"This notebook demonstrates the complete workflow for processing documents using LlamaCloud services:\n",
"1. **Parse** - Extract and convert documents to markdown\n",
"<a href=\"https://colab.research.google.com/github/run-llama/llama_cloud_services/blob/main/examples/misc/parse_classify_extract_workflow.ipynb\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a>\n",
"\n",
"This notebook shows a multi-step agentic document workflow that uses the **parsing**, **classification** and **extraction** modules in LlamaCloud, orchestrated through **LlamaIndex Workflows**. The workflow can take in a complex input document, parse it into clean markdown, classify it according to its subtype, and extract data according to a specified schema for that subtype. This allows you to automate document extraction of various types within the same workflow instead of having to manually separate the data beforehand. \n",
"\n",
"This notebook uses the following modules:\n",
"1. **Parse (LlamaParse)** - Extract and convert documents to markdown\n",
"2. **Classify** - Categorize documents based on their content\n",
"3. **Extract** - Extract structured data using the markdown as input via SourceText\n",
"3. **Extract (LlamaExtract)** - Extract structured data using the markdown as input via SourceText\n",
"4. **LlamaIndex Workflows** - Event-driven orchestration of the parse, classify and extract steps\n",
"\n",
"## Overview of the Workflow\n",
"\n",
"### 1. Parse Phase\n",
"- Use `LlamaParse` to convert documents (PDFs, Word docs, etc.) into structured formats\n",
"- Extract markdown content that preserves document structure\n",
"- Get both raw text and markdown representations\n",
"\n",
"### 2. Classify Phase\n",
"- Use `ClassifyClient` to categorize documents based on content\n",
"- Apply classification rules to route documents appropriately\n",
"- Handle different document types with specific processing logic\n",
"\n",
"### 3. Extract Phase\n",
"- Use `LlamaExtract` with `SourceText` to extract structured data\n",
"- Pass the markdown content as input for more accurate extraction\n",
"- Define custom schemas for structured data extraction\n",
"\n",
"Let's walk through each step with practical examples."
"The workflow is implemented as a proper LlamaIndex Workflow with separate steps for parsing, classification, and extraction, connected by typed events. This provides modularity, observability, and type safety."
"print(f\"📝 Created {len(classification_rules)} classification rules\")"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"### Try Classification Independently\n",
"\n",
"Let's test the classification on one of our parsed documents to see how it works:\n"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"🔍 Classifying financial document...\n",
" Document length: 1,338,499 characters\n",
"\n",
"✅ Classification Result:\n",
" Type: financial_document\n",
" Confidence: 100.00%\n",
" Reasoning: This document is a Form 10-K, which is an annual report required by the U.S. Securities and Exchange Commission (SEC) for publicly traded companies. It contains financial data, information about the c...\n",
"Let's create a function that demonstrates the complete workflow:"
"Now that we've seen how parsing works, let's build a complete 3-step workflow (Parse → Classify → Extract) using LlamaIndex Workflows. We'll define the workflow structure here, and you can see it in action below where we also demonstrate the classification and extraction modules independently.\n",
"\n",
"### Install Workflows Package\n",
"\n",
"First, let's install the LlamaIndex workflows package:\n"
"The workflow has been visualized and saved to `document_workflow.html`. You can open this file in a browser to see the interactive workflow diagram.\n"
" • key_business_segments: ['Mobility', 'Delivery', 'Freight', 'All Other (including former New Mobility, e-bikes, e-scooters, Advanced Technologies Group and other technology programs)']\n",
" • risk_factors: [\"The company faces numerous risk factors across its business operations and environment. The COVID-19 pandemic and related mitigation measures have adversely affected parts of the business, including reduced demand for Mobility offerings and creating ongoing uncertainties. The company's operational and financial performance is influenced by competitive pressure in the mobility, delivery, and logistics industries, characterized by well-established alternatives, low barriers to entry, and low switching costs. Driver classification risks exist if Drivers are deemed employees, workers, or quasi-employees rather than independent contractors, exposing the company to legal actions and financial liabilities globally. Competition challenges require the company to sometimes lower fares, offer incentives, and promotions, which impacts profitability. There are significant operating losses historically with substantial future operating expense increases anticipated, and the ability to achieve or maintain profitability is uncertain. Network value depends on maintaining critical mass among Drivers, consumers, merchants, shippers, and carriers, and failures to do so diminish platform attractiveness. Brand and reputation maintenance is critical, with exposure to negative publicity, media coverage, and risks from associated companies' brands or licensed brands in joint ventures.