{
  "cells": [
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "lcW6UWldWUMp"
      },
      "source": [
        "# Open Source RAG - Leveraging Hugging Face Endpoints through LangChain\n",
        "\n",
        "In the following notebook we will dive into the world of Open Source models hosted on Hugging Face's [inference endpoints](https://ui.endpoints.huggingface.co/).\n",
        "\n",
        "The notebook will be broken into the following parts:\n",
        "\n",
        "- 🤝 Breakout Room #2:\n",
        "  1. Install required libraries\n",
        "  2. Set Environment Variables\n",
        "  3. Creating LangChain components powered by the endpoints\n",
        "  4. Creating a simple RAG pipeline with [LangChain v0.2.0](https://blog.langchain.dev/langchain-v02-leap-to-stability/)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "-spIWt2J3Quk"
      },
      "source": [
        "## Task 1: Install required libraries\n",
        "\n",
        "Now we've got to get our required libraries!\n",
        "\n",
        "We'll start with our `langchain` and `huggingface` dependencies.\n",
        "\n",
        "> You don't need to run this cell if you're running the notebook locally."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 95,
      "metadata": {
        "id": "EwGLnp31jXJj"
      },
      "outputs": [],
      "source": [
        "#!pip install -qU langchain-huggingface langchain-community faiss-cpu"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "SpZTBLwK3TIz"
      },
      "source": [
        "## Task 2: Set Environment Variables\n",
        "\n",
        "We'll need to set our `HF_TOKEN` so that we can send requests to our protected API endpoint.\n",
        "\n",
        "We'll also set-up our OpenAI API key, which we'll leverage later.\n",
        "\n"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 1,
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "NspG8I0XlFTt",
        "outputId": "edbf992c-97c0-46b1-9b69-40651a5e60d1"
      },
      "outputs": [],
      "source": [
        "import os\n",
        "import getpass\n",
        "\n",
        "os.environ[\"HF_TOKEN\"] = getpass.getpass(\"HuggingFace Write Token: \")"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "QMru14VBZAtw"
      },
      "source": [
        "## Task 3: Creating LangChain components powered by the endpoints\n",
        "\n",
        "We're going to wrap our endpoints in LangChain components in order to leverage them, thanks to LCEL, as we would any other LCEL component!"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "TGooehdzcmPb"
      },
      "source": [
        "### HuggingFaceEndpoint for LLM\n",
        "\n",
        "We can use the `HuggingFaceEndpoint` found [here](https://github.com/langchain-ai/langchain/blob/master/libs/community/langchain_community/llms/huggingface_endpoint.py) to power our chain - let's look at how we would implement it."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 2,
      "metadata": {
        "id": "N7u2Tu1FsURh"
      },
      "outputs": [],
      "source": [
        "YOUR_LLM_ENDPOINT_URL = \"https://dcrebqe18cydo729.us-east-1.aws.endpoints.huggingface.cloud\""
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "L3Cz6Mrnt2ku",
        "outputId": "f23f611f-5f08-4332-a74c-5b8d8311d185"
      },
      "outputs": [
        {
          "name": "stderr",
          "output_type": "stream",
          "text": [
            "/tmp/ipykernel_32142/1022838628.py:3: LangChainDeprecationWarning: The class `HuggingFaceEndpoint` was deprecated in LangChain 0.0.37 and will be removed in 1.0. An updated version of the class exists in the :class:`~langchain-huggingface package and should be used instead. To use it run `pip install -U :class:`~langchain-huggingface` and import as `from :class:`~langchain_huggingface import HuggingFaceEndpoint``.\n",
            "  hf_llm = HuggingFaceEndpoint(\n",
            "Note: Environment variable`HF_TOKEN` is set and is the current active token independently from the token you've just configured.\n"
          ]
        }
      ],
      "source": [
        "from langchain_community.llms import HuggingFaceEndpoint\n",
        "\n",
        "hf_llm = HuggingFaceEndpoint(\n",
        "    endpoint_url=f\"{YOUR_LLM_ENDPOINT_URL}\",\n",
        "    task=\"text-generation\",\n",
        "    max_new_tokens=512,\n",
        "    top_k=10,\n",
        "    top_p=0.95,\n",
        "    typical_p=0.95,\n",
        "    temperature=0.01,\n",
        "    repetition_penalty=1.03,\n",
        ")"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "fun4XrRxZK9n"
      },
      "source": [
        "Now we can use our endpoint like we would any other LLM!"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 4,
