相似的那个，接下来就把该问题的Answer当做新问题的Answer交给用户。"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "3dae402c",
   "metadata": {},
   "source": [
    "&emsp;&emsp;这个任务的核心就是如何找到这个最相似的，涉及两个知识点：如何表示一个Question，以及如何查找到相似的Question。对于第一点，我们用API提供的Embedding表示，我们可以把它当做一个黑盒子，输入任意长度的文本，输出一个向量。查找相似问题则主要是用到相似度算法，语义相似度一般用cosine距离来衡量。"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "d71cbb7f",
   "metadata": {},
   "source": [
    "&emsp;&emsp;当然实际中可能会更加复杂一些，比如我们可能除了使用语义匹配，还会使用字词匹配（经典的做法）。而且，一般都会找到topN个相似的，然后对这topN个结果进行排序，选出最可能的那个。不过，这个前面我们举过例子了，完全可以通过ChatGPT来解决，让它帮你选出最好的那个。"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "a6db5f4b",
   "metadata": {},
   "source": [
    "### 3.2 聚类"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "a145fa4a",
   "metadata": {},
   "source": [
    "&emsp;&emsp;聚类的意思是把彼此相近的样本聚集在一起，本质也是在使用一种表示和相似度衡量来处理文本。比如我们有大量的未分类文本，如果事先能知道有几种类别，就可以用聚类的方法先将样本大致分一下。"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "049b10f4",
   "metadata": {},
   "source": [
    "&emsp;&emsp;我们使用Kaggle的DBPedia数据集：[DBPedia Classes | Kaggle](https://www.kaggle.com/datasets/danofer/dbpedia-classes?select=DBPEDIA_val.csv)。\n",
    "\n",
    "&emsp;&emsp;这个数据集对一段文本会给出三个不同层次级别的分类标签，我们这里用第一层的类别。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "id": "bcbf3c00",
   "metadata": {},
   "outputs": [],
   "source": [
    "import pandas as pd"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "id": "bb693b27",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(36003, 4)"
      ]
     },
     "execution_count": 2,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "df = pd.read_csv(\"./dataset/DBPEDIA_val.csv\")\n",
    "df.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "id": "1ddd54fd",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>text</th>\n",
       "      <th>l1</th>\n",
       "      <th>l2</th>\n",
       "      <th>l3</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>Li Curt is a station on the Bernina Railway li...</td>\n",
       "      <td>Place</td>\n",
       "      <td>Station</td>\n",
       "      <td>RailwayStation</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>Grafton State Hospital was a psychiatric hospi...</td>\n",
       "      <td>Place</td>\n",
       "      <td>Building</td>\n",
       "      <td>Hospital</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>The Democratic Patriotic Alliance of Kurdistan...</td>\n",
       "      <td>Agent</td>\n",
       "      <td>Organisation</td>\n",
       "      <td>PoliticalParty</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3</th>\n",
       "      <td>Ira Rakatansky (October 3, 1919 – March 4, 201...</td>\n",
       "      <td>Agent</td>\n",
       "      <td>Person</td>\n",
       "      <td>Architect</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>4</th>\n",
       "      <td>Universitatea Reșița is a women handball club ...</td>\n",
       "      <td>Agent</td>\n",
       "      <td>SportsTeam</td>\n",
       "      <td>HandballTeam</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "                                                text     l1            l2  \\\n",
       "0  Li Curt is a station on the Bernina Railway li...  Place       Station   \n",
       "1  Grafton State Hospital was a psychiatric hospi...  Place      Building   \n",
       "2  The Democratic Patriotic Alliance of Kurdistan...  Agent  Organisation   \n",
       "3  Ira Rakatansky (October 3, 1919 – March 4, 201...  Agent        Person   \n",
       "4  Universitatea Reșița is a women handball club ...  Agent    SportsTeam   \n",
       "\n",
       "               l3  \n",
       "0  RailwayStation  \n",
       "1        Hospital  \n",
       "2  PoliticalParty  \n",
       "3       Architect  \n",
       "4    HandballTeam  "
