examples/Structured_web_insights.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "<img src=\"https://raw.githubusercontent.com/instill-ai/cookbook/main/images/Logo.png\" alt=\"Instill Logo\" width=\"300\"/>"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Curate Structured Insights from the Web with **Instill VDP** on **Instill Cloud**\n",
    "\n",
    "In this notebook we will demonstrate how to generate structured insights from the web using either a Google Search query, or simply specifying a URL. This content complements our tutorial: [\"Mastering OpenAI's Structured Outputs\"](https://www.instill.tech/blog/mastering-openai-structured-outputs), where we provide a step-by-step guide for building an LLM-powered knowledge curation pipeline with **Instill VDP** on **Instill Cloud**, highlight best practices along the way.\n",
    "\n",
    "More specifically, we will be calling two pipelines in this notebook using our [Python SDK](https://www.instill.tech/docs/sdk/python):\n",
    "1. Full Pipeline: [structured-web-insights](https://instill.tech/instill-ai/pipelines/structured-web-insights/playground) (requires your own Google API key, see setup docs)\n",
    "2. URL-only Version: [structured-website-summary](https://instill.tech/instill-ai/pipelines/structured-web-insights-from-url/playground) (no Google search, input specific URL)\n",
    "\n",
    "We recommend that you first start by going through the [tutorial](https://www.instill.tech/blog/mastering-openai-structured-outputs) - **see the video below** - which walks you through the steps required to build the pipeline that we will be calling in this notebook!"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/jpeg": "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",
      "text/html": [
       "\n",
       "        <iframe\n",
       "            width=\"1000\"\n",
       "            height=\"600\"\n",
       "            src=\"https://www.youtube.com/embed/x0XsmuCrit8\"\n",
       "            frameborder=\"0\"\n",
       "            allowfullscreen\n",
       "            \n",
       "        ></iframe>\n",
       "        "
      ],
      "text/plain": [
       "<IPython.lib.display.YouTubeVideo at 0x104363ed0>"
      ]
     },
     "execution_count": 1,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "from IPython.display import YouTubeVideo\n",
    "\n",
    "YouTubeVideo('x0XsmuCrit8', width=1000, height=600)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Setup\n",
    "\n",
    "To execute all of the code in this notebook, you’ll need to create a free Instill Cloud account and setup an API Token. To create your account, please refer to our [quickstart guide](https://www.instill.tech/docs/quickstart). For generating your API Token, consult the [API Token Management](https://www.instill.tech/docs/core/token) page.\n",
    "\n",
    "**This will give you access to 10,000 free credits per month that you can use to make API calls with third-party AI vendors. Please see our [documentation](https://www.instill.tech/docs/cloud/credit) for further details.**"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "We will now install the latest Instill Python SDK, import the required libraries, and configure the SDK with a valid API token."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "!pip install instill-sdk==0.15.1 --quiet"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [],
   "source": [
    "from google.protobuf.json_format import MessageToDict\n",
    "from pprint import pprint\n",
    "from IPython.display import IFrame\n",
    "\n",
    "from instill.clients import init_pipeline_client\n",
    "pipeline = init_pipeline_client(api_token=\"YOUR_INSTILL_API_TOKEN\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 1. Structured Web Insights from URL\n",
    "\n",
    "Here, we will view and try out the [structured-web-insights-from-url](https://instill.tech/instill-ai/pipelines/structured-web-insights-from-url/playground) pipeline. Checkout the pipeline preview page and README in the IFrame below to see the structure of this pipeline and understand how it works!"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "\n",
       "        <iframe\n",
       "            width=\"1000\"\n",
       "            height=\"700\"\n",
       "            src=\"https://instill.tech/instill-ai/pipelines/structured-web-insights-from-url/preview\"\n",
       "            frameborder=\"0\"\n",
       "            allowfullscreen\n",
       "            \n",
       "        ></iframe>\n",
       "        "
      ],
      "text/plain": [
       "<IPython.lib.display.IFrame at 0x106bd6390>"
      ]
     },
     "execution_count": 6,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "IFrame('https://instill.tech/instill-ai/pipelines/structured-web-insights-from-url/preview', width=1000, height=700)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [],
   "source": [
    "response = pipeline.trigger(\n",
    "    namespace_id=\"instill-ai\",\n",
    "    pipeline_id=\"structured-web-insights-from-url\",\n",
    "    data=[{\"query\": \"I want to know about structured output best practices\",\n",
    "           \"website-url\": \"https://www.instill.tech/blog/llm-structured-outputs\"}],\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "{'author': 'George Strong',\n",
      " 'key_insights': [\"OpenAI's structured outputs API represents a significant \"\n",
      "                  'advancement in handling unstructured data.',\n",
      "                  'LLMs often struggle with reasoning tasks when subjected to '\n",
      "                  'strict format restrictions.',\n",
      "                  'Effective techniques for generating structured outputs '\n",
      "                  'include prompt engineering, output parsing, and error '\n",
      "                  'handling.',\n",
      "                  'A multistep pipeline approach can enhance the accuracy of '\n",
