{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "%load_ext autoreload\n",
    "%autoreload 2"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [],
   "source": [
    "from graph import create_graph\n",
    "\n",
    "graph = create_graph()\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/jpeg": "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",
      "text/plain": [
       "<IPython.core.display.Image object>"
      ]
     },
     "execution_count": 9,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "from IPython.display import Image\n",
    "\n",
    "Image(graph.get_graph().draw_mermaid_png())"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Checking memory or database...\n",
      "{'question': 'What did I ask you about?', 'generation': [HumanMessage(content=\"The total value of 'Cash and cash equivalents' reported on December 31, 2023, was $6,874 million.\", id='1b372d25-12a9-45e8-a813-0dcd3513c4b0')], 'wiki_search': True, 'documents': [Document(page_content='Table of Contents\\nAirbnb, Inc.\\nNotes to Condensed Consolidated Financial Statements (unaudited)\\nNote 3. Supplemental Financial Statement Information\\nCash, Cash Equivalents, and Restricted Cash\\nThe following table reconciles cash, cash equivalents, and restricted cash reported on the Company’s unaudited condensed consolidated balance sheets to the total amount\\npresented in the unaudited condensed consolidated statements of cash flows (in millions):\\nDecember 31,\\n2023March 31,\\n2024\\nCash and cash equivalents $ 6,874 $ 7,829 \\nCash and cash equivalents included in funds receivable and amounts held on behalf of customers 5,769 8,665 \\nRestricted cash included in prepaids and other current assets 24 35 \\nTotal cash, cash equivalents, and restricted cash presented in the unaudited condensed consolidated statements of cash flows $ 12,667 $ 16,529 \\nSupplemental disclosures of balance sheet information\\nSupplemental balance sheet information consisted of the following (in millions):\\nDecember 31,\\n2023March 31,\\n2024\\nOther assets, noncurrent:\\nProperty and equipment, net $ 160 $ 171', metadata={'source': './data/airbnb_10q_q1.pdf', 'page': 10, '_id': 'dcc76f23b4c6413ba4b1b8f6095ddc68', '_collection_name': 'rag-chroma'}), Document(page_content='unrealized loss position for more than twelve months as of December 31, 2023 and March 31, 2024, respectively.\\nThe following table summarizes the contractual maturities of the Company’s available-for-sale debt securities (in millions):\\nMarch 31, 2024\\nAmortized\\nCostEstimated\\nFair Value\\nDue within one year $ 1,489 $ 1,489 \\nDue after one year through five years 957 947 \\nDue after five years 96 92 \\nTotal $ 2,542 $ 2,528 \\nNote 5. Fair Value Measurements and Financial Instruments\\nThe following table summarizes the Company’s financial assets and liabilities measured at fair value on a recurring basis (in millions):\\nDecember 31, 2023\\nLevel 1 Level 2 Level 3 Total\\nAssets\\nCash and cash equivalents:\\nMoney market funds $ 2,018 $ — $ — $ 2,018 \\nCertificates of deposit — 1 — 1 \\nGovernment bonds — 115 — 115 \\nCommercial paper — 223 — 223 \\nCorporate debt securities — 12 — 12 \\n2,018 351 — 2,369 \\nShort-term investments:', metadata={'source': './data/airbnb_10q_q1.pdf', 'page': 12, '_id': '1ca887ac192949e0b166c0ba896d6cc1', '_collection_name': 'rag-chroma'}), Document(page_content='Page: Car Allowance Rebate System\\nSummary: The Car Allowance Rebate System (CARS), colloquially known as \"cash for clunkers\", was a $3 billion U.S. federal scrappage program intended to provide economic incentives to U.S. residents to purchase a new, more fuel-efficient vehicle when trading in a less fuel-efficient vehicle. The program was promoted as a post-recession stimulus program to boost auto sales (which had declined due to the effects of the 2007–2008 financial crisis, leading to the Great Recession and 2008–2010 automotive industry crisis) while putting more fuel-efficient vehicles on the roadways.\\nThe program officially started on July 1, 2009, the processing of claims began July 24, and the program ended on August 24, 2009, as the appropriated funds were exhausted, having scrapped 677,081 vehicles. The deadline for dealers to submit applications was August 25. According to estimates of the Department of Transportation, the initial $1 billion appropriated for the system was exhausted by July 30, 2009, well before the anticipated end date of November 1, 2009, due to very high demand. In response, Congress approved an additional $2 billion.