{ "cells": [ { "cell_type": "markdown", "id": "dfe37963-1af6-44fc-a841-8e462443f5e6", "metadata": {}, "source": [ "## Expert Knowledge Worker\n", "\n", "Features:\n", "- A question answering agent that is an expert knowledge worker\n", "- To be used by employees of Insurellm, an Insurance Tech company\n", "- The agent needs to be accurate and the solution should be low cost.\n", "\n", "This project will use RAG (Retrieval Augmented Generation) to ensure our question/answering assistant has high accuracy.\n", "\n", "Technology:\n", "- RAG: LangChain\n", "- Embedding model: OpenAIEmbeddings or HuggingFace sentence-transformers\n", "- Encoding method: Auto-encoding\n", "- Vector datastore: Chroma or FAISS\n", "- Vector DB visualization: Plotly\n", "- Dimensionality reduction technique: t-SNE\n", "\n", "# Dependencies" ] }, { "cell_type": "code", "execution_count": null, "id": "802137aa-8a74-45e0-a487-d1974927d7ca", "metadata": {}, "outputs": [], "source": [ "# imports\n", "\n", "import os\n", "import glob\n", "from dotenv import load_dotenv\n", "import gradio as gr\n", "from langchain.document_loaders import DirectoryLoader, TextLoader\n", "from langchain.text_splitter import RecursiveCharacterTextSplitter\n", "from langchain.schema import Document\n", "from langchain_openai import OpenAIEmbeddings, ChatOpenAI\n", "from langchain.embeddings import HuggingFaceEmbeddings\n", "from langchain_chroma import Chroma\n", "from langchain.vectorstores import FAISS\n", "import numpy as np\n", "from sklearn.manifold import TSNE\n", "import plotly.graph_objects as go\n", "from langchain.memory import ConversationBufferMemory\n", "from langchain.chains import ConversationalRetrievalChain" ] }, { "cell_type": "markdown", "id": "7187c181-5b17-4df7-b298-b7cb2b6d09f7", "metadata": {}, "source": [ "# Setup" ] }, { "cell_type": "code", "execution_count": null, "id": "58c85082-e417-4708-9efe-81a5d55d1424", "metadata": {}, "outputs": [], "source": [ "MODEL = \"gpt-4o-mini\"\n", "db_name = \"vector_db\"\n", "db_type = \"Chroma\"\n", "# db_type = \"FAISS\"\n", "embed_type = \"OpenAIEmbeddings\"\n", "# embed_type = \"sentence-transformers\"" ] }, { "cell_type": "code", "execution_count": null, "id": "ee78efcb-60fe-449e-a944-40bab26261af", "metadata": {}, "outputs": [], "source": [ "# Load environment variables\n", "\n", "load_dotenv(override=True)\n", "os.environ['OPENAI_API_KEY'] = os.getenv('OPENAI_API_KEY')" ] }, { "cell_type": "markdown", "id": "a2f0866b-5cfb-4ecd-87d1-6da872887dcd", "metadata": {}, "source": [ "# Create Knowledge Base for RAG\n", "\n", "## Load Company Documents\n", "\n", "Uses LangChain to read in a Knowledge Base of documents and to divide up documents into overlaping chunks." ] }, { "cell_type": "code", "execution_count": null, "id": "730711a9-6ffe-4eee-8f48-d6cfb7314905", "metadata": {}, "outputs": [], "source": [ "# Read in documents using LangChain's loaders\n", "# Take everything in all the sub-folders of our knowledgebase\n", "\n", "folders = glob.glob(\"../knowledge-base/*\")\n", "text_loader_kwargs = {'encoding': 'utf-8'}\n", "# text_loader_kwargs={'autodetect_encoding': True}\n", "\n", "documents = []\n", "for folder in folders:\n", " doc_type = os.path.basename(folder)\n", " loader = DirectoryLoader(folder, glob=\"**/*.md\", loader_cls=TextLoader, loader_kwargs=text_loader_kwargs)\n", " folder_docs = loader.load()\n", " for doc in folder_docs:\n", " doc.metadata[\"doc_type\"] = doc_type\n", " documents.append(doc)" ] }, { "cell_type": "code", "execution_count": null, "id": "7310c9c8-03c1-4efc-a104-5e89aec6db1a", "metadata": {}, "outputs": [], "source": [ "text_splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=200)\n", "chunks = text_splitter.split_documents(documents)" ] }, { "cell_type": "code", "execution_count": null, "id": "cd06e02f-6d9b-44cc-a43d-e1faa8acc7bb", "metadata": {}, "outputs": [], "source": [ "len(chunks)" ] }, { "cell_type": "code", "execution_count": null, "id": "2c54b4b6-06da-463d-bee7-4dd456c2b887", "metadata": {}, "outputs": [], "source": [ "doc_types = set(chunk.metadata['doc_type'] for chunk in chunks)\n", "print(f\"Document types found: {', '.join(doc_types)}\")" ] }, { "cell_type": "markdown", "id": "77f7d2a6-ccfa-425b-a1c3-5e55b23bd013", "metadata": {}, "source": [ "## Vector Embeddings\n", "\n", "Convert chunks of text into Vectors using OpenAIEmbeddings and store the Vectors in Chroma (or FAISS)." ] }, { "cell_type": "code", "execution_count": null, "id": "78998399-ac17-4e28-b15f-0b5f51e6ee23", "metadata": {}, "outputs": [], "source": [ "# Put the chunks of data into a Vector Store that associates a Vector Embedding with each chunk\n", "\n", "embeddings = None\n", "# OpenAIEmbeddings is OpenAI's vector embedding models\n", "if embed_type == \"OpenAIEmbeddings\":\n", " embeddings = OpenAIEmbeddings()\n", "\n", "# sentence-transformers is a free Vector embeddings model from HuggingFace\n", "elif embed_type == \"sentence-transformers\":\n", " embeddings = HuggingFaceEmbeddings(model_name=\"sentence-transformers/all-MiniLM-L6-v2\")\n", "\n", "if embeddings is None:\n", " print(\"ERROR: embeddings not set. Check embed_type is set to a valid model\")" ] }, { "cell_type": "markdown", "id": "64768521-a775-472c-83c5-0c0d715d44ac", "metadata": {}, "source": [ "## Create Vector Datastore" ] }, { "cell_type": "code", "execution_count": null, "id": "057868f6-51a6-4087-94d1-380145821550", "metadata": {}, "outputs": [], "source": [ "# Create vectorstore\n", "vectorstore = None\n", "\n", "# Chroma is a popular open source Vector Database based on SQLLite\n", "if db_type == \"Chroma\":\n", " # Delete vector DB if already exists\n", " if os.path.exists(db_name):\n", " Chroma(persist_directory=db_name, embedding_function=embeddings).delete_collection()\n", " \n", " # Create vectorstore\n", " vectorstore = Chroma.from_documents(documents=chunks, embedding=embeddings, persist_directory=db_name)\n", " \n", " print(f\"Vectorstore created with {vectorstore._collection.count()} documents\")\n", "\n", " # Get one vector and find how many dimensions it has\n", " collection = vectorstore._collection\n", " sample_embedding = collection.get(limit=1, include=[\"embeddings\"])[\"embeddings\"][0]\n", " dimensions = len(sample_embedding)\n", " print(f\"The vectors have {dimensions:,} dimensions\")\n", " \n", "# FAISS is an in-memory vector DB from Facebook\n", "elif db_type == \"FAISS\":\n", " # Create vectorstore\n", " vectorstore = FAISS.from_documents(chunks, embedding=embeddings)\n", " \n", " total_vectors = vectorstore.index.ntotal\n", " dimensions = vectorstore.index.d\n", " print(f\"There are {total_vectors} vectors with {dimensions:,} dimensions in the vector store\")\n", "\n", "else:\n", " print(\"ERROR: Vector datastore not created. Check db_type is set to a valid database\")" ] }, { "cell_type": "markdown", "id": "b0d45462-a818-441c-b010-b85b32bcf618", "metadata": {}, "source": [ "# Visualizing the Vector Store\n", "\n", "Humans are not very good at visualizing things with more than 3 dimensions so to visualize a vector datastore with thousands of dimesions. We need to use techniques like projecting down to reduce the dimensions to only 2 or 3 dimensions in a way that does the best possible job at separating things out to stay faithful to the multi-dimensional representation.\n", "\n", "For example, things that are far apart in these multiple dimensions will still be far apart even when projected down to 2 dimensions.