      ]
     },
     "execution_count": 3,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "df.head()"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "7a2daff3",
   "metadata": {},
   "source": [
    "查看一下类别数量："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "id": "013b32ae",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "Agent             18647\n",
       "Place              6855\n",
       "Species            3210\n",
       "Work               3141\n",
       "Event              2854\n",
       "SportsSeason        879\n",
       "UnitOfWork          263\n",
       "TopicalConcept      117\n",
       "Device               37\n",
       "Name: l1, dtype: int64"
      ]
     },
     "execution_count": 6,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "df.l1.value_counts()"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "cb146545",
   "metadata": {},
   "source": [
    "数量有点多，我们随机采样200条："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "id": "9c22c434",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "Agent             102\n",
       "Place              31\n",
       "Work               22\n",
       "Species            19\n",
       "Event              12\n",
       "SportsSeason       10\n",
       "UnitOfWork          3\n",
       "TopicalConcept      1\n",
       "Name: l1, dtype: int64"
      ]
     },
     "execution_count": 7,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "sdf = df.sample(200)\n",
    "sdf.l1.value_counts()"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "65691d50",
   "metadata": {},
   "source": [
    "&emsp;&emsp;为了便于观察，我们只保留3个数量差不多的类别：Place、Work和Species。（类型太多时，样本点会混在一块难以观察，大家可以自己尝试。）"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 98,
   "id": "06a37228",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(72, 6)"
      ]
     },
     "execution_count": 98,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "cdf = sdf[\n",
    "    (sdf.l1 == \"Place\") | (sdf.l1 == \"Work\") | (sdf.l1 == \"Species\")\n",
    "]\n",
    "cdf.shape"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "42ce2778",
   "metadata": {},
   "source": [
    "&emsp;&emsp;接下来先把文本变成向量："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 99,
   "id": "a3042ac4",
   "metadata": {},
   "outputs": [],
   "source": [
    "from openai.embeddings_utils import get_embedding, cosine_similarity\n",
    "import openai\n",
    "import numpy as np\n",
    "OPENAI_API_KEY = \"填入专属的API key\"\n",
    "openai.api_key = OPENAI_API_KEY"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "5c7a8186",
   "metadata": {},
   "source": [
    "&emsp;&emsp;我们前面提到过了，这里`get_embedding`可以支持多种模型（engine）选项，它默认的是：`text-similarity-davinci-001`，我们这里用另一个：`text-embedding-ada-002`，这个比较快（它的维度比前面的少很多）。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 100,
   "id": "03626b17",
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "<ipython-input-100-85f47cb77106>:1: SettingWithCopyWarning: \n",
      "A value is trying to be set on a copy of a slice from a DataFrame.\n",
      "Try using .loc[row_indexer,col_indexer] = value instead\n",
      "\n",
      "See the caveats in the documentation: https://pandas.pydata.org/pandas-docs/stable/user_guide/indexing.html#returning-a-view-versus-a-copy\n",
      "  cdf[\"embedding\"] = cdf.text.apply(lambda x: get_embedding(x, engine=\"text-embedding-ada-002\"))\n"
     ]
    }
   ],
   "source": [
    "cdf[\"embedding\"] = cdf.text.apply(lambda x: get_embedding(x, engine=\"text-embedding-ada-002\"))"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "ae9200b4",
   "metadata": {},
   "source": [
    "&emsp;&emsp;接下来用PCA（主成分分析）给降维，将原来的向量从1536维降到3维，便于显示。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 101,
   "id": "beb3f2d3",