      "                  'reasoning and structured output generation by separating '\n",
      "                  'these tasks.'],\n",
      " 'published_date': 'August 28, 2024',\n",
      " 'source': 'Instill AI Blog',\n",
      " 'summary': 'The article discusses best practices for generating structured '\n",
      "            'outputs from Large Language Models (LLMs), particularly in light '\n",
      "            \"of OpenAI's recent introduction of structured outputs in their \"\n",
      "            'API. It highlights the complexities involved in structuring '\n",
      "            'outputs, the challenges faced by LLMs under format restrictions, '\n",
      "            'and compares various tools and frameworks for generating '\n",
      "            'structured outputs. The article also proposes a multistep '\n",
      "            'pipeline approach using Instill VDP to effectively separate '\n",
      "            'reasoning tasks from structuring tasks, thereby improving '\n",
      "            'accuracy and output quality.',\n",
      " 'tags': ['Structured Outputs',\n",
      "          'Large Language Models',\n",
      "          'OpenAI',\n",
      "          'Prompt Engineering',\n",
      "          'Output Parsing',\n",
      "          'Error Handling',\n",
      "          'Instill VDP',\n",
      "          'Reasoning Tasks',\n",
      "          'Data Models',\n",
      "          'API'],\n",
      " 'title': 'The Best Way to Generate Structured Output from LLMs',\n",
      " 'url': 'https://www.instill.tech/blog/llm-structured-outputs'}\n"
     ]
    }
   ],
   "source": [
    "response = MessageToDict(response)\n",
    "\n",
    "structured_output = response[\"outputs\"][0][\"structured-output\"]\n",
    "pprint(structured_output)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "The Best Way to Generate Structured Output from LLMs \\| Blog\n",
      "\n",
      "Hacktoberfest 2024! [Win Instill AI Swag Pack!](https://www.instill.tech/blog/hacktoberfest-2024-opening)\n",
      "\n",
      "ExploreFor BusinessFor DevelopersCompany [Pricing](http://www.instill.tech/pricing)\n",
      "\n",
      "Get Started\n",
      "\n",
      "Blog\n",
      "\n",
      "![The theme of this tutorial](http://www.instill.tech/blog-assets/structured-outputs/scaffolding.png)\n",
      "\n",
      "# The Best Way to Generate Structured Output from LLMs\n",
      "\n",
      "Benchmarking LLM structured output performance with OpenAI, Instructor, Marvin, BAML, TypeChat, LangChain, and how to overcome reasoning deficiencies using a multi-step Instill VDP pipeline.\n",
      "\n",
      "August 28, 2024\n",
      "\n",
      "Insight\n",
      "\n",
      "![George Strong's github avatar](http://www.instill.tech/author-avatar/george.jpeg)\n",
      "\n",
      "[George Strong](https://github.com/georgewilliamstrong)\n",
      "\n",
      "Industries are eagerly capitalizing on Large Language Models (LLMs) to unlock\n",
      "the potential within their vast reserves of under-utilized unstructured data.\n",
      "Given that up to [80% of the worlds data is soon forecast to be\\\n",
      "unstructured](https://www.ibm.com/blog/managing-unstructured-data/), the drive\n",
      "to harness this wealth for innovation and new product development is immense.\n",
      "There is an ironic paradox here: LLMs, by their very design, output _more_\n",
      "unstructured text data to manage and keep on top of. That is, until very\n",
      "recently!\n",
      "\n",
      "Earlier this month, OpenAI announced that they now support [**Structured Outputs**\\\n",
      "**in the**\\\n",
      "**API**](https://openai.com/index/introducing-structured-outputs-in-the-api/) with\n",
      "general availability. The ability to distill and transform the creative and\n",
      "diverse unstructured outputs of LLMs into actionable and reliable structured\n",
      "data represents a huge milestone in the world of unstructured data ETL (Extract,\n",
      "Transform and Load).\n",
      "\n",
      "However, there's more to this story than meets the eye.\n",
      "\n",
      "## [\\#](http://www.instill.tech\\#the-complexities-of-structuring-outputs-from-llms) The Complexities of Structuring Outputs from LLMs\n",
      "\n",
      "Coincidentally, the day before OpenAI's announcement, a paper was published\n",
      "titled [Let Me Speak Freely? A Study on the Impact of Format Restrictions on\\\n",
      "Performance of Large Language Models](https://arxiv.org/abs/2408.02442v1), which\n",
      "offers a compelling counterpoint. They demonstrate that LLMs **struggle with**\n",
      "**reasoning tasks when they're placed under format restrictions**. Additionally,\n",
      "the _stricter_ these format restrictions are, the _more_ their reasoning\n",
      "performance drops, revealing a complex interplay between structuring outputs and\n",
      "model performance.\n",
      "\n",
      "Besides the new structured outputs API from OpenAI, there is also a multitude of\n",
      "existing LLM frameworks that have been developed to tease out structured outputs\n",
      "from conventional LLMs. Under the hood, each of these tools works by using a\n",
      "blend of the following techniques:\n",
      "\n",
      "### [\\#](http://www.instill.tech\\#1-prompt-engineering) 1\\. Prompt Engineering\n",
      "\n",
      "**Structured Prompts:** Well-crafted prompts are used to instruct the LLM to\n",
      "output data in a specific format, such as JSON. These prompts guide the model to\n",
      "generate responses that match the expected structure.\n",
      "\n",
      "**Example Guidance:** Prompts often include examples of the desired output\n",
      "format to further guide the model.\n",
      "\n",
      "### [\\#](http://www.instill.tech\\#2-output-parsing) 2\\. Output Parsing\n",
      "\n",
      "**Data Models:** The structured output is typically aligned with predefined data\n",
