NemoClaw Knowledge Wiki

Langextract - Prompt Engineer channel

4 min read

Langextract - Prompt Engineer channel

https://www.youtube.com/watch?v=dPL2vRDunMw The video introduces LangExtract, a new open-source project from Google that is a Gemini-powered information extraction library. [0:00] Its primary purpose is to convert unstructured text data into structured data. [0:06] Key Features and Capabilities:

  • Custom Schema: LangExtract allows users to define custom schemas for information extraction, enabling the extraction of specific information they are looking for. [0:13]
  • Visualization: It provides a nice visualization of the extracted information in an interactive HTML document. [0:20, 2:54]
  • Open Source: LangExtract is an open-source Python package available on GitHub. [0:24, 0:49]
  • LLM Backend Flexibility: It supports both cloud-based Large Language Models (LLMs) like Google’s Gemini family and open-source on-device models. [1:06]
  • Reliable Structured Outputs: Users can define their desired output using LangExtract’s data representation and provide “few-shot” examples. LangExtract leverages “Controlled Generation” in supported models like Gemini to ensure consistently structured outputs. [0:50]
  • Optimized Long-Context Information Extraction: It handles information retrieval from large documents by using a chunking strategy, parallel processing, and multiple extraction passes over smaller, focused contexts. [1:05]
  • Flexible Across Domains: It allows defining information extraction tasks for any domain with just a few well-chosen examples, without needing fine-tuning of an LLM. [1:05]
  • Utilizing LLM World Knowledge: Beyond grounded entities, LangExtract can leverage an LLM’s world knowledge to supplement extracted information, including implicit or inferred data. [1:06]

Usage and Examples: The video demonstrates how to use LangExtract with Python.

  1. Installation: Install the langextract package using pip install langextract. [1:21]
  2. Defining Prompt and Examples: Define a concise prompt describing what to extract and provide high-quality “few-shot” examples to guide the model. [1:25] The examples consist of ExampleData objects, each with text and extractions. [1:50] extractions define the extraction_class (e.g., “character”, “emotion”, “relationship”), extraction_text, and optional attributes (e.g., {“emotional_state”: “wonder”}). [2:03]
  3. Running Extraction: The lx.extract() function is called with input_text, prompt_description, examples, model_id (e.g., gemini-2.5-pro), and api_key. [2:31]
  4. Displaying Results: The extracted entities can be printed to the console, showing the class, text, and character positions. [4:29]
  5. Generating Visualization: An HTML file is generated using lx.visualization.visualize() to visually highlight extracted entities in context. [4:37]

Practical Examples Demonstrated:

  • Basic Entity Extraction (Apple Inc. Announcement): Extracts company, person, product, date, location, and price from a sample news text. [3:23]
  • Financial Report Extraction (TechCorp Quarterly Earnings Report): Demonstrates extracting financial and business entities, including company information, stock symbols, financial metrics (revenue, income, EPS, margins), M&A activities, product launches, customer metrics, executive quotes, and forward guidance. This example also shows two levels of data: the entity and its corresponding attributes (e.g., ticker for company, period/value/change for financial metrics). [5:48]
  • Competitive Intelligence Extraction (News Article): Extracts competitive intelligence like companies and their AI initiatives, investment and funding rounds, partnerships, product announcements, and key personnel. This also includes nested attributes for funding events and competitive moves. [6:33]
  • Customer Feedback Analysis (Product Reviews): Extracts customer feedback insights including customer segments, company names, product features (positive/negative), pricing, and quantitative metrics (ratings, savings). [6:46]
  • Advanced Relationship Extraction (Clinical Note): This is highlighted as a main power of LangExtract. It extracts medication information (name, dosage, route, frequency, duration) and their relationships to medical conditions. [7:03] It uses extraction_passes=2 to tell the LLM to process the information multiple times for better relationship detection, which can improve accuracy. [9:07] The output can be visualized as a knowledge graph showing relationships between conditions (red nodes), medications (green nodes), and patients (blue node). [10:13] For example, it shows which medication “treats” which condition. [10:18]

Disclaimer: The video explicitly states that LangExtract is not an officially supported Google product. It is an open-source project from some Googlers. [11:30] Usage is subject to the Apache 2.0 License, and for health-related applications, it is also subject to the Health AI Developer Foundations Terms of Use. [11:31]

Graph View