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "<ipython-input-101-3bc0b46a423a>:5: SettingWithCopyWarning: \n",
      "A value is trying to be set on a copy of a slice from a DataFrame.\n",
      "Try using .loc[row_indexer,col_indexer] = value instead\n",
      "\n",
      "See the caveats in the documentation: https://pandas.pydata.org/pandas-docs/stable/user_guide/indexing.html#returning-a-view-versus-a-copy\n",
      "  cdf[\"embed_vis\"] = vis_dims.tolist()\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "((72, 1536), (72, 3))"
      ]
     },
     "execution_count": 101,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "from sklearn.decomposition import PCA\n",
    "arr = np.array(cdf.embedding.to_list())\n",
    "pca = PCA(n_components=3)\n",
    "vis_dims = pca.fit_transform(arr)\n",
    "cdf[\"embed_vis\"] = vis_dims.tolist()\n",
    "arr.shape, vis_dims.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 108,
   "id": "19e6be0c",
   "metadata": {
    "scrolled": false
   },
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 576x576 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "%matplotlib inline\n",
    "import matplotlib.pyplot as plt\n",
    "import numpy as np\n",
    "\n",
    "fig, ax = plt.subplots(subplot_kw={\"projection\": \"3d\"}, figsize=(8, 8))\n",
    "cmap = plt.get_cmap(\"tab20\")\n",
    "categories = sorted(cdf.l1.unique())\n",
    "\n",
    "# 分别绘制每个类别\n",
    "for i, cat in enumerate(categories):\n",
    "    sub_matrix = np.array(cdf[cdf.l1 == cat][\"embed_vis\"].to_list())\n",
    "    x=sub_matrix[:, 0]\n",
    "    y=sub_matrix[:, 1]\n",
    "    z=sub_matrix[:, 2]\n",
    "    colors = [cmap(i/len(categories))] * len(sub_matrix)\n",
    "    ax.scatter(x, y, z, c=colors, label=cat)\n",
    "\n",
    "ax.legend(bbox_to_anchor=(1.2, 1))\n",
    "plt.show();"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "ec8c20d7",
   "metadata": {},
   "source": [
    "&emsp;&emsp;可以比较明显的看出，三个不同类型的数据分别在不同的位置。"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "4a4a5c96",
   "metadata": {},
   "source": [
    "### 3.3 推荐"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "81667fec",
   "metadata": {},
   "source": [
    "&emsp;&emsp;我们在很多APP或网站上都能看到推荐功能。比如在购物网站，每当你登陆或者选购一件商品后，系统就会给你推荐一些相关的产品。在这一小节中，我们就来做一个类似的应用，不过我们推荐的不是商品，而是文本，比如帖子、文章、新闻等。"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "c9c6d1b2",
   "metadata": {},
   "source": [
    "&emsp;&emsp;我们以新闻为例，先说一下基本逻辑：\n",
    "\n",
    "- 首先要有一个基础的文章库，可能包括标题、内容、标签等。\n",
    "- 计算已有文章的Embedding并存储。\n",
    "- 根据用户浏览记录，推荐和浏览记录最相似的文章。"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "8e08ebc0",
   "metadata": {},
   "source": [
    "&emsp;&emsp;使用如下的数据集：[AG News Classification Dataset | Kaggle](https://www.kaggle.com/datasets/amananandrai/ag-news-classification-dataset?select=train.csv)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "a4ffd147",
   "metadata": {},
   "source": [
    "&emsp;&emsp;看起来好像和前面的QA差不多，事实也的确如此，因为它们本质上都是相似匹配问题。只不过QA使用的是用户的Question去匹配已有知识库，而推荐是使用用户的浏览记录去匹配。但是很明显，推荐相比QA要更复杂一些，主要包括以下几个方面：\n",
    "\n",
    "- 刚开始用户没有记录时的推荐（一般行业称为冷启动问题）。\n",
    "- 除了相似还有其他要考虑的因素：比如热门内容、新内容、内容多样性、随时间变化的兴趣变化等等。\n",
    "- 编码（Embedding输入）问题：我们应该取标题呢，还是文章，还是简要描述或者摘要，还是都要计算。\n",
    "- 规模问题：推荐面临的量级一般会远超QA，除了横向扩展机器，是否能从流程和算法设计上提升效率。\n",
    "- 用户反馈对推荐系统的影响问题：用户反感或喜欢与文章本身并没有直接关系，比如用户喜欢体育新闻但讨厌中国足球。\n",
    "- 线上实时更新问题。"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "5afd0385",
   "metadata": {},
   "source": [
    "&emsp;&emsp;当然，一个完整的线上系统要考虑的因素可能更多。我们给大家列出这些只是希望读者在设计一个方案时能够充分调研和考虑，同时结合实际情况进行。反过来说，可能并不需要考虑上面的每个因素。所以，大家在实际操作时一定要活学活用，充分理解需求后再动手实施。"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "107c5d93",
   "metadata": {},
   "source": [
    "&emsp;&emsp;我们这里综合考虑上面的因素给大家一个比较简单的方案，但务必注意，其中每个模块的方案都不是唯一的。整体设计如下：\n",
