{
 "cells": [
  {
   "cell_type": "markdown",
   "id": "184878ec-80c1-4533-a7f8-f7fecbc3d8ad",
   "metadata": {},
   "source": [
    "# Build a RAG using a locally hosted NIM\n",
    "\n",
    "This notebook demonstrates how to build a RAG using NVIDIA NIM microservices. We locally host a Llama3-8b-instruct model using [NVIDIA NIM for LLMs](https://docs.nvidia.com/nim/large-language-models/latest/introduction.html) and connect to it using [LangChain NVIDIA AI Endpoints](https://python.langchain.com/docs/integrations/chat/nvidia_ai_endpoints/) package.\n",
    "\n",
    "We then create a vector store by downloading web pages and generating their embeddings using FAISS. We then showcase two different chat chains for querying the vector store. For this example, we use the NVIDIA Triton documentation website, though the code can be easily modified to use any other source. For the embedding model, we use [the GPU accelerated NV-Embed-QA model from NVIDIA API Catalog](https://build.nvidia.com/nvidia/embed-qa-4).\n",
    "\n",
    "### First stage is to load NVIDIA Triton documentation from the web, chunkify the data, and generate embeddings using FAISS\n",
    "\n",
    "To get started:\n",
    "\n",
    "1. Generate an [NGC CLI API key](https://org.ngc.nvidia.com/setup/personal-keys). This key will need to be passed to docker run in the next section as the NGC_API_KEY environment variable to download the appropriate models and resources when starting the NIM.\n",
    "\n",
    "2. Download and install the NGC CLI following the [NGC Setup steps](https://docs.ngc.nvidia.com/cli/index.html?_gl=1*22f68y*_gcl_au*MTE2NTMwMTA2NC4xNzE1NzY4NzE4). Follow the steps on that page to set the NGC CLI and docker client configs appropriately."
   ]
  },
  {
   "cell_type": "markdown",
   "id": "225c7185",
   "metadata": {},
   "source": [
    "Note: In order to run this notebook, you need to launch the NIM Docker container in the terminal outside of the web browser notebook environment. Run the commands in the first 3 cells from a terminal then begin with the 4th cell (curl inference command) within the notebook environment (web browser)."
   ]
  },
  {
   "cell_type": "markdown",
   "id": "39ef96bf",
   "metadata": {},
   "source": [
    "To pull the NIM container image from NGC, first authenticate with the NVIDIA Container Registry with the following command from your terminal."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "1308d67c",
   "metadata": {},
   "outputs": [],
   "source": [
    "!export NGC_API_KEY=\"Provide your api key here\"\n",
    "!docker login nvcr.io --username '$oauthtoken' --password \"${NGC_API_KEY}\""
   ]
  },
  {
   "cell_type": "markdown",
   "id": "9c2403b8",
   "metadata": {},
   "source": [
    "Set up location for caching the model artifacts. Export the following env variables from your terminal."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "06a23a6d",
   "metadata": {},
   "outputs": [],
   "source": [
    "!export LOCAL_NIM_CACHE=~/.cache/nim\n",
    "!mkdir -p \"$LOCAL_NIM_CACHE\"\n",
    "!chmod 777 \"$LOCAL_NIM_CACHE\""
   ]
  },
  {
   "cell_type": "markdown",
   "id": "19a7e489",
   "metadata": {},
   "source": [
    "Launch the NIM LLM microservice by executing this command from the terminal where you have exported all the environment variables."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "88f612dd",
   "metadata": {},
   "outputs": [],
   "source": [
    "!docker run -d --name meta-llama3-8b-instruct --gpus all -e NGC_API_KEY -v \"$LOCAL_NIM_CACHE:/opt/nim/.cache\" -u $(id -u) -p 8000:8000 nvcr.io/nim/meta/llama3-8b-instruct:1.0.0"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "37e2fcda",
   "metadata": {},
   "source": [
    "Before we continue and connect the NIM to LangChain, let's test it using a simple OpenAI completion request. You can execute this command and all the subsequent one after this from your web browser."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "af3bf04b",
   "metadata": {},
   "outputs": [],
   "source": [
    "!curl -X 'POST' \\\n",
    "    \"http://0.0.0.0:8000/v1/completions\" \\\n",
    "    -H \"accept: application/json\" \\\n",
    "    -H \"Content-Type: application/json\" \\\n",
    "    -d '{\"model\": \"meta/llama3-8b-instruct\", \"prompt\": \"Once upon a time\", \"max_tokens\": 64}'"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "79c2aa6f",