      "models. These models serve as blueprints to ensure the output data conforms to\n",
      "specific formats and structures.\n",
      "\n",
      "**Validation:** After the LLM generates the output, it is validated against\n",
      "these data models. If the output doesn’t match the expected structure, error\n",
      "handling mechanisms may be triggered to correct or retry the process.\n",
      "\n",
      "### [\\#](http://www.instill.tech\\#3-error-handling) 3\\. Error Handling\n",
      "\n",
      "**Retries and Corrections:** If the output doesn't meet the required format\n",
      "(e.g., due to syntax errors or unexpected structures), adjustments may be made\n",
      "to the prompt, or the output may be re-parsed to correct the issue.\n",
      "\n",
      "**Strict Parsing:** The framework includes strict parsing tools that ensure the\n",
      "output conforms exactly to the required schema, raising exceptions if the output\n",
      "is malformed.\n",
      "\n",
      "This raises the question: **How do these tools compare with OpenAI's structured**\n",
      "**outputs feature on a task that involves both reasoning and output format**\n",
      "**restrictions?**\n",
      "\n",
      "## [\\#](http://www.instill.tech\\#benchmarking-structured-output-tools-with-reasoning-tasks) Benchmarking Structured Output Tools with Reasoning Tasks\n",
      "\n",
      "To better understand the complexities and nuances of generating structured\n",
      "outputs from LLMs, we conducted an experiment using OpenAI's latest GPT-4o\n",
      "model. This experiment aimed to evaluate how OpenAI's new structured outputs in\n",
      "the API feature compared with various existing tools and frameworks on a complex\n",
      "task that involved both reasoning and producing outputs that are structured\n",
      "according to a predefined data model.\n",
      "\n",
      "The full details, code, and results of this experiment can be found in the\n",
      "accompanying notebook in the new [**Instill AI**\\\n",
      "**Cookbook**](https://github.com/instill-ai/cookbook). Launch it with Google Colab\n",
      "using the button below:\n",
      "\n",
      "[Open in Colab 🚀](https://colab.research.google.com/github/instill-ai/cookbook/blob/main/examples/Generating_structured_outputs.ipynb)\n",
      "\n",
      "### [\\#](http://www.instill.tech\\#the-task-combining-reasoning-and-structured-output) The Task: Combining Reasoning and Structured Output\n",
      "\n",
      "Our benchmark task was designed to challenge the LLM's ability to reason while\n",
      "adhering to a strict output structure. Inspired by the aforementioned\n",
      "[paper](https://arxiv.org/abs/2408.02442v1), we created a task that involved two\n",
      "main components:\n",
      "\n",
      "1. **Reasoning Problem:** The LLM was asked to calculate a freelancer's weekly\n",
      "earnings based on varying hourly rates, including overtime pay, as follows:\n",
      "\n",
      "\n",
      "\n",
      "`\n",
      "John Doe is a freelance software engineer. He charges a base rate of $50 per\n",
      "hour for the first 29 hours of work each week. For any additional hours, he\n",
      "charges 1.7 times his base hourly rate. This week, John worked on a project\n",
      "for 38 hours. How much will John Doe charge his client for the project this\n",
      "week?\n",
      "`\n",
      "\n",
      "2. **Structured Output:** In addition to solving the reasoning problem, the LLM\n",
      "was also asked to summarize information from a resume into a specific data\n",
      "model, shown below:\n",
      "\n",
      "\n",
      "\n",
      "`\n",
      "class DataModel(BaseModel):\n",
      "name: str\n",
      "email: str\n",
      "cost: float  # Answer to the reasoning problem, stored as a float\n",
      "experience: list[str]\n",
      "skills: list[str]\n",
      "`\n",
      "\n",
      "\n",
      "\n",
      "The LLM was instructed to store the result of the reasoning problem in the\n",
      "`cost` field of the data model, while the rest of the fields were to be\n",
      "populated with data extracted and summarized from the following resume text:\n",
      "\n",
      "\n",
      "\n",
      "`\n",
      "John Doe\n",
      "1234 Elm Street\n",
      "Springfield, IL 62701\n",
      "(123) 456-7890\n",
      "Email: john.doe@gmail.com\n",
      "Objective: To obtain a position as a software engineer.\n",
      "Education:\n",
      "Bachelor of Science in Computer Science\n",
      "University of Illinois at Urbana-Champaign\n",
      "May 2020 - May 2024\n",
      "Experience:\n",
      "Software Engineer Intern\n",
      "Google\n",
      "May 2022 - August 2022\n",
      "- Worked on the Google Search team\n",
      "- Developed new features for the search engine\n",
      "- Wrote code in Python and C++\n",
      "Software Engineer Intern\n",
      "Facebook\n",
      "May 2021 - August 2021\n",
      "- Worked on the Facebook Messenger team\n",
      "- Developed new features for the messenger app\n",
      "- Wrote code in Python and Java\n",
      "`\n",
      "\n",
      "### [\\#](http://www.instill.tech\\#existing-tools-for-structured-output-generation) Existing Tools for Structured Output Generation\n",
      "\n",
      "We evaluated several existing libraries and frameworks designed to help generate\n",
      "structured outputs from LLMs. The tools tested include:\n",
      "\n",
      "- [**Instructor**](https://python.useinstructor.com): A Python library built on\n",
      "Pydantic that facilitates generating structured output from LLMs.\n",
      "- [**Marvin**](https://www.askmarvin.ai): A tool for building reliable natural\n",
      "language interfaces.\n",
      "- [**BAML**](https://www.boundaryml.com): A domain-specific language for writing\n",
      "and testing LLM functions.\n",
      "- [**TypeChat**](https://microsoft.github.io/TypeChat/): A tool from Microsoft\n",
      "for getting well-typed responses from language models.\n",
