{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from tqdm.auto import tqdm\n",
    "import pandas as pd\n",
    "import time\n",
    "\n",
    "from langchain.document_loaders import PyMuPDFLoader\n",
    "from langchain.text_splitter import RecursiveCharacterTextSplitter\n",
    "\n",
    "pd.set_option(\"display.max_colwidth\", None)\n",
    "\n",
    "# Set ChatMistralAI API KEY\n",
    "# e.g., export MISTRAL_API_KEY==your_api_key_here\n",
    "# or save apy key in .env file\n",
    "from dotenv import load_dotenv\n",
    "load_dotenv()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Load pdf file\n",
    "filepath = \"data/documents/Brandt et al_2024_Kadi_info_page.pdf\"\n",
    "loader_module = PyMuPDFLoader\n",
    "loader = loader_module(filepath)\n",
    "document = loader.load()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Split docs into chunks\n",
    "text_splitter = RecursiveCharacterTextSplitter(\n",
    "    chunk_size=2000,\n",
    "    chunk_overlap=200,\n",
    "    add_start_index=True,\n",
    "    separators=[\"\\n\\n\", \"\\n\", \".\", \" \", \"\"],\n",
    ")\n",
    "\n",
    "docs_processed = []\n",
    "for doc in document:\n",
    "    docs_processed += text_splitter.split_documents([doc])\n",
    "\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Create LLM, here we use MistralAI\n",
    "from langchain_mistralai.chat_models import ChatMistralAI\n",
    "\n",
    "llm = ChatMistralAI(\n",
    "    model=\"mistral-large-latest\"\n",
    ")\n",
    "\n",
    "llm.invoke(\"hello\")  # test llm"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "QA_generation_prompt = \"\"\"\n",
    "Your task is to write a factoid question and an answer given a context.\n",
    "Your factoid question should be answerable with a specific, concise piece of factual information from the context.\n",
    "Your factoid question should be formulated in the same style as questions users could ask in a search engine. Users are usually scientific researchers in the field of materials science.\n",
    "This means that your factoid question MUST NOT mention something like \"according to the passage\" or \"context\".\n",
    "Please ask the specific question instead of the general question, like 'What is the key information in the given paragraph?'.\n",
    "\n",
    "Provide your answer as follows:\n",
    "\n",
    "Output:::\n",
    "Factoid question: (your factoid question)\n",
    "Answer: (your answer to the factoid question)\n",
    "\n",
    "Now here is the context.\n",
    "\n",
    "Context: {context}\\n\n",
    "Output:::\"\"\"\n",
    "\n",
    "# Or\n",
    "# Ref: https://mlflow.org/docs/latest/llms/rag/notebooks/question-generation-retrieval-evaluation.html\n",
    "# QA_generation_prompt = \"\"\"\n",
    "# Please generate a question asking for the key information in the given paragraph.\n",
    "# Also answer the questions using the information in the given paragraph.\n",
    "# Please ask the specific question instead of the general question, like\n",
    "# 'What is the key information in the given paragraph?'.\n",
    "# Please generate the answer using as much information as possible.\n",
    "# If you are unable to answer it, please generate the answer as 'I don't know.'\n",
    "\n",
    "# Provide your answer as follows:\n",
    "\n",
    "# Output:::\n",
    "# Factoid question: (your factoid question)\n",
    "# Answer: (your answer to the factoid question)\n",
    "\n",
    "# Now here is the context.\n",
    "\n",
    "# Context: {context}\\n\n",
    "# Output:::\"\"\""
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Generate QA pairs\n",
    "\n",
    "import random\n",
    "\n",
    "N_GENERATIONS = 5  # generate only 5 QA couples here for cost and time considerations\n",
    "\n",
    "print(f\"Generating {N_GENERATIONS} QA couples...\")\n",
    "\n",
    "outputs = []\n",
    "for sampled_context in tqdm(random.choices(docs_processed, k=N_GENERATIONS)):\n",
    "    # Generate QA pairs\n",
    "    output_QA_couple = llm.invoke(QA_generation_prompt.format(context=sampled_context.page_content)).content\n",
    "    try:\n",
    "        question = output_QA_couple.split(\"Factoid question: \")[-1].split(\"Answer: \")[0]\n",
    "        answer = output_QA_couple.split(\"Answer: \")[-1]\n",