    "\n",
    "- 用户注册登录时，让其选择感兴趣的类型（如体育、音乐、时尚等），我们通过这一步将用户框在一个大的范围内，同时用来顺道解决冷启动问题。\n",
    "- 给用户推荐内容时，在知道类别（用户注册时选择+浏览记录）后，应依次考虑时效性、热门程度、多样性等。\n",
    "- 考虑到性能问题，可以编码「标题+摘要」。\n",
    "- 对大类别进一步细分，只在细分类别里进行相似度计算。\n",
    "- 记录用户实时行为（如浏览Item、浏览时长、评论、收藏、点赞、转发等）。\n",
    "- 动态更新内容库，更新用户行为库。"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "6c5a81c4",
   "metadata": {},
   "source": [
    "&emsp;&emsp;在具体实施时，我们使用最常用的流程线方案：召回+排序。\n",
    "\n",
    "- 召回：通过各种不同属性或特征（如用户偏好、热点、行为等）先找到一批要推荐列表。\n",
    "- 排序：根据多样性、时效性、用户反馈、热门程度等属性对结果进行排序。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "id": "b2594de7",
   "metadata": {},
   "outputs": [],
   "source": [
    "from dataclasses import dataclass\n",
    "import pandas as pd"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "id": "b92fdab4",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(120000, 3)"
      ]
     },
     "execution_count": 2,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "df = pd.read_csv(\"./dataset/AG_News.csv\")\n",
    "df.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "id": "aa169a6d",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>Class Index</th>\n",
       "      <th>Title</th>\n",
       "      <th>Description</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>3</td>\n",
       "      <td>Wall St. Bears Claw Back Into the Black (Reuters)</td>\n",
       "      <td>Reuters - Short-sellers, Wall Street's dwindli...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>3</td>\n",
       "      <td>Carlyle Looks Toward Commercial Aerospace (Reu...</td>\n",
       "      <td>Reuters - Private investment firm Carlyle Grou...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>3</td>\n",
       "      <td>Oil and Economy Cloud Stocks' Outlook (Reuters)</td>\n",
       "      <td>Reuters - Soaring crude prices plus worries\\ab...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3</th>\n",
       "      <td>3</td>\n",
       "      <td>Iraq Halts Oil Exports from Main Southern Pipe...</td>\n",
       "      <td>Reuters - Authorities have halted oil export\\f...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>4</th>\n",
       "      <td>3</td>\n",
       "      <td>Oil prices soar to all-time record, posing new...</td>\n",
       "      <td>AFP - Tearaway world oil prices, toppling reco...</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "   Class Index                                              Title  \\\n",
       "0            3  Wall St. Bears Claw Back Into the Black (Reuters)   \n",
       "1            3  Carlyle Looks Toward Commercial Aerospace (Reu...   \n",
       "2            3    Oil and Economy Cloud Stocks' Outlook (Reuters)   \n",
       "3            3  Iraq Halts Oil Exports from Main Southern Pipe...   \n",
       "4            3  Oil prices soar to all-time record, posing new...   \n",
       "\n",
       "                                         Description  \n",
       "0  Reuters - Short-sellers, Wall Street's dwindli...  \n",
       "1  Reuters - Private investment firm Carlyle Grou...  \n",
       "2  Reuters - Soaring crude prices plus worries\\ab...  \n",
       "3  Reuters - Authorities have halted oil export\\f...  \n",
       "4  AFP - Tearaway world oil prices, toppling reco...  "
      ]
     },
     "execution_count": 3,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "df.head()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "id": "5532f89f",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "3    30000\n",
       "4    30000\n",
       "2    30000\n",
       "1    30000\n",
       "Name: Class Index, dtype: int64"
      ]
     },
     "execution_count": 21,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "df[\"Class Index\"].value_counts()"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "44389263",
