@arizeai/phoenix-mcp

{ "cells": [ { "cell_type": "markdown", "id": "9f853e403eabd4f8", "metadata": { "collapsed": false, "id": "9f853e403eabd4f8", "jupyter": { "outputs_hidden": false } }, "source": [ "<center>\n", " <p style=\"text-align:center\">\n", " <img alt=\"phoenix logo\" src=\"https://storage.googleapis.com/arize-phoenix-assets/assets/phoenix-logo-light.svg\" width=\"200\"/>\n", " <br>\n", " <a href=\"https://arize.com/docs/phoenix/\">Docs</a>\n", " |\n", " <a href=\"https://github.com/Arize-ai/phoenix\">GitHub</a>\n", " |\n", " <a href=\"https://arize-ai.slack.com/join/shared_invite/zt-2w57bhem8-hq24MB6u7yE_ZF_ilOYSBw#/shared-invite/email\">Community</a>\n", " </p>\n", "</center>\n", "<h1 align=\"center\">Tracing a LangGraph Agent</h1>\n", "\n", "LangGraph provides tools to easily define a structured AI Agent. However, it can be challenging to understand what is going on under the hood and to pinpoint the cause of issues. Phoenix makes your LLM applications *observable* by visualizing the underlying structure of each call to your query engine and surfacing problematic \"spans\" of execution based on latency, token count, or other evaluation metrics.\n", "\n", "In this tutorial, you will:\n", "- Build a simple SQL database agent using LangGraph,\n", "- Record trace data in OpenInference format,\n", "- Inspect the traces and spans of your application to identify sources of latency and cost\n", "\n", "ℹ️ This notebook requires an OpenAI API key." ] }, { "cell_type": "markdown", "id": "b5a87813ffe7e4d2", "metadata": { "collapsed": false, "id": "b5a87813ffe7e4d2", "jupyter": { "outputs_hidden": false } }, "source": [ "## Setup\n", "\n", "First let's install our required packages and set our API keys" ] }, { "cell_type": "code", "execution_count": null, "id": "4a4be247", "metadata": {}, "outputs": [], "source": [ "%pip install -U langgraph langchain_openai langchain_community arize-phoenix openinference-instrumentation-langchain" ] }, { "cell_type": "code", "execution_count": null, "id": "6c05a600f1afb5b6", "metadata": {}, "outputs": [], "source": [ "import getpass\n", "import os\n", "\n", "\n", "def _set_env(key: str):\n", " if key not in os.environ:\n", " os.environ[key] = getpass.getpass(f\"{key}:\")\n", "\n", "\n", "_set_env(\"OPENAI_API_KEY\")" ] }, { "cell_type": "markdown", "id": "877d8c85825089d8", "metadata": { "collapsed": false, "id": "877d8c85825089d8", "jupyter": { "outputs_hidden": false } }, "source": [ "## Configure the database\n", "\n", "We will be creating a SQLite database for this tutorial. SQLite is a lightweight database that is easy to set up and use. We will be loading the `chinook` database, which is a sample database that represents a digital media store.\n", "Find more information about the database [here](https://www.sqlitetutorial.net/sqlite-sample-database/).\n", "\n", "For convenience, we have hosted the database (`Chinook.db`) on a public GCS bucket." ] }, { "cell_type": "code", "execution_count": null, "id": "64b0bf1b14c2e902", "metadata": {}, "outputs": [], "source": [ "import requests\n", "\n", "url = \"https://storage.googleapis.com/benchmarks-artifacts/chinook/Chinook.db\"\n", "\n", "response = requests.get(url)\n", "\n", "if response.status_code == 200:\n", " # Open a local file in binary write mode\n", " with open(\"Chinook.db\", \"wb\") as file:\n", " # Write the content of the response (the file) to the local file\n", " file.write(response.content)\n", " print(\"File downloaded and saved as Chinook.db\")\n", "else:\n", " print(f\"Failed to download the file. Status code: {response.status_code}\")" ] }, { "cell_type": "markdown", "id": "61c8304aa5ceb6a5", "metadata": { "collapsed": false, "id": "61c8304aa5ceb6a5", "jupyter": { "outputs_hidden": false } }, "source": [ "We will use a handy SQL database wrapper available in the `langchain_community` package to interact with the database. The wrapper provides a simple interface to execute SQL queries and fetch results. We will also use the `langchain_openai` package to interact with the OpenAI API for language models later in the