langchain/docs/docs/integrations/chat/maritalk.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "<a href=\"https://colab.research.google.com/github/langchain-ai/langchain/blob/master/docs/docs/integrations/chat/maritalk.ipynb\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a>\n",
    "\n",
    "# Maritalk\n",
    "\n",
    "## Introduction\n",
    "\n",
    "MariTalk is an assistant developed by the Brazilian company [Maritaca AI](https://www.maritaca.ai).\n",
    "MariTalk is based on language models that have been specially trained to understand Portuguese well.\n",
    "\n",
    "This notebook demonstrates how to use MariTalk with LangChain through two examples:\n",
    "\n",
    "1. A simple example of how to use MariTalk to perform a task.\n",
    "2. LLM + RAG: The second example shows how to answer a question whose answer is found in a long document that does not fit within the token limit of MariTalk. For this, we will use a simple searcher (BM25) to first search the document for the most relevant sections and then feed them to MariTalk for answering."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Installation\n",
    "First, install the LangChain library (and all its dependencies) using the following command:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "!pip install langchain langchain-core langchain-community"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## API Key\n",
    "You will need an API key that can be obtained from chat.maritaca.ai (\"Chaves da API\" section)."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "\n",
    "### Example 1 - Pet Name Suggestions\n",
    "\n",
    "Let's define our language model, ChatMaritalk, and configure it with your API key."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from langchain_community.chat_models import ChatMaritalk\n",
    "from langchain_core.output_parsers import StrOutputParser\n",
    "from langchain_core.prompts.chat import ChatPromptTemplate\n",
    "\n",
    "llm = ChatMaritalk(\n",
    "    model=\"sabia-2-medium\",  # Available models: sabia-2-small and sabia-2-medium\n",
    "    api_key=\"\",  # Insert your API key here\n",
    "    temperature=0.7,\n",
    "    max_tokens=100,\n",
    ")\n",
    "\n",
    "output_parser = StrOutputParser()\n",
    "\n",
    "chat_prompt = ChatPromptTemplate.from_messages(\n",
    "    [\n",
    "        (\n",
    "            \"system\",\n",
    "            \"You are an assistant specialized in suggesting pet names. Given the animal, you must suggest 4 names.\",\n",
    "        ),\n",
    "        (\"human\", \"I have a {animal}\"),\n",
    "    ]\n",
    ")\n",
    "\n",
    "chain = chat_prompt | llm | output_parser\n",
    "\n",
    "response = chain.invoke({\"animal\": \"dog\"})\n",
    "print(response)  # should answer something like \"1. Max\\n2. Bella\\n3. Charlie\\n4. Rocky\""
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Example 2 - RAG + LLM: UNICAMP 2024 Entrance Exam Question Answering System\n",
    "For this example, we need to install some extra libraries:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "!pip install unstructured rank_bm25 pdf2image pdfminer-six pikepdf pypdf unstructured_inference fastapi kaleido uvicorn \"pillow<10.1.0\" pillow_heif -q"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Loading the database\n",
    "\n",
    "The first step is to create a database with the information from the notice. For this, we will download the notice from the COMVEST website and segment the extracted text into 500-character windows."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from langchain.document_loaders import OnlinePDFLoader\n",
    "from langchain.text_splitter import RecursiveCharacterTextSplitter\n",
    "\n",
    "# Loading the COMVEST 2024 notice\n",
    "loader = OnlinePDFLoader(\n",
    "    \"https://www.comvest.unicamp.br/wp-content/uploads/2023/10/31-2023-Dispoe-sobre-o-Vestibular-Unicamp-2024_com-retificacao.pdf\"\n",
    ")\n",
    "data = loader.load()\n",
    "\n",
    "text_splitter = RecursiveCharacterTextSplitter(\n",
    "    chunk_size=500, chunk_overlap=100, separators=[\"\\n\", \" \", \"\"]\n",
    ")\n",
    "texts = text_splitter.split_documents(data)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Creating a Searcher\n",
    "Now that we have our database, we need a searcher. For this example, we will use a simple BM25 as a search system, but this could be replaced by any other searcher (such as search via embeddings)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from langchain.retrievers import BM25Retriever\n",
    "\n",
    "retriever = BM25Retriever.from_documents(texts)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "#### Combining Search System + LLM\n",
    "Now that we have our searcher, we just need to implement a prompt specifying the task and invoke the chain."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from langchain.chains.question_answering import load_qa_chain\n",
    "\n",
    "prompt = \"\"\"Baseado nos seguintes documentos, responda a pergunta abaixo.\n",
    "\n",
    "{context}\n",
    "\n",
    "Pergunta: {query}\n",
    "\"\"\"\n",
    "\n",
    "qa_prompt = ChatPromptTemplate.from_messages([(\"human\", prompt)])\n",
    "\n",
    "chain = load_qa_chain(llm, chain_type=\"stuff\", verbose=True, prompt=qa_prompt)\n",
    "\n",
    "query = \"Qual o tempo máximo para realização da prova?\"\n",
    "\n",
    "docs = retriever.invoke(query)\n",
    "\n",
    "chain.invoke(\n",
    "    {\"input_documents\": docs, \"query\": query}\n",
    ")  # Should output something like: \"O tempo máximo para realização da prova é de 5 horas.\""
   ]
  }
 ],
 "metadata": {
  "language_info": {
   "name": "python"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}