Unstract: AI Document Parser: Extract Data from Complex PDFs + LLM Challenge! (Opensource)

Key Concepts

Unstructured data extraction, Large Language Models (LLMs), LLM Challenge, Hallucination avoidance, Open-source platform, Prompt Studio, Token Calculator, LLM Whisper, Document parsing, Data accuracy, Automation, JSON output, API metadata, Cost estimation.

Dealing with Unstructured Data

• The Problem: Working with unstructured data is frustrating due to the time-consuming manual effort required to sift through and organize it.
• The Solution: Extract valuable insights, improve data accessibility, and enable automation for reports, workflows, and transactions.
• Unstruct Introduction: Unstruct is an open-source, no-code platform designed for large language model (LLM) powered unstructured data extraction.
• Process:
  1. Upload a file (CSV, PDF, etc.).
  2. Specify prompts for the desired data (e.g., issuer name, address).
  3. Receive a combined JSON output with the extracted data.
• Benefits: Simplifies working with unstructured data, allowing users to focus on core tasks.
• Capabilities: Handles various document formats without manual annotations, processes bank statements from 200+ banks, and adapts to different layouts.

Token Calculator

• Function: A free tool within Unstruct's suite that calculates token usage and API costs across 400+ LLMs.
• Process:
  1. Select a provider and model.
  2. Upload a document or paste text.
  3. View the total tokens used, input cost, and output cost.

Large Language Model (LLM) Challenge

• Purpose: To increase extraction accuracy and avoid hallucination (AI generating false information).
• Methodology: Uses two LLMs: one to extract data and another to double-check it. If they disagree, the result is set to null.
• Benefits: Ensures more accurate data extraction, especially in production environments where reliability is critical (e.g., legal, finance, compliance).
• Implementation:
  1. Enable LLM Challenge in the settings tab.
  2. Select a challenger LLM (e.g., GPT-4, OAI).
  3. Create a workflow in the Prompt Studio and upload a document.
  4. Run LLMs to parse the desired information.
  5. Export the results as a tool.
  6. Select the tool in workflows, enable LLM Challenge, and save.
• Operation: LLM Challenge automatically runs during extraction, validating the output of the primary LLM.
• API Integration: The include_metadata parameter provides extra details, including LLM logs and cost, accessible under the challenge_data file.
• Transparency: Provides scores from the challenger LLM, indicating confidence in the extracted values.
• Cost Tracking: API metadata includes token usage and cost estimates for the challenger LLM.

LLM Whisper

• Function: An AI document parser designed to revolutionize complex PDF data extraction.
• Capabilities:
  • Processes and prepares complex documents for optimal LLM processing.
  • Connects to a vector database.
  • Preserves layouts.
  • Accurately handles checkboxes and radio buttons.
  • Optimizes token usage with auto-compaction.
  • Automatic mode switching to OCR.
  • Pre-processing customization and flexible deployment.
• Supported File Types: PDFs, images.
• Output Modes: Multiple output modes.
• Benefits: Ensures high-quality, precise, and cost-efficient results for LLM tasks.
• Example: Accurately extracts data, including radio button selections (e.g., "I am applying for an individual credit"), from complex contract reports.
• Application: Can automate data entry for accountants by accurately listing transactions from account activity reports.

System Requirements and Setup

• Unstruct can be installed locally if the system meets the specified requirements.
• A setup guide is available to access the Prompt Studio.

Conclusion

Unstruct is a powerful, open-source platform that leverages LLMs to simplify unstructured data extraction. The LLM Challenge feature enhances accuracy and reduces hallucination, making it suitable for production environments. LLM Whisper further optimizes document parsing, ensuring high-quality results for various file types. The platform offers tools for cost estimation and provides transparency through detailed metadata, empowering users to efficiently manage and utilize unstructured data.