\\n\\nOperational risks include historical workplace culture and compliance challenges, management complexity due to rapid growth, technological infrastructure issues potentially causing disruptions or poor user experience, and security or data privacy breaches that could impact revenue and reputation. Platform users may engage in or be subjected to criminal, violent, or dangerous activity leading to safety incidents and legal actions. New offerings and technologies investments are inherently risky without guaranteed benefits. Economic conditions, inflation, and increased costs (fuel, food, labor, energy) may negatively impact results. Regulatory risks are extensive and global, involving payment and financial services compliance, licensing, anti-money laundering laws, data privacy (GDPR, CCPA, LGPD), and labor laws. Legal and regulatory investigations and inquiries, including antitrust, FCPA, labor classification, data protection, and intellectual property matters, pose risks of fines, penalties, operational changes, and increased costs.\\n\\nGeopolitical and jurisdictional risks include operating limitations or bans in some locations, currency exchange risk, and complex evolving regulations with the potential for fines and loss of licenses or permits. Insurance risks include potential inadequacy of reserves, liability exposure from accidents or impersonation, and insurer insolvency. Driver qualification requirements and background checks may increase costs or fail to expose all relevant information, with associated insurance cost risks and potential for courtroom or regulatory challenges to pricing models.\\n\\nFinancial risks comprise significant accumulated deficits, requirement for additional capital with uncertain availability, debt obligations, tax exposure including uncertain positions and observed changes in tax laws, and volatility in common stock price with no expected cash dividends. Accounting judgments and estimates involve critical assumptions affecting reported financial metrics related to goodwill, revenue recognition, incentive accruals, and stock-based compensation. Cybersecurity risks include exposures to malware, ransomware, phishing, and other cyberattacks. Climate change presents physical and transitional risks that may impact operations and costs, and failure to meet climate commitments may have operational and reputational consequences.\\n\\nOther risks include potential liability under anti-corruption and anti-terrorism laws, adverse effects from defaults under debt agreements, limitations in takeover actions due to corporate governance provisions, and the impact of non-GAAP financial measure limitations. Overall, these diverse and interconnected risk factors contribute to significant uncertainty regarding the company's future business prospects, operating results, and financial condition.\"]\n",
" • revenue_2021: $17,455 and $21,764\n",
" • net_income_2021: $(496) to (700)\n",
" • key_business_segments: ['Borrower and the Restricted Subsidiaries', 'Holdings', 'Guarantors', 'Material Domestic Subsidiaries', 'Material Foreign Subsidiaries']\n",
" • risk_factors: ['Indemnification obligations of the borrower for losses, claims, damages, liabilities, and out-of-pocket expenses incurred by agents, lenders, arrangers, and related parties in connection with the agreement or loans, except in certain cases such as gross negligence, bad faith, willful misconduct, or material breach by the indemnitee.', \"Borrower not required to indemnify any indemnitee for settlements entered into without the borrower's consent.\", 'Limitation of liability for special, indirect, consequential, or punitive damages, and for damages from unauthorized use of information, except for direct damages resulting from gross negligence, bad faith, or willful misconduct.', 'Obligation of the borrower to indemnify the administrative agent for liabilities arising from performance of duties, except in cases of gross negligence, bad faith, or willful misconduct.', 'Limitations and conditions on assignments and participations of lender rights, including restrictions on assignments to disqualified institutions, loan parties, affiliates of loan parties, defaulting lenders, and natural persons.', 'Setoff rights for lenders and issuing banks after an event of default, allowing them to apply borrower deposits toward obligations under the agreement.', 'Potential for increased obligations under the agreement as a result of changes in law affecting payment terms.', 'Requirement for the borrower and guarantors to provide information to comply with anti-money laundering rules and the USA PATRIOT Act.']\n",
" • description: The LM317 is an adjustable three-pin, positive-voltage regulator capable of supplying up to 1.5A over an output voltage range of 1.25V to 37V. It features line and load regulation, internal current limiting, thermal overload protection, and safe operating area compensation.\n",
" • part_number: LM317, SLVS044Z\n",