      "- [**LangChain**](https://www.langchain.com): A Python library that integrates\n",
      "language models with data and APIs to build applications.\n",
      "\n",
      "[Outlines](https://outlines-dev.github.io/outlines/),\n",
      "[JSONformer](https://github.com/1rgs/jsonformer) and\n",
      "[Guidance](https://github.com/guidance-ai/guidance/tree/main) were also\n",
      "considered, however they were left out of this experiment as they had limited\n",
      "support for remote API calls or failed when integrating with the latest OpenAI\n",
      "API.\n",
      "\n",
      "### [\\#](http://www.instill.tech\\#experiment-setup) Experiment Setup\n",
      "\n",
      "The GPT-4o model from OpenAI was used by all tools. To ensure a reasonable\n",
      "amount of variation, its temperature was set to 1.0, with the exception of\n",
      "TypeChat which operates with a temperature of 0.0 and doesn't readily expose\n",
      "this parameter to the user.\n",
      "\n",
      "Each tool processed 50 inference requests based on the same example task\n",
      "described above. For more details on the experimental setup, please refer to the\n",
      "accompanying notebook.\n",
      "\n",
      "### [\\#](http://www.instill.tech\\#results) Results\n",
      "\n",
      "The performance of these tools was assessed based on three criteria:\n",
      "\n",
      "1. **Correctness of the Structured Output:** Whether the output was correctly\n",
      "formatted according to the data model.\n",
      "2. **Accuracy of the Reasoning Task:** Whether the tool could correctly\n",
      "calculate the `cost` based on the reasoning problem.\n",
      "3. **Mean Absolute Error:** The average absolute difference between the\n",
      "calculated `cost` and the ground truth value.\n",
      "\n",
      "The results of the experiment are summarized in the table below:\n",
      "\n",
      "**Tool****Correct Output Structure****Correct Reasoning****Mean Absolute Error**OpenAI (Text)❌✅$0.00OpenAI (Structured)✅❌$93.93Instructor✅❌$109.61Marvin✅❌$71.01BAML✅❌$72.94TypeChat✅❌$100.00LangChain✅❌$97.23\n",
      "\n",
      "_Please note that these were generated using the latest OpenAI GPT-4o model on_\n",
      "_the 27th of August 2024. Future experiments will produce variations in these_\n",
      "_results as the APIs and tools evolve._\n",
      "\n",
      "## [\\#](http://www.instill.tech\\#key-insights) Key Insights\n",
      "\n",
      "While OpenAI's GPT-4o model was able to calculate the correct reasoning result\n",
      "when generating unstructured text, it was unable to return the correct value\n",
      "when the output was constrained to the structured data model. Additionally, none\n",
      "of the existing tools were able to correctly solve the reasoning task while\n",
      "adhering to the structured output format.\n",
      "\n",
      "This aligns with findings from the [paper](https://arxiv.org/abs/2408.02442v1),\n",
      "which highlighted how format restrictions can impair reasoning abilities. It\n",
      "also shows that the challenges identified in the original study with\n",
      "GPT-3.5-turbo persist: a) with OpenAI's latest flagship model, and b) when this\n",
      "model is used in tandem with an array of existing output structuring tools.\n",
      "\n",
      "### [\\#](http://www.instill.tech\\#can-chain-of-thought-reasoning-help) Can Chain-of-Thought Reasoning Help?\n",
      "\n",
      "In the Structured Outputs [release\\\n",
      "post](https://openai.com/index/introducing-structured-outputs-in-the-api/),\n",
      "OpenAI suggest that the quality of the final response can be improved using\n",
      "chain-of-thought reasoning within the schema by defining a field called\n",
      "`reasoning_steps`. This approach involves having the model outline its reasoning\n",
      "process in this field before placing the final answer in the `cost` field.\n",
      "\n",
      "Interestingly, our tests with this approach - implemented in the\n",
      "[`structured-output-reasoning-cot`](https://instill.tech/george_strong/pipelines/structured-output-reasoning-cot/playground)\n",
      "pipeline - showed that, even when the reasoning was correctly detailed by the\n",
      "model in `reasoning_steps`, it was still unable to populate the `cost` field\n",
      "accurately.\n",
      "\n",
      "Despite advancements in model capabilities, combining complex reasoning with\n",
      "structured output generation remains a significant challenge. A nice quote from\n",
      "OpenAI's [release\\\n",
      "post](https://openai.com/index/introducing-structured-outputs-in-the-api/) hints\n",
      "at this problem, and provides some additional guidance:\n",
      "\n",
      "\" _Structured Outputs doesn't prevent all kinds of model mistakes. For example, the_\n",
      "_model may still make mistakes within the values of the JSON object (e.g.,_\n",
      "_getting a step wrong in a mathematical equation). If developers find mistakes,_\n",
      "_we recommend providing examples in the system instructions or splitting tasks_\n",
      "_into simpler subtasks._\"\n",
      "\n",
      "This motivates our next question: **How can we reliably overcome these reasoning**\n",
      "**and output structuring challenges?**\n",
      "\n",
      "## [\\#](http://www.instill.tech\\#the-solution-a-multistep-pipeline-built-with--instill-vdp) The Solution: A Multistep Pipeline Built with **💧 Instill VDP**\n",
      "\n",
      "Given the difficulties observed across all frameworks, a more robust approach is\n",
      "needed. Instead of attempting to solve the reasoning and structuring tasks\n",
      "simultaneously, we can use Instill Versatile Data Pipeline (VDP) to build a\n",
      "multistep pipeline that divides these into two distinct stages:\n",
      "\n",
      "1. **Reasoning Step:** In the first step, the LLM is tasked solely with solving\n",