   "metadata": {},
   "source": [
    "Now setup the LangChain flow by installing prerequisite libraries"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "5966ea5a",
   "metadata": {},
   "outputs": [],
   "source": [
    "!pip install --upgrade pip\n",
    "!pip install langchain==0.2.5\n",
    "!pip install langchain-nvidia-ai-endpoints==0.1.2\n",
    "!pip install faiss-gpu==1.7.2 # replace with faiss-cpu if you don't have a gpu"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "74b0c989",
   "metadata": {},
   "source": [
    "Set up NVIDIA API key, which you can get from the [API Catalog](https://build.nvidia.com/). This key will be used to communicate with GPU accelerated cloud hosted embedding model."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "79771ce9",
   "metadata": {},
   "outputs": [],
   "source": [
    "import getpass\n",
    "import os\n",
    "\n",
    "if not os.environ.get(\"NVIDIA_API_KEY\", \"\").startswith(\"nvapi-\"):\n",
    "    nvapi_key = getpass.getpass(\"Enter your NVIDIA API key: \")\n",
    "    assert nvapi_key.startswith(\"nvapi-\"), f\"{nvapi_key[:5]}... is not a valid key\"\n",
    "    os.environ[\"NVIDIA_API_KEY\"] = nvapi_key"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "5584e3b1",
   "metadata": {},
   "source": [
    "We can now connect with the deployed NIM LLM model in LangChain by specifying the base URL"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "35baa8c6",
   "metadata": {},
   "outputs": [],
   "source": [
    "from langchain_nvidia_ai_endpoints import ChatNVIDIA\n",
    "\n",
    "llm = ChatNVIDIA(base_url=\"http://0.0.0.0:8000/v1\", model=\"meta/llama3-8b-instruct\", temperature=0.1, max_tokens=1000, top_p=1.0)\n",
    "\n",
    "result = llm.invoke(\"What is the capital of France?\")\n",
    "print(result.content)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "f95b2753",
   "metadata": {},
   "source": [
    "Import all the required libraries for building the langchain agent."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "42bf2619-1ca3-4477-82b8-88c240dd87ad",
   "metadata": {},
   "outputs": [],
   "source": [
    "import os\n",
    "from langchain.chains import ConversationalRetrievalChain, LLMChain\n",
    "from langchain.chains.conversational_retrieval.prompts import CONDENSE_QUESTION_PROMPT, QA_PROMPT\n",
    "from langchain.chains.question_answering import load_qa_chain\n",
    "from langchain.memory import ConversationBufferMemory\n",
    "from langchain.vectorstores import FAISS\n",
    "from langchain.text_splitter import RecursiveCharacterTextSplitter\n",
    "from langchain_nvidia_ai_endpoints import ChatNVIDIA\n",
    "from langchain_nvidia_ai_endpoints import NVIDIAEmbeddings"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "id": "eb7f8822",
   "metadata": {},
   "source": [
    "Helper functions for loading html files, which we'll use to generate the embeddings. We'll use this later to load the relevant html documents from the Triton documentation website and convert to a vector store."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "f8097819",
   "metadata": {},
   "outputs": [],
   "source": [
    "import re\n",
    "from typing import List, Union\n",
    "\n",
    "import requests\n",
    "from bs4 import BeautifulSoup\n",
    "\n",
    "def html_document_loader(url: Union[str, bytes]) -> str:\n",
    "    \"\"\"\n",
    "    Loads the HTML content of a document from a given URL and return it's content.\n",
    "\n",
    "    Args:\n",
    "        url: The URL of the document.\n",
    "\n",
    "    Returns:\n",
    "        The content of the document.\n",
    "\n",
    "    Raises:\n",
    "        Exception: If there is an error while making the HTTP request.\n",
    "\n",
    "    \"\"\"\n",
    "    try:\n",
    "        response = requests.get(url)\n",
    "        html_content = response.text\n",
    "    except Exception as e:\n",
    "        print(f\"Failed to load {url} due to exception {e}\")\n",
    "        return \"\"\n",
    "\n",
    "    try:\n",
    "        # Create a Beautiful Soup object to parse html\n",
    "        soup = BeautifulSoup(html_content, \"html.parser\")\n",
    "\n",
    "        # Remove script and style tags\n",
    "        for script in soup([\"script\", \"style\"]):\n",
    "            script.extract()\n",
    "\n",
    "        # Get the plain text from the HTML document\n",