tutorial." ] }, { "cell_type": "code", "execution_count": null, "id": "1f1e1f4f86ed54", "metadata": {}, "outputs": [], "source": [ "from langchain_community.utilities import SQLDatabase\n", "\n", "db = SQLDatabase.from_uri(\"sqlite:///Chinook.db\")\n", "print(db.dialect)\n", "print(db.get_usable_table_names())\n", "db.run(\"SELECT * FROM Artist LIMIT 10;\")" ] }, { "cell_type": "markdown", "id": "6959e93141d8099c", "metadata": { "collapsed": false, "id": "6959e93141d8099c", "jupyter": { "outputs_hidden": false } }, "source": [ "## Utility functions\n", "\n", "We will define a few utility functions to help us with the agent implementation. Specifically, we will wrap a `ToolNode` with a fallback to handle errors and surface them to the agent." ] }, { "cell_type": "code", "execution_count": null, "id": "deae8460e4cf72b1", "metadata": {}, "outputs": [], "source": [ "from typing import Any\n", "\n", "from langchain_core.messages import ToolMessage\n", "from langchain_core.runnables import RunnableLambda, RunnableWithFallbacks\n", "from langgraph.prebuilt import ToolNode\n", "\n", "\n", "def create_tool_node_with_fallback(tools: list) -> RunnableWithFallbacks[Any, dict]:\n", " \"\"\"\n", " Create a ToolNode with a fallback to handle errors and surface them to the agent.\n", " \"\"\"\n", " return ToolNode(tools).with_fallbacks(\n", " [RunnableLambda(handle_tool_error)], exception_key=\"error\"\n", " )\n", "\n", "\n", "def handle_tool_error(state) -> dict:\n", " error = state.get(\"error\")\n", " tool_calls = state[\"messages\"][-1].tool_calls\n", " return {\n", " \"messages\": [\n", " ToolMessage(\n", " content=f\"Error: {repr(error)}\\n please fix your mistakes.\",\n", " tool_call_id=tc[\"id\"],\n", " )\n", " for tc in tool_calls\n", " ]\n", " }" ] }, { "cell_type": "markdown", "id": "d0196604f8cbb07b", "metadata": { "collapsed": false, "id": "d0196604f8cbb07b", "jupyter": { "outputs_hidden": false } }, "source": [ "## Define tools for the agent\n", "\n", "We will define a few tools that the agent will use to interact with the database.\n", "\n", "1. `list_tables_tool`: Fetch the available tables from the database\n", "2. `get_schema_tool`: Fetch the DDL for a table\n", "3. `db_query_tool`: Execute the query and fetch the results OR return an error message if the query fails\n", "\n", "For the first two tools, we will grab them from the `SQLDatabaseToolkit`, also available in the `langchain_community` package." ] }, { "cell_type": "code", "execution_count": null, "id": "452d049a3d2a4406", "metadata": {}, "outputs": [], "source": [ "from langchain_community.agent_toolkits import SQLDatabaseToolkit\n", "from langchain_openai import ChatOpenAI\n", "\n", "toolkit = SQLDatabaseToolkit(db=db, llm=ChatOpenAI(model=\"gpt-4o\"))\n", "tools = toolkit.get_tools()\n", "\n", "list_tables_tool = next(tool for tool in tools if tool.name == \"sql_db_list_tables\")\n", "get_schema_tool = next(tool for tool in tools if tool.name == \"sql_db_schema\")\n", "\n", "print(list_tables_tool.invoke(\"\"))\n", "\n", "print(get_schema_tool.invoke(\"Artist\"))" ] }, { "cell_type": "markdown", "id": "c16359edada327fa", "metadata": { "collapsed": false, "id": "c16359edada327fa", "jupyter": { "outputs_hidden": false } }, "source": [ "The third will be defined manually. For the `db_query_tool`, we will execute the query against the database and return the results." ] }, { "cell_type": "code", "execution_count": null, "id": "f7eb708ecb4c7cfc", "metadata": {}, "outputs": [], "source": [ "from langchain_core.tools import tool\n", "\n", "\n", "@tool\n", "def db_query_tool(query: str) -> str:\n", " \"\"\"\n", " Execute a SQL query against the database and get back the result.\n", " If the query is not correct, an error message will be returned.\n", " If an error is returned, rewrite the query, check the query, and try again.\n", " \"\"\"\n", " result = db.run_no_throw(query)\n", " if not result:\n", " return \"Error: Query failed. Please rewrite your query and try again.