      "the reasoning problem without any constraints on the output format. This\n",
      "allows the model to leverage its full reasoning capabilities without being\n",
      "hindered by strict formatting requirements.\n",
      "\n",
      "2. **Structuring Step:** In the second step, the unstructured output from the\n",
      "reasoning step is passed to a component that focuses exclusively on\n",
      "structuring the data into the required format.\n",
      "\n",
      "\n",
      "This approach ensures that the system is able to adhere to the specified data\n",
      "model, whilst also enabling the reasoning LLM to solve complex tasks,\n",
      "unencumbered by format constraints. The full results of the benchmarking\n",
      "experiment are shown in the Figure below:\n",
      "\n",
      "![Benchmarking experiment results comparing Multistep Instill VDP, OpenAI, Instructor, Marvin, BAML, TypeChat, and LangChain on reasoning and structured output generation.](http://www.instill.tech/blog-assets/structured-outputs/benchmark.png)\n",
      "\n",
      "Benchmarking experiment results comparing Multistep Instill VDP, OpenAI,\n",
      "Instructor, Marvin, BAML, TypeChat, and LangChain on reasoning and structured\n",
      "output generation.\n",
      "\n",
      "As can be seen, the multistep Instill VDP pipeline was able to consistently\n",
      "**achieve both correct reasoning and structured output where other techniques**\n",
      "**failed**.\n",
      "\n",
      "Despite requiring two LLM inferences, it is worth noting that the cost of this\n",
      "multistep pipeline approach will likely still be less than many of the\n",
      "structuring output tools that have been considered, as these often rely on\n",
      "making repeat API calls with modified prompts until the output can be\n",
      "successfully parsed into the required data model.\n",
      "\n",
      "### [\\#](http://www.instill.tech\\#try-it-yourself) Try it Yourself!\n",
      "\n",
      "Click the button below to view and try this pipeline out for yourself - **for free** \\- on the Instill Cloud platform!\n",
      "\n",
      "[Try on Instill Cloud ☁️](https://instill.tech/george_strong/pipelines/multistep-structured-output/playground)\n",
      "\n",
      "## [\\#](http://www.instill.tech\\#conclusion) Conclusion\n",
      "\n",
      "The article underscores the complexity of generating structured outputs from\n",
      "LLMs, particularly when reasoning tasks are involved. The new Structured Outputs\n",
      "in the API feature from OpenAI is a significant advancement, but it's not a\n",
      "silver bullet. As our results show, even the most advanced models and techniques\n",
      "can falter under strict output format constraints.\n",
      "\n",
      "The multi-step approach built with Instill VDP provides a practical solution to\n",
      "these challenges. By isolating the reasoning process from the structuring\n",
      "process, it allows LLMs to perform complex tasks without sacrificing accuracy or\n",
      "output quality. For businesses and developers looking to harness the full\n",
      "potential of LLMs in applications requiring structured data, this approach\n",
      "offers a reliable and cost-effective path forward.\n",
      "\n",
      "Stay tuned for the next blog post, where we'll show how you can integrate\n",
      "structured output generation with existing databases and applications using\n",
      "Instill VDP! 🚀\n",
      "\n",
      "### Similar Articles\n",
      "\n",
      "[![The theme of this tutorial](http://www.instill.tech/blog-assets/modular-ai/modular_architecture.png)\\\n",
      "![Building Better AI: How Modularity Shapes Successful AI Solutions's author's github avatar](http://www.instill.tech/author-avatar/george.jpeg)\\\n",
      "\\\n",
      "George Strong\\\n",
      "\\\n",
      "July 18, 2024\\\n",
      "\\\n",
      "Insight\\\n",
      "\\\n",
      "**Building Better AI: How Modularity Shapes Successful AI Solutions** \\\n",
      "Learn why modularity is vital for success in the dynamic landscape of AI, and how we embrace this principle at Instill AI.](http://www.instill.tech/blog/modularity-in-ai) [![The theme of this tutorial](http://www.instill.tech/blog-assets/unstructured-data-etl/lego-theme-thumbnail.png)\\\n",
      "![Missing Piece in Modern Data Stack: Unstructured Data ETL's author's github avatar](http://www.instill.tech/author-avatar/xiaofei-du.png)\\\n",
      "\\\n",
      "Xiaofei Du\\\n",
      "\\\n",
      "February 03, 2022\\\n",
      "\\\n",
      "Insight\\\n",
      "\\\n",
      "**Missing Piece in Modern Data Stack: Unstructured Data ETL** \\\n",
      "To fix the broken value chain in the modern data stack for unstructured data, we propose integrating with a component, named unstructured data ETL.](http://www.instill.tech/blog/unstructured-data-etl)\n",
      "\n",
      "## On this page\n",
      "\n",
      "- [The Complexities of Structuring Outputs from LLMs](http://www.instill.tech#the-complexities-of-structuring-outputs-from-llms)\n",
      "- [1\\. Prompt Engineering](http://www.instill.tech#1-prompt-engineering)\n",
      "- [2\\. Output Parsing](http://www.instill.tech#2-output-parsing)\n",
      "- [3\\. Error Handling](http://www.instill.tech#3-error-handling)\n",
      "- [Benchmarking Structured Output Tools with Reasoning Tasks](http://www.instill.tech#benchmarking-structured-output-tools-with-reasoning-tasks)\n",
      "- [The Task: Combining Reasoning and Structured Output](http://www.instill.tech#the-task-combining-reasoning-and-structured-output)\n",
      "- [Existing Tools for Structured Output Generation](http://www.instill.tech#existing-tools-for-structured-output-generation)\n",
      "- [Experiment Setup](http://www.instill.tech#experiment-setup)\n",
      "- [Results](http://www.instill.tech#results)\n",
      "- [Key Insights](http://www.instill.tech#key-insights)\n",