    "        text = soup.get_text()\n",
    "\n",
    "        # Remove excess whitespace and newlines\n",
    "        text = re.sub(\"\\s+\", \" \", text).strip()\n",
    "\n",
    "        return text\n",
    "    except Exception as e:\n",
    "        print(f\"Exception {e} while loading document\")\n",
    "        return \"\""
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "id": "b6a2d1e0",
   "metadata": {},
   "source": [
    "Read html files and split text in preparation for embedding generation\n",
    "Note chunk_size value must match the specific LLM used for embedding genetation\n",
    "\n",
    "Make sure to pay attention to the chunk_size parameter in TextSplitter. Setting the right chunk size is critical for RAG performance, as much of a RAG’s success is based on the retrieval step finding the right context for generation. The entire prompt (retrieved chunks + user query) must fit within the LLM’s context window. Therefore, you should not specify chunk sizes too big, and balance them out with the estimated query size. For example, while OpenAI LLMs have a context window of 8k-32k tokens, Llama3 is limited to 8k tokens. Experiment with different chunk sizes, but typical values should be 100-600, depending on the LLM."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "56aa8900",
   "metadata": {},
   "outputs": [],
   "source": [
    "def create_embeddings(embedding_path: str = \"./data/nv_embedding\"):\n",
    "\n",
    "    embedding_path = \"./data/nv_embedding\"\n",
    "    print(f\"Storing embeddings to {embedding_path}\")\n",
    "\n",
    "    # List of web pages containing NVIDIA Triton technical documentation\n",
    "    urls = [\n",
    "         \"https://docs.nvidia.com/deeplearning/triton-inference-server/user-guide/docs/index.html\",\n",
    "         \"https://docs.nvidia.com/deeplearning/triton-inference-server/user-guide/docs/getting_started/quickstart.html\",\n",
    "         \"https://docs.nvidia.com/deeplearning/triton-inference-server/user-guide/docs/user_guide/model_repository.html\",\n",
    "         \"https://docs.nvidia.com/deeplearning/triton-inference-server/user-guide/docs/user_guide/model_analyzer.html\",\n",
    "         \"https://docs.nvidia.com/deeplearning/triton-inference-server/user-guide/docs/user_guide/architecture.html\",\n",
    "    ]\n",
    "\n",
    "    documents = []\n",
    "    for url in urls:\n",
    "        document = html_document_loader(url)\n",
    "        documents.append(document)\n",
    "\n",
    "\n",
    "    text_splitter = RecursiveCharacterTextSplitter(\n",
    "        chunk_size=1000,\n",
    "        chunk_overlap=0,\n",
    "        length_function=len,\n",
    "    )\n",
    "    texts = text_splitter.create_documents(documents)\n",
    "    index_docs(url, text_splitter, texts, embedding_path)\n",
    "    print(\"Generated embedding successfully\")"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "id": "4d4e2097",
   "metadata": {},
   "source": [
    "Generate embeddings using NVIDIA Retrieval QA Embedding NIM and NVIDIA AI Endpoints for LangChain and save embeddings to offline vector store in the /embed directory for future re-use"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "fc7e6a93",
   "metadata": {},
   "outputs": [],
   "source": [
    "def index_docs(url: Union[str, bytes], splitter, documents: List[str], dest_embed_dir) -> None:\n",
    "    \"\"\"\n",
    "    Split the document into chunks and create embeddings for the document\n",
    "\n",
    "    Args:\n",
    "        url: Source url for the document.\n",
    "        splitter: Splitter used to split the document\n",
    "        documents: list of documents whose embeddings needs to be created\n",
    "        dest_embed_dir: destination directory for embeddings\n",
    "\n",
    "    Returns:\n",
    "        None\n",
    "    \"\"\"\n",
    "    embeddings = NVIDIAEmbeddings(model=\"NV-Embed-QA\", truncate=\"END\")\n",
    "\n",
    "    for document in documents:\n",
    "        texts = splitter.split_text(document.page_content)\n",
    "\n",
    "        # metadata to attach to document\n",
    "        metadatas = [document.metadata]\n",
    "\n",
    "        # create embeddings and add to vector store\n",
    "        if os.path.exists(dest_embed_dir):\n",