   "metadata": {},
   "source": [
    "&emsp;&emsp;根据数据集介绍（上面的链接），四个类型分别是：1-World, 2-Sports, 3-Business, 4-Sci/Tech，每个类型有30000条数据，共12万条。接下来，我们将使用上面已经介绍的知识来做一个简单的流水线系统。"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "20146b61",
   "metadata": {},
   "source": [
    "&emsp;&emsp;为了便于运行，依然取100条sample作为示例："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "id": "bb785111",
   "metadata": {},
   "outputs": [],
   "source": [
    "sdf = df.sample(100)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "id": "aeb62886",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "2    28\n",
       "4    26\n",
       "1    24\n",
       "3    22\n",
       "Name: Class Index, dtype: int64"
      ]
     },
     "execution_count": 10,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "sdf[\"Class Index\"].value_counts()"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "ff52441d",
   "metadata": {},
   "source": [
    "&emsp;&emsp;首先维护一个用户偏好和行为记录："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 35,
   "id": "7a93b84c",
   "metadata": {},
   "outputs": [],
   "source": [
    "from typing import List\n",
    "@dataclass\n",
    "class User:\n",
    "    \n",
    "    user_name: str\n",
    "\n",
    "@dataclass\n",
    "class UserPrefer:\n",
    "    \n",
    "    user_name: str\n",
    "    prefers: List[int]\n",
    "\n",
    "\n",
    "@dataclass\n",
    "class Item:\n",
    "    \n",
    "    item_id: str\n",
    "    item_props: dict\n",
    "\n",
    "\n",
    "@dataclass\n",
    "class Action:\n",
    "    \n",
    "    action_type: str\n",
    "    action_props: dict\n",
    "\n",
    "\n",
    "@dataclass\n",
    "class UserAction:\n",
    "    \n",
    "    user: User\n",
    "    item: Item\n",
    "    action: Action\n",
    "    action_time: str"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 232,
   "id": "509c10ee",
   "metadata": {},
   "outputs": [],
   "source": [
    "u1 = User(\"u1\")\n",
    "up1 = UserPrefer(\"u1\", [1, 2])\n",
    "# sdf.iloc[1] 正好是sport（类别为2）\n",
    "i1 = Item(\"i1\", {\n",
    "    \"id\": 1, \n",
    "    \"catetory\": \"sport\",\n",
    "    \"title\": \"Swimming: Shibata Joins Japanese Gold Rush\", \n",
    "    \"description\": \"\\\n",
    "    ATHENS (Reuters) - Ai Shibata wore down French teen-ager  Laure Manaudou to win the women's 800 meters \\\n",
    "    freestyle gold  medal at the Athens Olympics Friday and provide Japan with  their first female swimming \\\n",
    "    champion in 12 years.\", \n",
    "    \"content\": \"content\"\n",
    "})\n",
    "a1 = Action(\"浏览\", {\n",
    "    \"open_time\": \"2023-04-01 12:00:00\", \n",
    "    \"leave_time\": \"2023-04-01 14:00:00\",\n",
    "    \"type\": \"close\",\n",
    "    \"duration\": \"2hour\"\n",
    "})\n",
    "ua1 = UserAction(u1, i1, a1, \"2023-04-01 12:00:00\")"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "0170ac93",
   "metadata": {},
   "source": [
    "&emsp;&emsp;计算所有文本的Embedding，这一步和之前一样："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 77,
   "id": "ab142d4e",
   "metadata": {},
   "outputs": [],
   "source": [
    "from openai.embeddings_utils import get_embedding, cosine_similarity\n",
    "from sklearn.metrics.pairwise import cosine_similarity\n",
    "import openai\n",
    "import numpy as np\n",
    "\n",
    "OPENAI_API_KEY = \"填入专属的API key\"\n",
    "openai.api_key = OPENAI_API_KEY"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "id": "0def4599",
   "metadata": {},
   "outputs": [],
   "source": [
    "sdf[\"embedding\"] = sdf.apply(lambda x: \n",
    "                             get_embedding(x.Title + x.Description, engine=\"text-embedding-ada-002\"), axis=1)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "390b98e5",