\n", "\n", "[t-distributed stochastic neighbor embedding (t-SNE)](https://en.wikipedia.org/wiki/T-distributed_stochastic_neighbor_embedding) is a nonlinear dimensionality reduction technique for embedding high-dimensional data for visualization in a low-dimensional space of two or three dimensions.\n", "\n", "## Configure Visualization" ] }, { "cell_type": "code", "execution_count": null, "id": "b98adf5e-d464-4bd2-9bdf-bc5b6770263b", "metadata": {}, "outputs": [], "source": [ "# Prework\n", "if db_type == \"Chroma\":\n", " result = collection.get(include=['embeddings', 'documents', 'metadatas'])\n", " vectors = np.array(result['embeddings'])\n", " documents = result['documents']\n", " doc_types = [metadata['doc_type'] for metadata in result['metadatas']]\n", " colors = [['blue', 'green', 'red', 'orange'][['products', 'employees', 'contracts', 'company'].index(t)] for t in doc_types]\n", "\n", "elif db_type == \"FAISS\":\n", " vectors = []\n", " documents = []\n", " doc_types = []\n", " colors = []\n", " color_map = {'products':'blue', 'employees':'green', 'contracts':'red', 'company':'orange'}\n", " \n", " for i in range(total_vectors):\n", " vectors.append(vectorstore.index.reconstruct(i))\n", " doc_id = vectorstore.index_to_docstore_id[i]\n", " document = vectorstore.docstore.search(doc_id)\n", " documents.append(document.page_content)\n", " doc_type = document.metadata['doc_type']\n", " doc_types.append(doc_type)\n", " colors.append(color_map[doc_type])\n", " \n", " vectors = np.array(vectors)\n", "\n", "else:\n", " print(\"ERROR: Vector datastore not created. Check db_type is set to a valid database\")" ] }, { "cell_type": "markdown", "id": "bb279701-0086-44aa-a2da-14341aecf529", "metadata": {}, "source": [ "## Reduce the dimensionality to 2D" ] }, { "cell_type": "code", "execution_count": null, "id": "427149d5-e5d8-4abd-bb6f-7ef0333cca21", "metadata": {}, "outputs": [], "source": [ "# We humans find it easier to visalize things in 2D!\n", "# Reduce the dimensionality of the vectors to 2D using t-SNE\n", "# (t-distributed stochastic neighbor embedding)\n", "\n", "tsne = TSNE(n_components=2, random_state=42)\n", "reduced_vectors = tsne.fit_transform(vectors)\n", "\n", "# Create the 2D scatter plot\n", "fig = go.Figure(data=[go.Scatter(\n", " x=reduced_vectors[:, 0],\n", " y=reduced_vectors[:, 1],\n", " mode='markers',\n", " marker=dict(size=5, color=colors, opacity=0.8),\n", " text=[f\"Type: {t}
Text: {d[:100]}...\" for t, d in zip(doc_types, documents)],\n", " hoverinfo='text'\n", ")])\n", "\n", "fig.update_layout(\n", " title=f'2D {db_type} Vector Store Visualization',\n", " scene=dict(xaxis_title='x',yaxis_title='y'),\n", " width=800,\n", " height=600,\n", " margin=dict(r=20, b=10, l=10, t=40)\n", ")\n", "\n", "fig.show()" ] }, { "cell_type": "markdown", "id": "e2b724f3-e3ad-4d42-bfa4-a89386d6414e", "metadata": {}, "source": [ "## Reduce the dimensionality to 3D" ] }, { "cell_type": "code", "execution_count": null, "id": "e1418e88-acd5-460a-bf2b-4e6efc88e3dd", "metadata": {}, "outputs": [], "source": [ "# 3D representation isn't as easy to navigate\n", "\n", "tsne = TSNE(n_components=3, random_state=42)\n", "reduced_vectors = tsne.fit_transform(vectors)\n", "\n", "# Create the 3D scatter plot\n", "fig = go.Figure(data=[go.Scatter3d(\n", " x=reduced_vectors[:, 0],\n", " y=reduced_vectors[:, 1],\n", " z=reduced_vectors[:, 2],\n", " mode='markers',\n", " marker=dict(size=5, color=colors, opacity=0.8),\n", " text=[f\"Type: {t}