    "            update = FAISS.load_local(folder_path=dest_embed_dir, embeddings=embeddings, allow_dangerous_deserialization=True)\n",
    "            update.add_texts(texts, metadatas=metadatas)\n",
    "            update.save_local(folder_path=dest_embed_dir)\n",
    "        else:\n",
    "            docsearch = FAISS.from_texts(texts, embedding=embeddings, metadatas=metadatas)\n",
    "            docsearch.save_local(folder_path=dest_embed_dir)"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "id": "44650d71",
   "metadata": {},
   "source": [
    "### Second stage is to load the embeddings from the vector store and build a RAG using NVIDIAEmbeddings\n",
    "\n",
    "Create the embeddings model using NVIDIA Retrieval QA Embedding NIM from the API Catalog. This model represents words, phrases, or other entities as vectors of numbers and understands the relation between words and phrases. See here for reference: https://build.nvidia.com/nvidia/embed-qa-4"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "10db1c5c-f515-460f-bf23-5d68f195e52b",
   "metadata": {},
   "outputs": [],
   "source": [
    "\n",
    "\n",
    "create_embeddings()\n",
    "\n",
    "embedding_model = NVIDIAEmbeddings(model=\"NV-Embed-QA\", truncate=\"END\")\n"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "73f9f5e2",
   "metadata": {},
   "source": [
    "Load documents from vector database using FAISS"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "d55bb79e-5bb6-409d-8ead-3c3006aeb2ed",
   "metadata": {},
   "outputs": [],
   "source": [
    "# Embed documents\n",
    "embedding_path = \"./data/nv_embedding\"\n",
    "docsearch = FAISS.load_local(folder_path=embedding_path, embeddings=embedding_model, allow_dangerous_deserialization=True)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "7614e948-aab6-40d5-bf14-4f8ba99b1329",
   "metadata": {},
   "source": [
    "Create a ConversationalRetrievalChain chain using a local NIM. We'll use the Llama3 8B NIM we created and deployed locally, add memory for chat history, and connect to the vector store via the embedding model. See here for reference: https://python.langchain.com/docs/modules/chains/popular/chat_vector_db#conversationalretrievalchain-with-streaming-to-stdout"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "94b49b4a",
   "metadata": {},
   "outputs": [],
   "source": [
    "llm = ChatNVIDIA(base_url=\"http://0.0.0.0:8000/v1\", model=\"meta/llama3-8b-instruct\", temperature=0.1, max_tokens=1000, top_p=1.0)\n",
    "\n",
    "memory = ConversationBufferMemory(memory_key=\"chat_history\", return_messages=True)\n",
    "\n",
    "qa_prompt=QA_PROMPT\n",
    "\n",
    "doc_chain = load_qa_chain(llm, chain_type=\"stuff\", prompt=QA_PROMPT)\n",
    "\n",
    "qa = ConversationalRetrievalChain.from_llm(\n",
    "    llm=llm,\n",
    "    retriever=docsearch.as_retriever(),\n",
    "    chain_type=\"stuff\",\n",
    "    memory=memory,\n",
    "    combine_docs_chain_kwargs={'prompt': qa_prompt},\n",
    ")"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "f60f2240",
   "metadata": {},
   "source": [
    "Now try asking a question about Triton with the simpler chain. Compare the answer to the result with previous complex chain model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "9add6e2e",
   "metadata": {},
   "outputs": [],
   "source": [
    "query = \"What is Triton?\"\n",
    "result = qa({\"question\": query})\n",
    "print(result.get(\"answer\"))"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "43b1cddd",
   "metadata": {},
   "source": [
    "Ask another question about Triton"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "62654a9a",
   "metadata": {},
   "outputs": [],
   "source": [
    "query = \"Does Triton support ONNX?\"\n",
    "result = qa({\"question\": query})\n",
    "print(result.get(\"answer\"))"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "f178ac86",
   "metadata": {},
   "source": [
    "Finally showcase chat capabilites by asking a question about the previous query"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "781e058e",
   "metadata": {},
   "outputs": [],
   "source": [
    "query = \"But why?\"\n",
    "result = qa({\"question\": query})\n",
    "print(result.get(\"answer\"))"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3 (ipykernel)",
   "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": "3.10.13"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}