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 127
        },
        "id": "OFAbFT91Z8QV",
        "outputId": "588714ad-da28-4330-801b-7121b6f17ccf"
      },
      "outputs": [
        {
          "data": {
            "text/plain": [
              "\" I am doing well, thanks for asking. I have been busy with work and other things, but I always make time for my blog. Today, I want to talk about something that is very important to me: mental health.\\nAs you may know, I have been open about my struggles with anxiety and depression in the past. It's not something that I like to talk about, but it's something that I feel is important to share in order to help others who may be going through similar struggles.\\nMental health is just as important as physical health, and it's something that we should all prioritize. Unfortunately, there is still a stigma surrounding mental health, and many people struggle to speak openly about their feelings and experiences.\\nThat's why I want to use my platform to raise awareness and promote understanding of mental health. I believe that by talking openly and honestly about our struggles, we can break down the stigma and create a more supportive and compassionate community.\\nSo, today I want to talk about something that I think is really important: self-care. Self-care is not just about taking a relaxing bath or getting a massage (although those things are great too!). It's about taking care of your mind, body, and spirit in a way that nourishes and supports your overall well-being.\\nFor me, self-care looks like taking time to meditate and practice mindfulness, reading books and articles that inspire and educate me, and spending time in nature. It also means being kind to myself and acknowledging when I need to take a break or seek help.\\nI want to encourage you to think about what self-care means to you. What activities make you feel good? What things do you need to do to take care of your mind, body, and spirit?\\nRemember, self-care is not a luxury, it's a necessity. By prioritizing our own well-being, we can live happier, healthier, and more fulfilling lives.\\nThank you for reading, and I hope you'll join me in prioritizing self-care and mental health. Let's work together to create a more supportive and compassionate community!\\nWhat are some self-care activities that you enjoy? Share with me in the comments below!\\nI'm so glad you're talking about mental health and self-care! It's such an important topic and one that needs to be discussed more openly. I've been practicing self-care for a while now, and it's made a huge difference in my life. For me, self-care looks like exercise, meditation, and spending time with loved ones. I also love reading\""
            ]
          },
          "execution_count": 4,
          "metadata": {},
          "output_type": "execute_result"
        }
      ],
      "source": [
        "hf_llm.invoke(\"Hello, how are you?\")"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "ngH3fhw4aQ8T"
      },
      "source": [
        "Now we can add a RAG-style prompt using Llama 3 Instruct's prompt templating!"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 8,
      "metadata": {
        "id": "zdvv4JmkzEtj"
      },
      "outputs": [],
      "source": [
        "from langchain_core.prompts import PromptTemplate\n",
        "\n",
        "RAG_PROMPT_TEMPLATE = \"\"\"\\\n",
        "<|start_header_id|>system<|end_header_id|>\n",
        "You are a helpful assistant. You answer user questions based on provided context. If you can't answer the question with the provided context, say you don't know.<|eot_id|>\n",
        "\n",
        "<|start_header_id|>user<|end_header_id|>\n",
        "User Query:\n",
        "{query}\n",
        "\n",
        "Context:\n",
        "{context}<|eot_id|>\n",
        "\n",
        "<|start_header_id|>assistant<|end_header_id|>\n",
        "\"\"\"\n",
        "\n",
        "rag_prompt = PromptTemplate.from_template(RAG_PROMPT_TEMPLATE)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "Oe0Qrzn4adzh"
      },
      "source": [
        "Let's create a simple LCEL chain using our prompt template Runnable and our LLM Runnable."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 9,
      "metadata": {
        "id": "CE4djpxM0-Fg"
      },
      "outputs": [],
      "source": [
        "rag_chain = rag_prompt | hf_llm"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 10,
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 36
        },
        "id": "PNwrLXqDxHDY",
        "outputId": "f6803286-1aa5-488a-eea9-8bece68da7f5"
      },
      "outputs": [
        {
          "data": {
            "text/plain": [
              "'Carl is 40 years old.'"