      "- [Can Chain-of-Thought Reasoning Help?](http://www.instill.tech#can-chain-of-thought-reasoning-help)\n",
      "- [The Solution: A Multistep Pipeline Built with 💧 Instill VDP](http://www.instill.tech#the-solution-a-multistep-pipeline-built-with--instill-vdp)\n",
      "- [Try it Yourself!](http://www.instill.tech#try-it-yourself)\n",
      "- [Conclusion](http://www.instill.tech#conclusion)\n",
      "\n",
      "## Contribution\n",
      "\n",
      "[Provide feedback for this page](https://github.com/instill-ai/instill-core/issues/new/choose)\n",
      "\n",
      "![blurred spot](http://www.instill.tech/images/bottom-cta/3.svg)![beam](http://www.instill.tech/images/bottom-cta/2.svg)\n",
      "\n",
      "![line](http://www.instill.tech/images/bottom-cta/1.svg)\n",
      "\n",
      "AI infrastructure for Enterprise\n",
      "\n",
      "[Get Started](https://instill.tech/login) [Book a Demo](https://cal.com/patrick-sheridan-5oacrl/30min)\n"
     ]
    }
   ],
   "source": [
    "print(response[\"outputs\"][0][\"scraped-content\"])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 2. Structured Web Insights from Google Search\n",
    "\n",
    "Here, we will view and try out the [structured-web-insights](https://instill.tech/instill-ai/pipelines/structured-web-insights/playground) pipeline. Checkout the pipeline preview page and README in the IFrame below to see the structure of this pipeline and understand how it works!\n",
    "\n",
    "As explained in Step 1 of the [tutorial](https://www.instill.tech/blog/mastering-openai-structured-outputs), you will need to configure your Google Search and API key first."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "\n",
       "        <iframe\n",
       "            width=\"1000\"\n",
       "            height=\"700\"\n",
       "            src=\"https://instill.tech/instill-ai/pipelines/structured-web-insights/preview\"\n",
       "            frameborder=\"0\"\n",
       "            allowfullscreen\n",
       "            \n",
       "        ></iframe>\n",
       "        "
      ],
      "text/plain": [
       "<IPython.lib.display.IFrame at 0x106bdb610>"
      ]
     },
     "execution_count": 10,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "IFrame('https://instill.tech/instill-ai/pipelines/structured-web-insights/preview', width=1000, height=700)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [],
   "source": [
    "response = pipeline.trigger(\n",
    "    namespace_id=\"instill-ai\",\n",
    "    pipeline_id=\"structured-web-insights\",\n",
    "    data=[{\"search-query\": \"What does Instill AI do?\",\n",
    "           \"google-cse-id\": \"INSERT YOUR GOOGLE CSE ID HERE\"}], # Insert your Google CSE ID here\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "{'author': 'Not available',\n",
      " 'key_insights': ['Instill AI provides a comprehensive solution for making '\n",
      "                  'unstructured data AI-ready.',\n",
      "                  'The platform includes features for data cleaning, '\n",
      "                  'transformation, and integration, making it easier to manage '\n",
      "                  'complex data flows.',\n",
      "                  'It offers an API-first approach, allowing developers to '\n",
      "                  'easily retrieve relevant data for AI applications.',\n",
      "                  'Users can switch between different AI models and vendors, '\n",
      "                  'optimizing costs and performance.',\n",
      "                  'The platform is designed to be secure and enterprise-ready, '\n",
      "                  'with options for cloud-native or on-premises deployment.'],\n",
      " 'published_date': 'Not available',\n",
      " 'source': 'Instill AI website (https://www.instill.tech/)',\n",
      " 'summary': 'Instill AI offers a full-stack AI solution designed to streamline '\n",
      "            'data operations and make unstructured data AI-ready. It provides '\n",
      "            'tools for data transformation, retrieval, and model management, '\n",
      "            'enabling businesses to efficiently utilize their data for AI '\n",
      "            'applications. The platform addresses common challenges in data '\n",
      "            'handling, such as data cleaning, integration, and scaling, while '\n",
      "            'ensuring security and compliance.',\n",
      " 'tags': ['AI solutions',\n",
      "          'unstructured data',\n",
      "          'data transformation',\n",
      "          'API',\n",
      "          'enterprise AI',\n",
      "          'data cleaning',\n",
      "          'model management',\n",
      "          'cloud-native',\n",
      "          'security',\n",
      "          'automation',\n",
      "          'workflow automation'],\n",
      " 'title': 'Instill AI',\n",
      " 'url': 'https://www.instill.tech/'}\n"
     ]
    }
   ],
   "source": [
    "response = MessageToDict(response)\n",
    "\n",
    "structured_output = response[\"outputs\"][0][\"structured-output\"]\n",
    "pprint(structured_output)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Instill AI\n",
      "\n",
      "![beam](http://www.instill.tech/images/landing-page/hero/1.svg)\n",
      "\n",
      "![beam](http://www.instill.tech/images/landing-page/hero/3.svg)![beam](http://www.instill.tech/images/landing-page/hero/4.svg)\n",
      "\n",
      "![beam](http://www.instill.tech/images/landing-page/hero/2.svg)![beam](http://www.instill.tech/images/landing-page/hero/4.svg)![beam](http://www.instill.tech/images/landing-page/hero/8.svg)![beam](http://www.instill.tech/images/landing-page/hero/10.svg)\n",
      "\n",
      "![beam](http://www.instill.tech/images/landing-page/hero/9.svg)![beam](http://www.instill.tech/images/landing-page/hero/8.svg)![beam](http://www.instill.tech/images/landing-page/hero/10.svg)\n",
      "\n",