\\n\\n\\n\\nPage: Vietnamese cash\\nSummary: The Vietnamese cash (chữ Hán: 文錢 văn tiền; chữ Nôm: 銅錢 đồng tiền; French: sapèque), also called the sapek or sapèque, is a cast round coin with a square hole that was an official currency of Vietnam from the Đinh dynasty in 970 until the Nguyễn dynasty in 1945, and remained in circulation in North Vietnam until 1948. The same type of currency circulated in China, Japan, Korea, and Ryūkyū for centuries. Though the majority of Vietnamese cash coins throughout history were copper coins, lead, iron (from 1528) and zinc (from 1740) coins also circulated alongside them often at fluctuating rates (with 1 copper cash being worth 10 zinc cash in 1882). Coins made from metals of lower intrinsic value were introduced because of various superstitions involving Vietnamese people burying cash coins, as the problem of people burying cash coins became too much for the government. Almost all coins issued by government mints tended to be buried mere months after they had entered circulation. The Vietnamese government began issuing coins made from an alloy of zinc, lead, and tin. As these cash coins tended to be very fragile, they would decompose faster if buried, which caused the Vietnamese people to stop burying their coins.\\nThe inscriptions of Vietnamese cash coins can be written in either the Viết chéo (曰湥 / 曰袑, top-bottom-right-left) style or the Viết thuận (曰順, clockwise, top-right-bottom-left) style.\\n\\nPage: Cash transfer\\nSummary: A cash transfer is a direct transfer payment of money to an eligible person. Cash transfers are either unconditional cash transfers or conditional cash transfers. They may be provided by organisations funded by private donors, or a local or regional government.\\nCash transfers constitute a critical element in the realm of global social policy, addressing needs ranging from poverty alleviation to crisis response. This article distinguishes between two main types of cash transfers: humanitarian cash transfers, grounded in international humanitarian principles, and social assistance cash transfers, integral to long-term social welfare strategies.\\nHumanitarian cash transfers are aligned with the principles of humanity, neutrality, impartiality, and independence, which are fundamental to humanitarian aid. These principles, endorsed by UN General Assembly resolutions 46/182 and 58/114, ensure that humanitarian aid, including cash transfers, prioritizes human suffering and assists the most vulnerable without favoritism or discrimination. Humanity underscores the need to address human suffering universally. Neutrality ensures that aid does not favor any party in a conflict or dispute. Impartiality dictates that aid is based solely on need, without discrimination. Independence highlights the autonomy of humanitarian objectives from political, economic, military, or other objectives. Adhering to these principles, humanitarian cash transfers provide life-saving assistance in emergencies like natural disasters, conflicts, and famines, focusing on short-term, immediate relief.\\nConversely, social assistance cash transfers are part of broader social protection systems aimed at reducing long-term poverty and vulnerability. These transfers target various demographic groups, including the unemployed, single parents, and individuals facing disabilities or old age challenges. By enhancing the capacity of individuals to manage socioeconomic risks and promoting efficient labor markets, social assistance cash transfers contribute to social equity and stability.\\nWhile both humanitarian and social assistance cash transfers aim to provide monetary support to those in need, they differ significantly in their objectives, implementation, and impact. Humanitarian cash transfers are typically responsive, addressing urgent needs in crisis situation')]}\n",
      "Memory sufficient. Generating answer.\n",
      "Node 'start':\n",
      "\n",
      "---\n",
      "\n",
      "Generating answer...\n",
      "Node 'generate':\n",
      "\n",
      "---\n",
      "\n",
      "You asked about the topic of cash.\n"
     ]
    }
   ],
   "source": [
    "# Run\n",
    "config = {\"configurable\": {\"thread_id\": \"3\"}}\n",
    "\n",
    "inputs = {\"question\": \"What did I ask you about?\"}\n",
    "async for output in graph.astream(inputs, config=config):\n",
    "    for key, value in output.items():\n",
    "        # Node\n",
    "        print(f\"Node '{key}':\")\n",
    "        # Optional: print full state at each node\n",