   "metadata": {},
   "source": [
    "&emsp;&emsp;我们先处理一下召回："
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 214,
   "id": "dc1590b5",
   "metadata": {},
   "outputs": [],
   "source": [
    "import random\n",
    "class Recall:\n",
    "    \n",
    "    def __init__(self, df: pd.DataFrame):\n",
    "        self.data = df\n",
    "    \n",
    "    def user_prefer_recall(self, user, n):\n",
    "        up = self.get_user_prefers(user)\n",
    "        idx = random.randrange(0, len(up.prefers))\n",
    "        return self.pick_by_idx(idx, n)\n",
    "    \n",
    "    def hot_recall(self, n):\n",
    "        # 随机进行示例\n",
    "        df = self.data.sample(n)\n",
    "        return df\n",
    "    \n",
    "    def user_action_recall(self, user, n):\n",
    "        actions = self.get_user_actions(user)\n",
    "        interest = self.get_most_interested_item(actions)\n",
    "        recoms = self.recommend_by_interest(interest, n)\n",
    "        return recoms\n",
    "    \n",
    "    def get_most_interested_item(self, user_action):\n",
    "        \"\"\"\n",
    "        可以选近一段时间内用户交互时间、次数、评论（相关属性）过的Item\n",
    "        \"\"\"\n",
    "        # 就是sdf的第2行，idx为1的那条作为最喜欢（假设）\n",
    "        # 是一条游泳相关的Item\n",
    "        idx = user_action.item.item_props[\"id\"]\n",
    "        im = self.data.iloc[idx]\n",
    "        return im\n",
    "    \n",
    "    def recommend_by_interest(self, interest, n):\n",
    "        cate_id = interest[\"Class Index\"]\n",
    "        q_emb = interest[\"embedding\"]\n",
    "        # 确定类别\n",
    "        base = self.data[self.data[\"Class Index\"] == cate_id]\n",
    "        # 此处可以复用QA那一段代码，用给定embedding计算base中embedding的相似度\n",
    "        base_arr = np.array(\n",
    "            [v.embedding for v in base.itertuples()]\n",
    "        )\n",
    "        q_arr = np.expand_dims(q_emb, 0)\n",
    "        sims = cosine_similarity(base_arr, q_arr)\n",
    "        # 排除掉自己\n",
    "        idxes = sims.argsort(0).squeeze()[-(n+1):-1]\n",
    "        return base.iloc[reversed(idxes.tolist())]\n",
    "    \n",
    "    def pick_by_idx(self, category, n):\n",
    "        df = self.data[self.data[\"Class Index\"] == category]\n",
    "        return df.sample(n)\n",
    "    \n",
    "    def get_user_actions(self, user):\n",
    "        dct = {\"u1\": ua1}\n",
    "        return dct[user.user_name]\n",
    "    \n",
    "    def get_user_prefers(self, user):\n",
    "        dct = {\"u1\": up1}\n",
    "        return dct[user.user_name]\n",
    "    \n",
    "    def run(self, user):\n",
    "        ur = self.user_action_recall(user, 5)\n",
    "        if len(ur) == 0:\n",
    "            ur = self.user_prefer_recall(user, 5)\n",
    "        hr = self.hot_recall(3)\n",
    "        return pd.concat([ur, hr], axis=0)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 224,
   "id": "2a43c274",
   "metadata": {},
   "outputs": [],
   "source": [
    "r = Recall(sdf)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 225,
   "id": "1b2993db",
   "metadata": {},
   "outputs": [],
   "source": [
    "rd = r.run(u1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 227,
   "id": "a8f4e021",
   "metadata": {
    "scrolled": false
   },
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>Class Index</th>\n",
       "      <th>Title</th>\n",
       "      <th>Description</th>\n",
       "      <th>embedding</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>12120</th>\n",
       "      <td>2</td>\n",
       "      <td>Olympics Wrap: Another Doping Controversy Surf...</td>\n",
       "      <td>ATHENS (Reuters) - Olympic chiefs ordered Hun...</td>\n",