      "![blurred spot](http://www.instill.tech/images/landing-page/3.svg)![blurred spot](http://www.instill.tech/images/landing-page/4.svg)![beam](http://www.instill.tech/images/landing-page/1.svg)![beam](http://www.instill.tech/images/landing-page/2.svg)![blurred spot](http://www.instill.tech/images/landing-page/5.svg)![blurred spot](http://www.instill.tech/images/landing-page/6.svg)\n",
      "\n",
      "![beam](http://www.instill.tech/images/landing-page/7.svg)![beam](http://www.instill.tech/images/landing-page/8.svg)\n",
      "\n",
      "![beam](http://www.instill.tech/images/landing-page/9.svg)\n",
      "\n",
      "![beam](http://www.instill.tech/images/landing-page/10.svg)![beam](http://www.instill.tech/images/landing-page/12.svg)![beam](http://www.instill.tech/images/landing-page/11.svg)\n",
      "\n",
      "ExploreFor BusinessFor DevelopersCompany [Pricing](http://www.instill.tech/pricing)\n",
      "\n",
      "Get Started\n",
      "\n",
      "Make Your Data\n",
      "\n",
      "AI-Ready\n",
      "\n",
      "Meet Instill Core, a source-available full-stack AI solution for tech teams—streamline your data operations and unlock the potential of unstructured data.All with one line of code.\n",
      "\n",
      "[Get Started](https://instill.tech/login) [GitHub](https://github.com/instill-ai/instill-core)\n",
      "\n",
      "Docker Compose\n",
      "\n",
      "Helm Chart\n",
      "\n",
      "```\n",
      "\n",
      "```\n",
      "\n",
      "Status Quo\n",
      "\n",
      "Building a basic RAG pipeline is easy, but preprocessing unstructured data and always retrieving highly relevant results, takes effort.\n",
      "\n",
      "Complex Data Cleaning\n",
      "\n",
      "80% of enterprise data is unstructured in unusable format and difficult to clean.\n",
      "\n",
      "Broken or Rigid Data Flow\n",
      "\n",
      "Data flows for AI can be very complex, requiring rich integrations and and customized automation.\n",
      "\n",
      "Hallucinations in answers\n",
      "\n",
      "Stuck with LLMs that produce inaccurate, irrelevant or even fabricated answers.\n",
      "\n",
      "Struggling to Scale\n",
      "\n",
      "You're bogged down by infrastructure maintenance work that has little business impact.\n",
      "\n",
      "Expensive\n",
      "\n",
      "It's tough to control budgets and get quick results with complex AI systems.\n",
      "\n",
      "Fragmented Tools\n",
      "\n",
      "Your data, models, workflows and operations are fragmented across a galaxy of tools.\n",
      "\n",
      "Solution\n",
      "\n",
      "The World's First Full-stack AI solution\n",
      "\n",
      "Making your data RAG-ready out-of-the-box\n",
      "\n",
      "[GitHub](https://github.com/instill-ai/instill-core) [Discover Stack](http://www.instill.tech/docs/welcome)\n",
      "\n",
      "Make your data AI-ready\n",
      "\n",
      "Transform Unstructured Data into AI-Ready Augmented Data Catalog\n",
      "\n",
      "**Instill Catalog** easily converts documents, images, audio, and video into a unified format. It simplifies data cleaning, reduces errors, automates updates, and ensures your data is RAG-ready for AI applications. With built-in data lineage support, you can track data origins, transformations, and flows, ensuring transparency and maintaining data integrity.\n",
      "\n",
      "Text\n",
      "\n",
      "Image\n",
      "\n",
      "Audio\n",
      "\n",
      "Video\n",
      "\n",
      "PDF\n",
      "\n",
      "JSON\n",
      "\n",
      "CSV\n",
      "\n",
      "More\n",
      "\n",
      "![Data transform chart](http://www.instill.tech/images/landing-page/full-stack-solution/slides/desktop/1.svg)\n",
      "\n",
      "[How it works](http://www.instill.tech/docs/artifact/upload-files)\n",
      "\n",
      "```\n",
      "\n",
      "```\n",
      "\n",
      "API-first & Developer-friendly\n",
      "\n",
      "Precise Retrieval Grounded in Your Data via Simple APIs\n",
      "\n",
      "Retrieve relevant results grounded in your data via simple APIs provided by Instill Catalog. Ideal for developers building intelligent search and Q&A services, such as AI assistants, with no deep technical expertise in LLM or RAG required.\n",
      "\n",
      "![Data transform chart](http://www.instill.tech/images/landing-page/full-stack-solution/slides/desktop/2.svg)\n",
      "\n",
      "[How it works](http://www.instill.tech/docs/artifact/search)\n",
      "\n",
      "```\n",
      "\n",
      "```\n",
      "\n",
      "Flexible switching of models across vendors\n",
      "\n",
      "Switch Models Across Vendors or Run Any Open-source Model to Save Costs\n",
      "\n",
      "Flexibly use state-of-the-art models across different vendors, or you can run any open-source AI models, automatically scale up and down, ensuring reliable compute resources without manual maintenance. Our high-performance platform serves your AI models with ease. **From Setup to Scaling, We've Got Infra Covered.**\n",
      "\n",
      "[How it works](http://www.instill.tech/docs/model/introduction)\n",
      "\n",
      "![Data transform chart](http://www.instill.tech/images/landing-page/full-stack-solution/slides/desktop/5.svg)\n",
      "\n",
      "Model Vendors\n",
      "\n",
      "![OpenAI logo](http://www.instill.tech/assets/openai.svg)OpenAI\n",
      "\n",
      "![Anthropic logo](http://www.instill.tech/assets/anthropic.svg)Anthropic\n",
      "\n",
      "![Cohere logo](http://www.instill.tech/assets/cohere.svg)Cohere\n",
      "\n",
      "![Mistral AI logo](http://www.instill.tech/assets/mistral-ai.svg)Mistral AI\n",
      "\n",
      "![Hugging Face logo](http://www.instill.tech/assets/hugging-face.svg)Hugging\n",
      "\n",
      "Face\n",
      "\n",
      "More\n",
      "\n",
      "Instill Model\n",
      "\n",
      "🌋LLaVa\n",
      "\n",
      "![Meta logo](http://www.instill.tech/assets/meta.svg)Llama 3\n",
      "\n",
      "![Meta logo](http://www.instill.tech/assets/meta.svg)Llamacode\n",
      "\n",