    "        # print(value)\n",
    "    print(\"\\n---\\n\")\n",
    "\n",
    "# Final generation\n",
    "print(value[\"generation\"])\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from rich import print\n",
    "\n",
    "%load_ext rich"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "! pip install langchain_community tiktoken langchain-openai langchainhub chromadb langchain langgraph tavily-python"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Actual functions"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Wikipedia Tool"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from langchain_community.tools import WikipediaQueryRun\n",
    "from langchain_community.utilities import WikipediaAPIWrapper\n",
    "\n",
    "api_wrapper = WikipediaAPIWrapper(top_k_results=5, doc_content_chars_max=1000)\n",
    "\n",
    "tool = WikipediaQueryRun(api_wrapper=api_wrapper)\n",
    "print(tool.invoke(\"What's a state in computer science?\"))\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Load documents"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from langchain.text_splitter import RecursiveCharacterTextSplitter\n",
    "from langchain_community.document_loaders import WebBaseLoader\n",
    "from langchain_community.vectorstores import Qdrant\n",
    "from langchain_openai import OpenAIEmbeddings\n",
    "\n",
    "urls = [\n",
    "    \"https://lilianweng.github.io/posts/2023-06-23-agent/\",\n",
    "    \"https://lilianweng.github.io/posts/2023-03-15-prompt-engineering/\",\n",
    "    \"https://lilianweng.github.io/posts/2023-10-25-adv-attack-llm/\",\n",
    "]\n",
    "\n",
    "docs = [WebBaseLoader(url).load() for url in urls]\n",
    "docs_list = [item for sublist in docs for item in sublist]\n",
    "\n",
    "text_splitter = RecursiveCharacterTextSplitter.from_tiktoken_encoder(\n",
    "    chunk_size=250, chunk_overlap=0\n",
    ")\n",
    "doc_splits = text_splitter.split_documents(docs_list)\n",
    "\n",
    "# Add to vectorDB\n",
    "vectorstore = Qdrant.from_documents(\n",
    "    documents=doc_splits,\n",
    "    collection_name=\"rag-chroma\",\n",
    "    location=\":memory:\",\n",
    "    embedding=OpenAIEmbeddings(),\n",
    ")\n",
    "\n",
    "retriever = vectorstore.as_retriever()\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Grader"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from langchain_core.pydantic_v1 import BaseModel, Field\n",
    "from langchain_openai import ChatOpenAI\n",
    "from langchain_core.prompts import ChatPromptTemplate\n",
    "\n",
    "\n",
    "## Grading model\n",
    "class DocumentGrade(BaseModel):\n",
    "    \"\"\"Binary score for relevance check on retrieved documents\"\"\"\n",
    "\n",
    "    binary_score: str = Field(\n",
    "        description='Document is relevant to the question, \"yes\" or \"no\"'\n",
    "    )\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# LLM with function call\n",
    "llm = ChatOpenAI(model=\"gpt-3.5-turbo-0125\", temperature=0)\n",
    "structured_llm_grader = llm.with_structured_output(DocumentGrade)\n",
    "structured_llm_grader\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Grader Prompts\n",
    "system = \"\"\"You are a grader assessing relevance of a retrieved document to a user question. \\n \n",
    "    If the document contains keyword(s) or semantic meaning related to the question, grade it as relevant. \\n\n",
    "    Give a binary score 'yes' or 'no' score to indicate whether the document is relevant to the question.\"\"\"\n",
    "\n",
    "grade_prompt = ChatPromptTemplate.from_messages(\n",
    "    [\n",
    "        (\"system\", system),\n",
    "        (\"human\", \"Retrieved document: \\n\\n {document} \\n\\n User question: {question}\"),\n",
    "    ]\n",
    ")\n",
    "\n",
    "retrieval_grader = grade_prompt | structured_llm_grader\n",
    "\n",
    "question = \"agent memory\"\n",
    "docs = retriever.invoke(question)\n",
    "doc_txt = docs[1].page_content\n",
    "print(retrieval_grader.invoke({\"question\": question, \"document\": doc_txt}))\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Generate"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "agent_prompt = \"\"\"You are an assistant for question-answering tasks. Use the following pieces of retrieved context to answer the question. If you don't know the answer, just say that you don't know. Use three sentences maximum and keep the answer concise.\\nQuestion: {question} \\nContext: {context} \\nAnswer:\"\"\""