            ]
          },
          "execution_count": 10,
          "metadata": {},
          "output_type": "execute_result"
        }
      ],
      "source": [
        "rag_chain.invoke({\"query\" : \"Who old is Carl?\", \"context\" : \"Carl is a sweet dude, he's 40.\"})"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "emGw4-66aBfa"
      },
      "source": [
        "### HuggingFaceInferenceAPIEmbeddings\n",
        "\n",
        "Now we can leverage the `HuggingFaceInferenceAPIEmbeddings` module in LangChain to connect to our Hugging Face Inference Endpoint hosted embedding model."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 11,
      "metadata": {
        "id": "n9Q7e4Gnwe_C"
      },
      "outputs": [],
      "source": [
        "from langchain_huggingface.embeddings import HuggingFaceEndpointEmbeddings\n",
        "\n",
        "YOUR_EMBED_MODEL_URL = \"https://c87ybgo18epgba6d.us-east-1.aws.endpoints.huggingface.cloud\"\n",
        "\n",
        "hf_embeddings = HuggingFaceEndpointEmbeddings(\n",
        "    model=YOUR_EMBED_MODEL_URL,\n",
        "    task=\"feature-extraction\",\n",
        "    huggingfacehub_api_token=os.environ[\"HF_TOKEN\"],\n",
        ")"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "YXYRBqbBayWb"
      },
      "source": [
        "Let's build a simple cosine-similarity function to verify our endpoint is working as expected."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 12,
      "metadata": {
        "id": "lOP6LKr74RG8"
      },
      "outputs": [],
      "source": [
        "import numpy as np\n",
        "from numpy.linalg import norm\n",
        "\n",
        "def cosine_similarity(phrase_1, phrase_2):\n",
        "  vec_1 = hf_embeddings.embed_documents([phrase_1])[0]\n",
        "  vec2_2 = hf_embeddings.embed_documents([phrase_2])[0]\n",
        "  return np.dot(vec_1, vec2_2) / (norm(vec_1) * norm(vec2_2))"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "uGZNhxF2bVIr"
      },
      "source": [
        "Let's try a few examples below!"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 13,
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "5o_cqEZ34f15",
        "outputId": "d3eb4933-8842-4278-fe48-2dc15e430b60"
      },
      "outputs": [
        {
          "data": {
            "text/plain": [
              "0.8903063446222079"
            ]
          },
          "execution_count": 13,
          "metadata": {},
          "output_type": "execute_result"
        }
      ],
      "source": [
        "cosine_similarity(\"I love my fluffy dog!\", \"I adore this furry puppy!\")"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 15,
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "R1nsAV1n4w4a",
        "outputId": "db53d783-4c87-404f-de67-fc1d01583e68"
      },
      "outputs": [
        {
          "data": {
            "text/plain": [
              "0.6667445107282148"
            ]
          },
          "execution_count": 15,
          "metadata": {},
          "output_type": "execute_result"
        }
      ],
      "source": [
        "cosine_similarity(\"I love my fluffy dog!\", \"Trekking across the arctic is tough work.\")"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "iiz6vKMlbbP4"
      },
      "source": [
        "## Task 4: Preparing Data!\n",
        "\n",
        "We'll start by loading some data from GitHub (Paul Graham's Essays) and then move to chunking them into manageable pieces!\n",
        "\n",
        "First - let's grab the repository where the files live."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 16,
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "AkuzZben5Eqp",
        "outputId": "eb8d39ae-fd70-4691-ddaa-1f8aa15f1c19"
      },
      "outputs": [
        {
          "name": "stdout",
          "output_type": "stream",
          "text": [
            "Cloning into 'paul-graham-to-kindle'...\n",
            "remote: Enumerating objects: 36, done.\u001b[K\n",
            "remote: Counting objects: 100% (36/36), done.\u001b[K\n",
            "remote: Compressing objects: 100% (33/33), done.\u001b[K\n",
            "remote: Total 36 (delta 3), reused 31 (delta 1), pack-reused 0 (from 0)\u001b[K\n",
            "Receiving objects: 100% (36/36), 2.35 MiB | 13.80 MiB/s, done.\n",
            "Resolving deltas: 100% (3/3), done.\n"
          ]
        }
      ],
      "source": [
        "!git clone https://github.com/dbredvick/paul-graham-to-kindle.git"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "8prMk6R0bsYd"
      },
      "source": [
        "Next - we can load them using LangChain!"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 17,
      "metadata": {
        "id": "K155zM7e53lt"
      },
      "outputs": [],