      "More\n",
      "\n",
      "Unstructured data ETL\n",
      "\n",
      "Orchestrate pipelines to pre-process and transform data to keep it relevant\n",
      "\n",
      "Fix broken data flows and improve data quality to your AI applications. Create DAGs of dependencies, connect third-party data sources, AI models, operations, and applications. Automate ETL processes to manage and orchestrate the entire pipeline in a single interface, powering your applications and business efficiently.\n",
      "\n",
      "![GCS](http://www.instill.tech/assets/gcs.svg)GCS\n",
      "\n",
      "![BigQuery](http://www.instill.tech/assets/bigquery.svg)BigQuery\n",
      "\n",
      "![Gmail](http://www.instill.tech/assets/gmail.svg)Gmail\n",
      "\n",
      "![HubSpot](http://www.instill.tech/assets/hubspot.svg)HubSpot\n",
      "\n",
      "![Slack logo](http://www.instill.tech/assets/slack.svg)Slack\n",
      "\n",
      "[See integrations](http://www.instill.tech/integrations)\n",
      "\n",
      "![Data transform chart](http://www.instill.tech/images/landing-page/full-stack-solution/slides/desktop/3.svg)\n",
      "\n",
      "One Platform. Run Anywhere. Zero Hassle.\n",
      "\n",
      "Forget about gluing different tools together. Our platform integrates seamlessly with your systems, making scaling your AI applications effortless. It streamlines data handling, enhances AI accuracy, continuously optimizes through monitoring, and effectively manages infrastructure. This allows you to expand your AI applications efficiently and cost-effectively. Enjoy native compatibility at every stage for hassle-free maintenance.\n",
      "\n",
      "[How it works](http://www.instill.tech/docs/core/introduction)\n",
      "\n",
      "![Data transform stages](http://www.instill.tech/images/landing-page/full-stack-solution/slides/desktop/4.svg)![Data transform stages](http://www.instill.tech/images/landing-page/full-stack-solution/slides/mobile/4.svg)\n",
      "\n",
      "Secure and Enterprise-ready\n",
      "\n",
      "Meet Security and Operational Requirements to Ship AI Faster\n",
      "\n",
      "Cloud-native\n",
      "\n",
      "Fully managed on your choice of public cloud, Bring Your Own Cloud (BYOC), or on-premises.\n",
      "\n",
      "![Instill Cloud logo](http://www.instill.tech/assets/instill-cloud.svg)Instill Cloud\n",
      "\n",
      "![Google Cloud logo](http://www.instill.tech/assets/google-cloud.svg)Google Cloud\n",
      "\n",
      "![Amazon Web Services logo](http://www.instill.tech/assets/aws.svg)Amazon Web Services\n",
      "\n",
      "![Microsoft Azure logo](http://www.instill.tech/assets/azure.svg)Microsoft Azure\n",
      "\n",
      "On Premises\n",
      "\n",
      "Security & Privacy\n",
      "\n",
      "Control your data securely with TLS encryption and strict retention policies.\n",
      "\n",
      "![SOC 2 logo](http://www.instill.tech/assets/aicpa-soc.png)SOC 2\n",
      "\n",
      "![GDPR logo](http://www.instill.tech/assets/gdpr.png)GDPR\n",
      "\n",
      "Want results like these?\n",
      "\n",
      "[Book a Demo](https://cal.com/patrick-sheridan-5oacrl/30min)\n",
      "\n",
      "10X\n",
      "\n",
      "Faster to develop and ship your AI applications\n",
      "\n",
      "30%\n",
      "\n",
      "Capacity boost to break down data and AI team silos\n",
      "\n",
      "$2M\n",
      "\n",
      "Saved on R&D budget for more efficient resource use\n",
      "\n",
      "Use Cases\n",
      "\n",
      "TransformingUse CasesAcross Sectors\n",
      "\n",
      "Find out how your business can benefit from AI\n",
      "\n",
      "Customer Support\n",
      "\n",
      "AI Assistant for Customer Support\n",
      "\n",
      "AI Customer Assistant provides 24/7 support, automates inquiries, and personalizes recommendations.\n",
      "\n",
      "AI Assistant\n",
      "\n",
      "Workflow Automation\n",
      "\n",
      "AI Document Analysis Assistant\n",
      "\n",
      "Upload any Document and engage in AI-powered conversations.\n",
      "\n",
      "AI Assistant\n",
      "\n",
      "Agriculture\n",
      "\n",
      "Agricultural Biotechnology AI Image Analyzer\n",
      "\n",
      "Automatically detect, classify, and analyze visual data with high precision and efficiency.\n",
      "\n",
      "Computer Vision\n",
      "\n",
      "Workflow Automation\n",
      "\n",
      "Healthcare\n",
      "\n",
      "AI Infection Diagnosis Assistant\n",
      "\n",
      "Automate data extraction, analysis, and interpretation for medical diagnosis.\n",
      "\n",
      "AI Assistant\n",
      "\n",
      "Finance\n",
      "\n",
      "AI Market Analysis Assistant\n",
      "\n",
      "Predictive Market Intelligence to extract trends, generate insights, and receive tailored market strategies.\n",
      "\n",
      "AI Assistant\n",
      "\n",
      "Workflow Automation\n",
      "\n",
      "Workflow Automation\n",
      "\n",
      "AI-powered Workflow Automation\n",
      "\n",
      "Office Workflow Automation streamlines various administrative tasks.\n",
      "\n",
      "Workflow Automation\n",
      "\n",
      "[Explore More](http://www.instill.tech/use-cases)\n",
      "\n",
      "![blurred spot](http://www.instill.tech/images/bottom-cta/3.svg)![beam](http://www.instill.tech/images/bottom-cta/2.svg)\n",
      "\n",
      "![line](http://www.instill.tech/images/bottom-cta/1.svg)\n",
      "\n",
      "AI infrastructure for Enterprise\n",
      "\n",
      "[Get Started](https://instill.tech/login) [Book a Demo](https://cal.com/patrick-sheridan-5oacrl/30min)\n"
     ]
    }
   ],
   "source": [
    "print(response[\"outputs\"][0][\"scraped-content\"])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [],
   "source": [
    "pipeline.close()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "instill",
   "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": "3.11.9"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}