\"\n", " return result\n", "\n", "\n", "print(db_query_tool.invoke(\"SELECT * FROM Artist LIMIT 10;\"))" ] }, { "cell_type": "markdown", "id": "f1d66db8b8621639", "metadata": { "collapsed": false, "id": "f1d66db8b8621639", "jupyter": { "outputs_hidden": false } }, "source": [ "While not strictly a tool, we will prompt an LLM to check for common mistakes in the query and later add this as a node in the workflow." ] }, { "cell_type": "code", "execution_count": null, "id": "293017e8f05ac2b3", "metadata": {}, "outputs": [], "source": [ "from langchain_core.prompts import ChatPromptTemplate\n", "\n", "query_check_system = \"\"\"You are a SQL expert with a strong attention to detail.\n", "Double check the SQLite query for common mistakes, including:\n", "- Using NOT IN with NULL values\n", "- Using UNION when UNION ALL should have been used\n", "- Using BETWEEN for exclusive ranges\n", "- Data type mismatch in predicates\n", "- Properly quoting identifiers\n", "- Using the correct number of arguments for functions\n", "- Casting to the correct data type\n", "- Using the proper columns for joins\n", "\n", "If there are any of the above mistakes, rewrite the query. If there are no mistakes, just reproduce the original query.\n", "\n", "You will call the appropriate tool to execute the query after running this check.\"\"\"\n", "\n", "query_check_prompt = ChatPromptTemplate.from_messages(\n", " [(\"system\", query_check_system), (\"placeholder\", \"{messages}\")]\n", ")\n", "query_check = query_check_prompt | ChatOpenAI(model=\"gpt-4o\", temperature=0).bind_tools(\n", " [db_query_tool], tool_choice=\"required\"\n", ")\n", "\n", "query_check.invoke({\"messages\": [(\"user\", \"SELECT * FROM Artist LIMIT 10;\")]})" ] }, { "cell_type": "markdown", "id": "66f88452151e8188", "metadata": { "collapsed": false, "id": "66f88452151e8188", "jupyter": { "outputs_hidden": false } }, "source": [ "## Define the workflow\n", "\n", "We will then define the workflow for the agent. The agent will first force-call the `list_tables_tool` to fetch the available tables from the database, then follow the steps mentioned at the beginning of the tutorial." ] }, { "cell_type": "markdown", "id": "2fd9e41c-95c3-47aa-9a12-80b78cc7ac2d", "metadata": { "id": "2fd9e41c-95c3-47aa-9a12-80b78cc7ac2d" }, "source": [ "<div class=\"admonition note\">\n", " <p class=\"admonition-title\">Using Pydantic with LangChain</p>\n", " <p>\n", " This notebook uses Pydantic v2 <code>BaseModel</code>, which requires <code>langchain-core >= 0.3</code>. Using <code>langchain-core < 0.3</code> will result in errors due to mixing of Pydantic v1 and v2 <code>BaseModels</code>.\n", " </p>\n", "</div>" ] }, { "cell_type": "code", "execution_count": null, "id": "90d04ceea7b6b010", "metadata": {}, "outputs": [], "source": [ "from typing import Annotated, Literal\n", "\n", "from langchain_core.messages import AIMessage\n", "from langchain_openai import ChatOpenAI\n", "from langgraph.graph import END, START, StateGraph\n", "from langgraph.graph.message import AnyMessage, add_messages\n", "from pydantic import BaseModel, Field\n", "from typing_extensions import TypedDict\n", "\n", "\n", "# Define the state for the agent\n", "class State(TypedDict):\n", " messages: Annotated[list[AnyMessage], add_messages]\n", "\n", "\n", "# Define a new graph\n", "workflow = StateGraph(State)\n", "\n", "\n", "# Add a node for the first tool call\n", "def first_tool_call(state: State) -> dict[str, list[AIMessage]]:\n", " return {\n", " \"messages\": [\n", " AIMessage(\n", " content=\"\",\n", " tool_calls=[\n", " {\n", " \"name\": \"sql_db_list_tables\",\n", " \"args\": {},\n", " \"id\": \"tool_abcd123\",\n", " }\n", " ],\n", " )\n", " ]\n", " }\n", "\n", "\n", "def model_check_query(state: State) -> dict[str, list[AIMessage]]:\n", " \"\"\"\n", " Use this tool to double-check if your query is correct before executing it.