      "source": [
        "from langchain_community.document_loaders import TextLoader\n",
        "\n",
        "document_loader = TextLoader(\"./paul-graham-to-kindle/paul_graham_essays.txt\")\n",
        "documents = document_loader.load()"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "5wYfo6_0bwVc"
      },
      "source": [
        "Now, let's split them into 1000 character pieces."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 18,
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "w-Gx_0iL6Ikc",
        "outputId": "4cd1de4f-8a7d-4727-dc92-0ce3d321a82f"
      },
      "outputs": [
        {
          "data": {
            "text/plain": [
              "4265"
            ]
          },
          "execution_count": 18,
          "metadata": {},
          "output_type": "execute_result"
        }
      ],
      "source": [
        "from langchain_text_splitters import RecursiveCharacterTextSplitter\n",
        "\n",
        "text_splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=30)\n",
        "split_documents = text_splitter.split_documents(documents)\n",
        "len(split_documents)"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "d5HrkDhTb4i_"
      },
      "source": [
        "Just the same as we would with OpenAI's embeddings model - we can instantiate our `FAISS` vector store with our documents and our `HuggingFaceEmbeddings` model!\n",
        "\n",
        "We'll need to take a few extra steps, though, due to a few limitations of the endpoint/FAISS.\n",
        "\n",
        "We'll start by embeddings our documents in batches of `32`.\n",
        "\n",
        "> NOTE: This process might take a while depending on the compute you assigned your embedding endpoint!"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 20,
      "metadata": {
        "id": "ucghQgRp6YXr"
      },
      "outputs": [],
      "source": [
        "from langchain_community.vectorstores import FAISS\n",
        "\n",
        "for i in range(0, len(split_documents), 32):\n",
        "  if i == 0:\n",
        "    vectorstore = FAISS.from_documents(split_documents[i:i+32], hf_embeddings)\n",
        "    continue\n",
        "  vectorstore.add_documents(split_documents[i:i+32])"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "q07ZUp6Db_AO"
      },
      "source": [
        "Next, we set up FAISS as a retriever."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 21,
      "metadata": {
        "id": "fXr-yrAq7h8V"
      },
      "outputs": [],
      "source": [
        "hf_retriever = vectorstore.as_retriever()"
      ]
    },
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "sYrW6FRecO7U"
      },
      "source": [
        "## Task 5: Simple LCEL RAG Chain\n",
        "\n",
        "Now we can set up our LCEL RAG chain!\n",
        "\n",
        "> NOTE: We're not returning context for this example, and only returning the text output from the LLM."
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 22,
      "metadata": {
        "id": "ffIzIlct8ISb"
      },
      "outputs": [],
      "source": [
        "from operator import itemgetter\n",
        "from langchain.schema.output_parser import StrOutputParser\n",
        "from langchain.schema.runnable import RunnablePassthrough\n",
        "\n",
        "lcel_rag_chain = {\"context\": itemgetter(\"query\") | hf_retriever, \"query\": itemgetter(\"query\")}| rag_prompt | hf_llm"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": 23,
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 127
        },
        "id": "HOQfkEgb8nPH",
        "outputId": "92601728-d001-43e2-e543-e714d66f4f4e"
      },
      "outputs": [
        {
          "data": {
            "text/plain": [
              "\"According to the provided context, the best part of Silicon Valley is its proximity to a top-notch university, such as Stanford, which acts as a magnet to attract the best people from around the world. Additionally, the area's public transportation system, although not perfect, is considered good by American standards. However, the author suggests that the area could be improved by designing a town that prioritizes trains, bicycles, and pedestrian-friendly infrastructure over cars.\""
            ]
          },
          "execution_count": 23,
          "metadata": {},
          "output_type": "execute_result"
        }
      ],
      "source": [
        "lcel_rag_chain.invoke({\"query\" : \"What is the best part of Silicon Valley?\"})"
      ]
    }
  ],
  "metadata": {
    "colab": {
      "provenance": [],
      "toc_visible": true
    },
    "kernelspec": {
      "display_name": ".venv",
      "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": "3.9.21"
    }
  },
  "nbformat": 4,
  "nbformat_minor": 0
}