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "### Generate\n",
    "\n",
    "from langchain_core.output_parsers import StrOutputParser\n",
    "\n",
    "# Prompt\n",
    "prompt = ChatPromptTemplate.from_messages([(\"human\", agent_prompt)])\n",
    "prompt\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Generator LLM\n",
    "llm = ChatOpenAI(model_name=\"gpt-3.5-turbo\", temperature=0)\n",
    "\n",
    "\n",
    "# Post-processing\n",
    "def format_docs(docs):\n",
    "    return \"\\n\\n\".join(doc.page_content for doc in docs)\n",
    "\n",
    "\n",
    "# Chain\n",
    "rag_chain = prompt | llm | StrOutputParser()\n",
    "\n",
    "# Run\n",
    "generation = rag_chain.invoke({\"context\": docs, \"question\": question})\n",
    "print(generation)\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "rag_chain"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Question Re-writer"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Question rewriter LLM\n",
    "llm = ChatOpenAI(model=\"gpt-3.5-turbo-0125\", temperature=0)\n",
    "\n",
    "# Prompt\n",
    "system = \"\"\"You are a question re-writer that converts an input question to a better version that is optimized \\n \n",
    "     for web search. Look at the input and try to reason about the underlying semantic intent / meaning.\"\"\"\n",
    "\n",
    "re_write_prompt = ChatPromptTemplate.from_messages(\n",
    "    [\n",
    "        (\"system\", system),\n",
    "        (\n",
    "            \"human\",\n",
    "            \"Here is the initial question: \\n\\n {question} \\n Formulate an improved question.\",\n",
    "        ),\n",
    "    ]\n",
    ")\n",
    "\n",
    "question_rewriter = re_write_prompt | llm | StrOutputParser()\n",
    "question_rewriter.invoke({\"question\": question})\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Graph"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## State"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from typing import List\n",
    "from typing_extensions import TypedDict\n",
    "from langchain.schema import Document\n",
    "\n",
    "class GraphState(TypedDict):\n",
    "    \"\"\"\n",
    "    Represents the state of our graph.\n",
    "\n",
    "    Attributes:\n",
    "        question: question\n",
    "        generation: LLM generation\n",
    "        web_search: whether to add search\n",
    "        documents: list of documents\n",
    "    \"\"\"\n",
    "\n",
    "    question: str\n",
    "    generation: str\n",
    "    wiki_search: str\n",
    "    documents: List[str]\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Nodes"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "def retrieve(state):\n",
    "    print(\"Retrieving documents...\")\n",
    "\n",
    "    question = state[\"question\"]\n",
    "\n",
    "    docs = retriever.invoke(question)\n",
    "\n",
    "    return {\"question\": question, \"documents\": docs}"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "def generate(state):\n",
    "    print(\"Generating answer...\")\n",
    "\n",
    "    question = state[\"question\"]\n",
    "    documents = state[\"documents\"]\n",
    "\n",
    "    generation = rag_chain.invoke({\"context\": documents, \"question\": question})\n",
    "    return {\"documents\": documents, \"question\": question, \"generation\": generation}"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "def grade_documents(state):\n",
    "    print(\"Grading documents...\")\n",
    "\n",
    "    question = state[\"question\"]\n",
    "    documents = state[\"documents\"]\n",
    "\n",
    "    filtered_docs = []\n",
    "    search_wikipedia = False\n",
    "\n",
    "    for doc in documents:\n",
    "        score = retrieval_grader.invoke(\n",
    "            {\"question\": question, \"document\": doc.page_content}\n",
    "        )\n",
    "\n",
    "        grade = score.binary_score\n",
    "\n",
    "        if grade == \"yes\":\n",
    "            print(\"Document is relevant to the question.\")\n",
    "            filtered_docs.append(doc)\n",
    "        else:\n",
    "            print(\"Document is not relevant to the question.\")\n",
    "            search_wikipedia = True\n",
    "\n",
    "            continue\n",
    "\n",
    "    return {\n",
    "        \"documents\": filtered_docs,\n",
    "        \"question\": question,\n",
    "        \"wiki_search\": search_wikipedia,\n",
    "    }\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "def rewrite_query(state):\n",
    "    print(\"Rewriting question...\")\n",