       "      <td>[0.013697294518351555, 0.012140628881752491, 0...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>5905</th>\n",
       "      <td>2</td>\n",
       "      <td>Saturday Night #39;s Alright for Blighty</td>\n",
       "      <td>Matthew Pinsents coxless four team, sailor Ben...</td>\n",
       "      <td>[-0.012345104478299618, -0.0025237693917006254...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>29729</th>\n",
       "      <td>2</td>\n",
       "      <td>Beijing Paralympic Games to be fabulous: IPC P...</td>\n",
       "      <td>The 13th Summer Paralympic Games in 2008 in Be...</td>\n",
       "      <td>[-0.009852061048150063, 0.017894696444272995, ...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>27215</th>\n",
       "      <td>2</td>\n",
       "      <td>Dent tops Luczak to win at China Open</td>\n",
       "      <td>Taylor Dent defeated Australian qualifier Pete...</td>\n",
       "      <td>[-0.004778657108545303, 0.014275987632572651, ...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>72985</th>\n",
       "      <td>2</td>\n",
       "      <td>Rusedski through in St Petersburg</td>\n",
       "      <td>Greg Rusedski eased into the second round of t...</td>\n",
       "      <td>[-0.007127437274903059, 0.0025771241635084152,...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>28344</th>\n",
       "      <td>3</td>\n",
       "      <td>Delta pilots wary of retirements</td>\n",
       "      <td>Union says pilots may retire en masse to get p...</td>\n",
       "      <td>[-0.03769957274198532, -0.032835111021995544, ...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>80374</th>\n",
       "      <td>2</td>\n",
       "      <td>Everett powerless in loss to Prince George</td>\n",
       "      <td>Besides the final score, there is only one sta...</td>\n",
       "      <td>[-0.014837506227195263, -0.015726948156952858,...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>64648</th>\n",
       "      <td>4</td>\n",
       "      <td>New Screening Technology Is Nigh</td>\n",
       "      <td>Machines built to find weapons hidden in cloth...</td>\n",
       "      <td>[-0.020757483318448067, -0.017689339816570282,...</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "       Class Index                                              Title  \\\n",
       "12120            2  Olympics Wrap: Another Doping Controversy Surf...   \n",
       "5905             2           Saturday Night #39;s Alright for Blighty   \n",
       "29729            2  Beijing Paralympic Games to be fabulous: IPC P...   \n",
       "27215            2              Dent tops Luczak to win at China Open   \n",
       "72985            2                  Rusedski through in St Petersburg   \n",
       "28344            3                   Delta pilots wary of retirements   \n",
       "80374            2         Everett powerless in loss to Prince George   \n",
       "64648            4                   New Screening Technology Is Nigh   \n",
       "\n",
       "                                             Description  \\\n",
       "12120   ATHENS (Reuters) - Olympic chiefs ordered Hun...   \n",
       "5905   Matthew Pinsents coxless four team, sailor Ben...   \n",
       "29729  The 13th Summer Paralympic Games in 2008 in Be...   \n",
       "27215  Taylor Dent defeated Australian qualifier Pete...   \n",
       "72985  Greg Rusedski eased into the second round of t...   \n",
       "28344  Union says pilots may retire en masse to get p...   \n",
       "80374  Besides the final score, there is only one sta...   \n",
       "64648  Machines built to find weapons hidden in cloth...   \n",
       "\n",
       "                                               embedding  \n",