" • description: The LM317 is an adjustable three-pin, positive-voltage regulator capable of supplying more than 1.5A (typically up to 1.5A) over an output voltage range of 1.25V to 37V. The device requires only two external resistors to set the output voltage. It features a typical line regulation of 0.01% and typical load regulation of 0.1%. The LM317 includes current limiting, thermal overload protection, and safe operating area protection. Overload protection remains functional even if the ADJUST pin is disconnected. The regulator is used in applications such as constant-current battery-charger circuits, slow turn-on 15V regulator circuits, AC voltage-regulator circuits, current-limited charger circuits, and high-current and adjustable regulator circuits. It is available in packages including SOT-223 (DCY), TO-220 (KCS), and TO-263 (KTT).\n",
" • key_features: ['Adjustable output voltage: 1.25V to 37V', 'Output current up to 1.5A', 'Line regulation: 0.01%/V (typical)', 'Load regulation: 0.1% (typical)', 'Internal short-circuit current limiting', 'Thermal overload protection', 'Output safe-area compensation', 'High power supply rejection ratio (PSRR): 80dB at 120Hz (new chip)', 'Available in SOT-223, TO-263, and TO-220 packages']\n",
" • applications: ['Multifunction printers', 'AC drive power stage modules', 'Electricity meters', 'Servo drive control modules', 'Merchant network and server power supply units']\n",
" • maximum_current: 4.0\n",
" • key_features: ['Adjustable output voltage range: 1.25V to 37V', 'Output current up to 1.5A (up to 4A with external pass elements)', 'Line regulation: typically 0.01%/V', 'Load regulation: typically 0.1%', 'Internal short-circuit current limiting / Current limiting', 'Thermal overload protection / Thermal shutdown', 'Output safe-area compensation / Safe operating area protection', 'PSRR: 80dB at 120Hz for CADJ = 10μF (new chip)', 'NPN Darlington output drive', 'Programmable feedback', 'Multiple package options (SOT-223, TO-220, TO-263)', 'Can be used in constant-current, battery-charging, and regulator applications']\n",
" • applications: ['Multifunction printers, AC drive power stage modules, Electricity meters, Servo drive control modules, Merchant network and server PSU, Adjustable voltage regulator, 0V to 30V regulator circuit, Regulator circuit with improved ripple rejection, Precision current-limiter, Tracking preregulator, 1.25V to 20V regulator, Battery charger circuit, Constant-current battery charger circuits, Slow turn-on regulator, AC voltage-regulator, Current-limited charger circuits, High-current adjustable regulator circuits, General-purpose adjustable power supply']\n",
"\n",
"✨ Workflow completed successfully!\n",
"\n",
"📚 Key Learnings:\n",
" • Parse: Converted documents to clean markdown format\n",
"This notebook demonstrated the complete **Parse → Classify → Extract** workflow using LlamaCloud services:\n",
"The notebook shows you how to build an e2e document **Classify → Extract** workflow using LlamaCloud. This uses some of our core building blocks around **classification** interleaved with **document extraction**.\n",
" - Converts documents to clean, structured markdown\n",
@@ -715,38 +1104,17 @@
"\n",
"3. **LlamaExtract with SourceText** (`llama_cloud_services.extract.extract.LlamaExtract`, `SourceText`):\n",
" - Extracts structured data using custom Pydantic schemas\n",
" - **SourceText** allows using markdown content as input instead of raw files\n",
" - Provides much better extraction accuracy when using processed markdown\n",
" - You can either feed in the file directly (in which case parsing will happen under the hood), or the parsed text through the **SourceText** object (which is the case in this example) \n",
"\n",
"### Workflow Benefits:\n",
"\n",
"- **Better Accuracy**: Using markdown from parsing provides cleaner, more structured input for extraction\n",
"- **Automatic Routing**: Classification allows different processing logic for different document types\n",
"- **Structured Output**: Custom schemas ensure consistent, structured data extraction\n",
"- **Flexible Input**: SourceText supports text content, file paths, and bytes\n",
"\n",
"### Key Insights:\n",
"\n",
"1. **SourceText is the bridge**: It allows you to pass the clean markdown content from parsing directly to extraction\n",
"2. **Markdown improves extraction**: Pre-processed markdown provides much better context than raw PDFs\n",
"3. **Classification enables smart routing**: Different document types can use different extraction schemas\n",
"4. **End-to-end automation**: The entire workflow can be automated for production use\n",
"\n",
"This approach is ideal for production document processing pipelines where you need to:\n",
"- Process various document types automatically\n",
"- Extract structured data consistently\n",
"- Maintain high accuracy and reliability\n",
"- Handle documents at scale\n",
"\n",
"The combination of these three services provides a powerful, flexible document processing pipeline that can handle complex, real-world document processing requirements."
"**Benefits of an e2e workflow**: The main benefit of doing Classify -> Extract, instead of only Extract, is the fact that you can handle documents of different types/different expected schemas within the same workflow, without having to separate out the data before and running separate extractions on each data subset. "
"If `retrieve_image_nodes` is set to true, both `retrieve_page_screenshot_nodes` and `retrieve_page_figure_nodes` must also be set to true or omitted.",
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
Jerry Liu