    "\n",
    "    question = state[\"question\"]\n",
    "    documents = state[\"documents\"]\n",
    "\n",
    "    rewritten_question = question_rewriter.invoke({\"question\": question})\n",
    "\n",
    "    return {\"question\": rewritten_question, \"documents\": documents}\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "def search_wikipedia(state):\n",
    "    print(\"Searching Wikipedia...\")\n",
    "\n",
    "    question = state[\"question\"]\n",
    "    documents = state[\"documents\"]\n",
    "\n",
    "    wiki_search = tool.invoke(question)\n",
    "\n",
    "    wiki_results = Document(page_content=wiki_search)\n",
    "\n",
    "    documents.append(wiki_results)\n",
    "\n",
    "    return {\"question\": question, \"documents\": documents}\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Edges"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "def generate_or_not(state):\n",
    "    print(\"Determining whether to query Wikipedia...\")\n",
    "\n",
    "    question = state[\"question\"]\n",
    "    wiki_search = state[\"wiki_search\"]\n",
    "    documents = state[\"documents\"]\n",
    "\n",
    "    if wiki_search:\n",
    "        print(\"Rewriting query and supplementing information from Wikipedia...\")\n",
    "        return \"rewrite_query\"\n",
    "    \n",
    "    else:\n",
    "        print(\"Relevant documents found.\")\n",
    "        return \"generate\""
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Build and compile"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from langgraph.graph import StateGraph, END\n",
    "from langgraph.checkpoint.aiosqlite import AsyncSqliteSaver\n",
    "import aiosqlite\n",
    "\n",
    "workflow = StateGraph(GraphState)\n",
    "\n",
    "workflow.add_node(\"retrieve\", retrieve)\n",
    "workflow.add_node(\"grade_documents\", grade_documents)\n",
    "workflow.add_node(\"rewrite_query\", rewrite_query)\n",
    "workflow.add_node(\"search_wikipedia\", search_wikipedia)\n",
    "workflow.add_node(\"generate\", generate)\n",
    "\n",
    "workflow.set_entry_point(\"retrieve\")\n",
    "workflow.add_edge(\"retrieve\", \"grade_documents\")\n",
    "workflow.add_conditional_edges(\n",
    "    \"grade_documents\",\n",
    "    generate_or_not,\n",
    "    {\"rewrite_query\": \"rewrite_query\", \"generate\": \"generate\"},\n",
    ")\n",
    "\n",
    "workflow.add_edge(\"rewrite_query\", \"search_wikipedia\")\n",
    "workflow.add_edge(\"search_wikipedia\", \"generate\")\n",
    "workflow.add_edge(\"generate\", END)\n",
    "\n",
    "checkpoints = aiosqlite.connect(\"./checkpoints/checkpoint.sqlite\")\n",
    "memory = AsyncSqliteSaver(checkpoints) \n",
    "\n",
    "app = workflow.compile(checkpointer=memory)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "ename": "NameError",
     "evalue": "name 'app' is not defined",
     "output_type": "error",
     "traceback": [
      "\u001b[1;31m---------------------------------------------------------------------------\u001b[0m",
      "\u001b[1;31mNameError\u001b[0m                                 Traceback (most recent call last)",
      "Cell \u001b[1;32mIn[4], line 3\u001b[0m\n\u001b[0;32m      1\u001b[0m \u001b[38;5;28;01mfrom\u001b[39;00m \u001b[38;5;21;01mIPython\u001b[39;00m\u001b[38;5;21;01m.\u001b[39;00m\u001b[38;5;21;01mdisplay\u001b[39;00m \u001b[38;5;28;01mimport\u001b[39;00m Image\n\u001b[1;32m----> 3\u001b[0m Image(\u001b[43mapp\u001b[49m\u001b[38;5;241m.\u001b[39mget_graph()\u001b[38;5;241m.\u001b[39mdraw_mermaid_png())\n",
      "\u001b[1;31mNameError\u001b[0m: name 'app' is not defined"
     ]
    }
   ],
   "source": [
    "from IPython.display import Image\n",
    "\n",
    "Image(app.get_graph().draw_mermaid_png())"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "app"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Run\n",
    "config = {'configurable': {'thread_id': '1'}}\n",
    "\n",
    "inputs = {\"question\": \"What did I ask you about previously?\"}\n",
    "async for output in app.astream(inputs, config=config):\n",
    "    for key, value in output.items():\n",
    "        # Node\n",
    "        print(f\"Node '{key}':\")\n",
    "        # Optional: print full state at each node\n",
    "        print(value)\n",
    "    print(\"\\n---\\n\")\n",
    "\n",
    "# Final generation\n",
    "print(value[\"generation\"])\n",
    "\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "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": "3.11.0"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}