\n", " \"\"\"\n", " return {\"messages\": [query_check.invoke({\"messages\": [state[\"messages\"][-1]]})]}\n", "\n", "\n", "workflow.add_node(\"first_tool_call\", first_tool_call)\n", "\n", "# Add nodes for the first two tools\n", "workflow.add_node(\"list_tables_tool\", create_tool_node_with_fallback([list_tables_tool]))\n", "workflow.add_node(\"get_schema_tool\", create_tool_node_with_fallback([get_schema_tool]))\n", "\n", "# Add a node for a model to choose the relevant tables based on the question and available tables\n", "model_get_schema = ChatOpenAI(model=\"gpt-4o\", temperature=0).bind_tools([get_schema_tool])\n", "workflow.add_node(\n", " \"model_get_schema\",\n", " lambda state: {\n", " \"messages\": [model_get_schema.invoke(state[\"messages\"])],\n", " },\n", ")\n", "\n", "\n", "# Describe a tool to represent the end state\n", "class SubmitFinalAnswer(BaseModel):\n", " \"\"\"Submit the final answer to the user based on the query results.\"\"\"\n", "\n", " final_answer: str = Field(..., description=\"The final answer to the user\")\n", "\n", "\n", "# Add a node for a model to generate a query based on the question and schema\n", "query_gen_system = \"\"\"You are a SQL expert with a strong attention to detail.\n", "\n", "Given an input question, output a syntactically correct SQLite query to run, then look at the results of the query and return the answer.\n", "\n", "DO NOT call any tool besides SubmitFinalAnswer to submit the final answer.\n", "\n", "When generating the query:\n", "\n", "Output the SQL query that answers the input question without a tool call.\n", "\n", "Unless the user specifies a specific number of examples they wish to obtain, always limit your query to at most 5 results.\n", "You can order the results by a relevant column to return the most interesting examples in the database.\n", "Never query for all the columns from a specific table, only ask for the relevant columns given the question.\n", "\n", "If you get an error while executing a query, rewrite the query and try again.\n", "\n", "If you get an empty result set, you should try to rewrite the query to get a non-empty result set.\n", "NEVER make stuff up if you don't have enough information to answer the query... just say you don't have enough information.\n", "\n", "If you have enough information to answer the input question, simply invoke the appropriate tool to submit the final answer to the user.\n", "\n", "DO NOT make any DML statements (INSERT, UPDATE, DELETE, DROP etc.) to the database.\"\"\"\n", "query_gen_prompt = ChatPromptTemplate.from_messages(\n", " [(\"system\", query_gen_system), (\"placeholder\", \"{messages}\")]\n", ")\n", "query_gen = query_gen_prompt | ChatOpenAI(model=\"gpt-4o\", temperature=0).bind_tools(\n", " [SubmitFinalAnswer]\n", ")\n", "\n", "\n", "def query_gen_node(state: State):\n", " message = query_gen.invoke(state)\n", "\n", " # Sometimes, the LLM will hallucinate and call the wrong tool. We need to catch this and return an error message.\n", " tool_messages = []\n", " if message.tool_calls:\n", " for tc in message.tool_calls:\n", " if tc[\"name\"] != \"SubmitFinalAnswer\":\n", " tool_messages.append(\n", " ToolMessage(\n", " content=f\"Error: The wrong tool was called: {tc['name']}. Please fix your mistakes. Remember to only call SubmitFinalAnswer to submit the final answer. Generated queries should be outputted WITHOUT a tool call.\",\n", " tool_call_id=tc[\"id\"],\n", " )\n", " )\n", " else:\n", " tool_messages = []\n", " return {\"messages\": [message] + tool_messages}\n", "\n", "\n", "workflow.add_node(\"query_gen\", query_gen_node)\n", "\n", "# Add a node for the model to check the query before executing it\n", "workflow.add_node(\"correct_query\", model_check_query)\n", "\n", "# Add node for executing the query\n", "workflow.add_node(\"execute_query\", create_tool_node_with_fallback([db_query_tool]))\n", "\n", "\n", "# Define a conditional edge to decide whether to continue or end the workflow\n", "def should_continue(state: State) -> Literal[END, \"correct_query\", \"query_gen\"]:\n", " messages = state[\"messages\"]\n", " last_message = messages[-1]\n", " # If there is a tool call, then we finish\n", " if getattr(last_message, \"tool_calls\", None):\n", " return END\n", " if last_message.content.startswith(\"Error:\"):\n", " return \"query_gen\"\n", " else:\n", " return \"correct_query\"\n", "\n", "\n", "# Specify the edges between the