Text: {d[:100]}...\" for t, d in zip(doc_types, documents)],\n", " hoverinfo='text'\n", ")])\n", "\n", "fig.update_layout(\n", " title=f'3D {db_type} Vector Store Visualization',\n", " scene=dict(xaxis_title='x', yaxis_title='y', zaxis_title='z'),\n", " width=900,\n", " height=700,\n", " margin=dict(r=20, b=10, l=10, t=40)\n", ")\n", "\n", "fig.show()" ] }, { "cell_type": "markdown", "id": "9468860b-86a2-41df-af01-b2400cc985be", "metadata": {}, "source": [ "# Expert Knowledge Worker\n", "\n", "Use LangChain to bring it all together by creating a conversation chain with RAG and memory.\n", "\n", "Key abstractions in LangChain:\n", "- LLM: represents abstraction around a model\n", "- Retriever: interface onto somthing like a vector store used for RAG retrieval\n", "- Memory: represents a history of a conversation with a chatbot in memory\n", "\n", "Because LangChain abstracts the reprentation of the LLM, retriever and memory the code is the same for any model and knowledge base.\n", "\n", "Note: ok to ignore _Deprecation Warning_ for now; LangChain are not expected to remove ConversationBufferMemory any time soon.\n", "\n", "## Setup" ] }, { "cell_type": "code", "execution_count": null, "id": "129c7d1e-0094-4479-9459-f9360b95f244", "metadata": {}, "outputs": [], "source": [ "# create a new Chat with OpenAI\n", "llm = ChatOpenAI(temperature=0.7, model_name=MODEL)\n", "\n", "# set up the conversation memory for the chat\n", "memory = ConversationBufferMemory(memory_key='chat_history', return_messages=True)\n", "\n", "# the retriever is an abstraction over the VectorStore that will be used during RAG\n", "retriever = vectorstore.as_retriever()\n", "\n", "# putting it together: set up the conversation chain with the GPT 4o-mini LLM, the vector store and memory\n", "conversation_chain = ConversationalRetrievalChain.from_llm(llm=llm, retriever=retriever, memory=memory)" ] }, { "cell_type": "code", "execution_count": null, "id": "968e7bf2-e862-4679-a11f-6c1efb6ec8ca", "metadata": {}, "outputs": [], "source": [ "query = \"Can you describe Insurellm in a few sentences\"\n", "result = conversation_chain.invoke({\"question\":query})\n", "print(result[\"answer\"])" ] }, { "cell_type": "markdown", "id": "990a2917-562c-461a-8ce9-a8ad8ad1646d", "metadata": {}, "source": [ "## Clear Memory\n", "\n", "Clear the memory from the testing and restart conversation chain for UI." ] }, { "cell_type": "code", "execution_count": null, "id": "e6eb99fb-33ec-4025-ab92-b634ede03647", "metadata": {}, "outputs": [], "source": [ "# clear the memory and restart conversation chain for UI\n", "memory = ConversationBufferMemory(memory_key='chat_history', return_messages=True)\n", "conversation_chain = ConversationalRetrievalChain.from_llm(llm=llm, retriever=retriever, memory=memory)" ] }, { "cell_type": "markdown", "id": "bbbcb659-13ce-47ab-8a5e-01b930494964", "metadata": {}, "source": [ "## Functions" ] }, { "cell_type": "code", "execution_count": null, "id": "c3536590-85c7-4155-bd87-ae78a1467670", "metadata": {}, "outputs": [], "source": [ "# Wrapping in a function - note that history isn't used, as the memory is in the conversation_chain\n", "\n", "def chat(message, history):\n", " result = conversation_chain.invoke({\"question\": message})\n", " return result[\"answer\"]" ] }, { "cell_type": "markdown", "id": "b655d3da-277b-45a9-8113-747314ec0889", "metadata": {}, "source": [ "## UI" ] }, { "cell_type": "code", "execution_count": null, "id": "b252d8c1-61a8-406d-b57a-8f708a62b014", "metadata": {}, "outputs": [], "source": [ "# And in Gradio:\n", "\n", "view = gr.ChatInterface(chat, type=\"messages\", examples=[\"what is insurellm?\",\"what did avery do before?\", \"does insurellm offer any products in the auto industry space?\"], title=\"Insurellm Expert Knowledge Worker\").launch(inbrowser=True)" ] }, { "cell_type": "code", "execution_count": null, "id": "5435b2b9-935c-48cd-aaf3-73a837ecde49", "metadata": {}, "outputs": [], "source": [] } ], "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": "3.11.12" } }, "nbformat": 4, "nbformat_minor": 5 }