       "12120  [0.013697294518351555, 0.012140628881752491, 0...  \n",
       "5905   [-0.012345104478299618, -0.0025237693917006254...  \n",
       "29729  [-0.009852061048150063, 0.017894696444272995, ...  \n",
       "27215  [-0.004778657108545303, 0.014275987632572651, ...  \n",
       "72985  [-0.007127437274903059, 0.0025771241635084152,...  \n",
       "28344  [-0.03769957274198532, -0.032835111021995544, ...  \n",
       "80374  [-0.014837506227195263, -0.015726948156952858,...  \n",
       "64648  [-0.020757483318448067, -0.017689339816570282,...  "
      ]
     },
     "execution_count": 227,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# 共8个，5个用户行为推荐、3个热门\n",
    "rd"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "48a8002c",
   "metadata": {},
   "source": [
    "&emsp;&emsp;需要再次说明的是，上面的只是一个大致的流程，实际中有很多细节或优化点需要注意，比如：\n",
    "\n",
    "- 建数据库表（上面的`get_`其实都是查表）\n",
    "- 将Item、User和Action也进行Embedding，全部使用Embedding后再做召回\n",
    "- 对『感兴趣`get_most_interested_item`』更多的优化，考虑更多行为和反馈，召回更多不同类型条目\n",
    "- 性能和自动更新数据的考虑\n",
    "- 线上评测，A/B等"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "765ca69f",
   "metadata": {},
   "source": [
    "&emsp;&emsp;可以发现，我们虽然只做了召回一步，但其中涉及到的内容已经远远不止之前QA那一点了，QA用到的东西可能只是其中一小部分。不过事无绝对，即便是QA任务也可能根据实际情况不同需要做很多优化，比如召回+排序。但总体来说，类似推荐这样比较综合的系统相对来说会更加复杂一些。"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "07e10ecc",
   "metadata": {},
   "source": [
    "&emsp;&emsp;后面就是排序了，这一步需要区分不同的应用场景，可以做或不做，做的话（就是对刚刚得到的列表进行排序）也可以简单或复杂。比如简单地按发布时间，复杂的综合考虑多样性、时效性、用户反馈、热门程度等多种属性。具体操作时，可以直接按相关属性排序，也可以用模型排序。这里就不再继续深入了。"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "56189400",
   "metadata": {},
   "source": [
    "## 相关文献\n",
    "\n",
    "- 【1】[浅析文本分类 —— 情感分析与自然语言处理 | Yam](https://yam.gift/2021/10/27/NLP/2021-10-27-Senta/)\n",
    "- 【2】[句子表征综述 | Yam](https://yam.gift/2022/03/27/NLP/2022-03-27-Sentence-Representation-Summarization/)\n",
    "- 【3】[NLP 表征的历史与未来 | Yam](https://yam.gift/2020/12/12/NLP/2020-12-12-NLP-Representation-History-Future/)\n",
    "\n",
    "\n",
    "memo：\n",
    "\n",
    "- [openai-cookbook/Semantic_text_search_using_embeddings.ipynb at main · openai/openai-cookbook](https://github.com/openai/openai-cookbook/blob/main/examples/Semantic_text_search_using_embeddings.ipynb)\n",
    "- [openai-cookbook/getting-started-with-redis-and-openai.ipynb at main · openai/openai-cookbook](https://github.com/openai/openai-cookbook/blob/main/examples/vector_databases/redis/getting-started-with-redis-and-openai.ipynb)\n",
    "- [openai-cookbook/Visualizing_embeddings_in_3D.ipynb at main · openai/openai-cookbook](https://github.com/openai/openai-cookbook/blob/main/examples/Visualizing_embeddings_in_3D.ipynb)\n",
    "\n",
    "注①：感谢DataWhale成员于沼懿指出此处不严谨描述。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "abd26c9c",
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.9.13"
  },
  "latex_envs": {
   "LaTeX_envs_menu_present": true,
   "autoclose": false,
   "autocomplete": true,
   "bibliofile": "biblio.bib",
   "cite_by": "apalike",
   "current_citInitial": 1,
   "eqLabelWithNumbers": true,
   "eqNumInitial": 1,
   "hotkeys": {
    "equation": "Ctrl-E",
    "itemize": "Ctrl-I"
   },
   "labels_anchors": false,
   "latex_user_defs": false,
   "report_style_numbering": false,
   "user_envs_cfg": false
  },
  "toc": {
   "base_numbering": 1,
   "nav_menu": {},
   "number_sections": false,
   "sideBar": true,
   "skip_h1_title": false,
   "title_cell": "Table of Contents",
   "title_sidebar": "Contents",
   "toc_cell": true,
   "toc_position": {
    "height": "calc(100% - 180px)",
    "left": "10px",
    "top": "150px",
    "width": "175px"
   },
   "toc_section_display": true,
   "toc_window_display": true
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}