nodes\n", "workflow.add_edge(START, \"first_tool_call\")\n", "workflow.add_edge(\"first_tool_call\", \"list_tables_tool\")\n", "workflow.add_edge(\"list_tables_tool\", \"model_get_schema\")\n", "workflow.add_edge(\"model_get_schema\", \"get_schema_tool\")\n", "workflow.add_edge(\"get_schema_tool\", \"query_gen\")\n", "workflow.add_conditional_edges(\n", " \"query_gen\",\n", " should_continue,\n", ")\n", "workflow.add_edge(\"correct_query\", \"execute_query\")\n", "workflow.add_edge(\"execute_query\", \"query_gen\")\n", "\n", "# Compile the workflow into a runnable\n", "app = workflow.compile()" ] }, { "cell_type": "markdown", "id": "6c344ae086ba8d22", "metadata": { "collapsed": false, "id": "6c344ae086ba8d22", "jupyter": { "outputs_hidden": false } }, "source": [ "## Visualize the graph" ] }, { "cell_type": "code", "execution_count": null, "id": "4f200d1813897000", "metadata": {}, "outputs": [], "source": [ "from IPython.display import Image, display\n", "from langchain_core.runnables.graph import MermaidDrawMethod\n", "\n", "display(\n", " Image(\n", " app.get_graph().draw_mermaid_png(\n", " draw_method=MermaidDrawMethod.API,\n", " )\n", " )\n", ")" ] }, { "cell_type": "markdown", "id": "7679154e", "metadata": { "id": "7679154e" }, "source": [ "### Auto-Instrument Agent\n", "\n", "This example uses a hosted Phoenix instance. If you don't have one already, create one for free [here](https://phoenix.arize.com) to get your API key.\n", "\n", "If you'd rather self-host Phoenix, you can do so by following the instructions [here](https://arize.com/docs/phoenix/deployment)." ] }, { "cell_type": "code", "execution_count": null, "id": "294df67b", "metadata": {}, "outputs": [], "source": [ "_set_env(\"PHOENIX_API_KEY\")" ] }, { "cell_type": "code", "execution_count": null, "id": "9fb84dee", "metadata": {}, "outputs": [], "source": [ "from openinference.instrumentation.langchain import LangChainInstrumentor\n", "\n", "from phoenix.otel import register\n", "\n", "os.environ[\"PHOENIX_COLLECTOR_ENDPOINT\"] = \"https://app.phoenix.arize.com\"\n", "os.environ[\"PHOENIX_CLIENT_HEADERS\"] = f\"api_key={os.getenv('PHOENIX_API_KEY')}\"\n", "\n", "tracer_provider = register()\n", "LangChainInstrumentor().instrument(tracer_provider=tracer_provider)" ] }, { "cell_type": "markdown", "id": "bdf78dc68548522c", "metadata": { "collapsed": false, "id": "bdf78dc68548522c", "jupyter": { "outputs_hidden": false } }, "source": [ "## Run the agent" ] }, { "cell_type": "code", "execution_count": null, "id": "85958809-03c5-4e52-97cc-e7c0ae986f60", "metadata": {}, "outputs": [], "source": [ "messages = app.invoke({\"messages\": [(\"user\", \"Which albums did Aerosmith release?\")]})\n", "json_str = messages[\"messages\"][-1].tool_calls[0][\"args\"][\"final_answer\"]\n", "json_str" ] }, { "cell_type": "markdown", "id": "b50378cc", "metadata": { "id": "b50378cc" }, "source": [ "## Next Steps\n", "\n", "You should now see traces in the [Phoenix UI](https://app.phoenix.arize.com)!\n", "\n", "These traces contain information about the execution of the agent, including the tool calls, the LLM calls, and the overall execution path.\n", "\n", "From here you can use the traces you've captured to:\n", "- Visualize the execution path of your agent, track token usage, latency, and errors\n", "- [Evaluate various metrics of your agent's execution](https://arize.com/docs/phoenix/evaluation/how-to-evals/evaluating-phoenix-traces), such as the quality of its tool outputs and overall path\n", "- [Add annotations](https://arize.com/docs/phoenix/tracing/how-to-tracing/capture-feedback#send-annotations-to-phoenix) to traces to capture feedback or additional context on each execution\n", "- Use your annotations and evaluations to [create datasets](https://arize.com/docs/phoenix/datasets-and-experiments/how-to-datasets/creating-datasets) to power [experiments](https://arize.com/docs/phoenix/datasets-and-experiments/how-to-experiments/run-experiments), or export for [fine-tuning](https://arize.com/docs/phoenix/datasets-and-experiments/how-to-datasets/exporting-datasets)." ] } ], "metadata": { "language_info": { "name": "python" } }, "nbformat": 4, "nbformat_minor": 5 }