    "        assert len(answer) < 500, \"Answer is too long\"\n",
    "        outputs.append(\n",
    "            {\n",
    "                \"context\": sampled_context.page_content,\n",
    "                \"question\": question,\n",
    "                \"answer\": answer,\n",
    "                \"source_doc\": sampled_context.metadata[\"source\"],\n",
    "            }\n",
    "        )\n",
    "        time.sleep(3)  # sleep for llm rate limitation\n",
    "    except:\n",
    "        time.sleep(3)  # sleep for llm rate limitation\n",
    "        continue"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "reference_df = pd.DataFrame(outputs)\n",
    "display(reference_df.head(1))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# build a simple rag chain\n",
    "from langchain_huggingface import HuggingFaceEmbeddings\n",
    "from langchain.vectorstores import FAISS\n",
    "\n",
    "chunk_size=1024\n",
    "chunk_overlap=256\n",
    "splitter = RecursiveCharacterTextSplitter(\n",
    "    separators=[\"\\n\\n\", \"\\n\"], chunk_size=chunk_size, chunk_overlap=chunk_overlap\n",
    ")\n",
    "doc_chunks = splitter.split_documents(document)\n",
    "\n",
    "embeddings = HuggingFaceEmbeddings(model_name=\"all-mpnet-base-v2\")\n",
    "\n",
    "vectorstore = FAISS.from_documents(doc_chunks, embedding=embeddings)\n",
    "\n",
    "retriever = vectorstore.as_retriever()\n",
    "\n",
    "from langchain.chains import RetrievalQA\n",
    "\n",
    "rag_chain = RetrievalQA.from_llm(\n",
    "        llm=llm, retriever=retriever, return_source_documents=True\n",
    "    )"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Prepare evaluation data set\n",
    "def prepare_eval_dataset(reference_df, rag_chain):\n",
    "    \n",
    "    print(\"now loading evaluation dataset...\")\n",
    "    from datasets import Dataset\n",
    "    # Read reference file\n",
    "    df = reference_df\n",
    "\n",
    "    # Add anwsers from rag_chain\n",
    "    questions = df[\"question\"].values\n",
    "    ground_truth = []\n",
    "    for a in df[\"answer\"].values:\n",
    "        ground_truth.append(a)  # [a] for older version of ragas\n",
    "    answers = []\n",
    "    contexts = []\n",
    "\n",
    "    # Get anwswers from rag_chain\n",
    "    print(\"now getting anwsers from QA llm...\")\n",
    "    for query in questions:\n",
    "        results = rag_chain({\"query\": query})\n",
    "        answers.append(results[\"result\"])\n",
    "        contexts.append([docs.page_content for docs in results[\"source_documents\"]])\n",
    "        time.sleep(3)  # sleep for llm rate limitation\n",
    "\n",
    "    # To dict\n",
    "    data = {\n",
    "        \"question\": questions,\n",
    "        \"answer\": answers,\n",
    "        \"contexts\": contexts,\n",
    "        \"ground_truth\": ground_truth,\n",
    "    }\n",
    "\n",
    "    # Convert dict to dataset\n",
    "    dataset = Dataset.from_dict(data)\n",
    "    return dataset\n",
    "\n",
    "eval_dataset = prepare_eval_dataset(reference_df, rag_chain)\n",
    "eval_dataset\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Ragas evaluation\n",
    "from ragas.llms import LangchainLLMWrapper\n",
    "eval_llm = LangchainLLMWrapper(llm)\n",
    "\n",
    "from ragas import evaluate\n",
    "from ragas.metrics import (\n",
    "    faithfulness,\n",
    "    answer_relevancy,\n",
    "    context_recall,\n",
    "    context_precision,\n",
    "    answer_correctness,\n",
    ")\n",
    "result_eval_df = evaluate(\n",
    "    dataset=eval_dataset,\n",
    "    metrics=[\n",
    "        context_precision,\n",
    "        context_recall,\n",
    "        faithfulness,\n",
    "        answer_relevancy,\n",
    "        answer_correctness,\n",
    "    ],\n",
    "    llm=eval_llm, embeddings=embeddings,\n",
    "    raise_exceptions=False,\n",
    ")\n",
    "\n",
    "result_eval_df = result_eval_df.to_pandas()  # can take a while"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Check results\n",
    "result_eval_df\n",
    "# if you get NaN in results, check \"Frequently Asked Questions\" in Ragas for help"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "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": "3.12.1"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}