NemoClaw Knowledge Wiki

Langextract Sam Witteveen

5 min read

https://www.youtube.com/watch?v=t-53fouKqWI The video introduces LangExtract, a new open-source Python library from Google, designed for information extraction from unstructured text using Gemini models. The speaker contrasts this with traditional Natural Language Processing (NLP) methods and the challenges of using large generative Large Language Models (LLMs) for such specific, non-generative tasks. Context and Problem Statement:

  • (0:23) Standard NLP tasks like text classification, sentiment analysis, Named Entity Recognition (NER), and disambiguation are not generative uses of LLMs.
  • (0:54) Historically, these tasks were often handled by BERT models, which are encoder-only transformers. BERT models were excellent for fine-tuning on specific downstream tasks and had a relatively small context window (e.g., 512 tokens).
  • (1:27) The original Transformer architecture is split into an encoder (BERT-like, for understanding/encoding) and a decoder (GPT-like, for generation). While decoder-only LLMs like GPT have seen rapid advancements, (3:50) the speaker argues they are often “too big, slow, private, and expensive” for many practical, discriminative NLP jobs like classification, retrieval, and entity extraction.
  • (4:20) A significant cost comparison is provided: filtering 15 trillion tokens cost 1 million, despite the lower per-token cost. This highlights the practical efficiency advantage of encoder-only models for certain workloads.

LangExtract’s Solution and Key Features:

  • (5:01) LangExtract is presented as a lightweight, open-source Python library designed to empower developers to programmatically and reliably extract precise, structured information.
  • (5:12) Its core features include: Precise source grounding: Every extracted entity is mapped back to its exact character offsets in the original text, allowing for traceability and visual verification. Reliable structured outputs: Users define their desired output using LangExtract’s data representation and provide “few-shot” examples, leveraging Controlled Generation in supported models like Gemini to ensure consistently structured results. This means defining a schema for the output. Optimized long-context information extraction: Handles information retrieval from very large documents (e.g., 100,000 words) by using a chunking strategy, parallel processing, and multiple extraction passes over smaller, focused contexts. Interactive visualization: Provides an easy way to generate self-contained HTML visualizations of extracted entities in context, useful for reviewing annotations. Flexible support for LLM backends: Works with preferred cloud-based LLMs (like Google’s Gemini family) or open-source on-device models (e.g., Llama, Gemma, Chinese models). Flexible across domains: Information extraction tasks for any domain can be defined with just a few well-chosen “few-shot” examples, eliminating the need for extensive fine-tuning of an LLM.

How it Works (Code Demonstration):

  • (7:51) The process starts with installing the langextract library. If using Gemini, an API key is required.
  • The user defines a concise prompt specifying what information to extract (e.g., “Extract characters, emotions, and relationships”).
  • (8:00) High-quality “few-shot” examples are provided. These examples include sample text and the expected structured extractions (e.g., character name “ROMEO” with emotional_state: "wonder").
  • (8:10) The lx.extract function is then called, passing in the input text to be processed, the prompt description, the defined examples, and the chosen Gemini model (e.g., gemini-2.5-pro or gemini-2.5-flash).
  • (8:27) The results can be saved to a JSONL file and interactively visualized as an HTML file, highlighting extracted entities directly in the text.

Practical Examples and Use Cases:

  • Shakespeare Example (7:51): Demonstrates basic entity and relationship extraction from famous lines from Romeo and Juliet.
  • Medical Reports (8:40): Shows how LangExtract can process clinical text to extract structured information like “condition,” “dosage,” “frequency,” and “medication,” and their relationships (e.g., medication_group).
  • RadExtract Demo (9:11): An interactive Hugging Face demo showcasing LangExtract for structuring radiology reports. It transforms unstructured report text into structured findings, with grounded highlights.
  • TechCrunch Article Analysis (9:33): The speaker demonstrates extracting people’s names, AI models, products, and company names from a TechCrunch article. (14:44) Using gemini-2.5-flash for efficiency, the extraction processed 11,412 characters, found 74 entities (4 unique types) in about 25 seconds. (15:05) The raw output is a complex structured object, but it can be easily parsed. The speaker shows filtering to print only person names (with associated companies) and company names (with character intervals). (16:28) Demonstrates using Python sets to get unique company names mentioned (e.g., Google DeepMind, Google, TechCrunch, AI2, OpenAI, Perplexity, xAI, Anthropic, Meta Superintelligence Labs). (17:09) Shows extraction of AI models and products, noting that o1 and Strawberry were classified as AI models, while ChatGPT and Codex were products. Some ambiguity is observed (e.g., ChatGPT classified as both product and AI model), which the speaker suggests could be improved by refining the prompt.

Conclusion:

  • (18:43) LangExtract allows users to define custom extraction tasks and output schemas, making it flexible for various domains like medicine, finance, and engineering.
  • (19:40) Overall, LangExtract is presented as a valuable and practical library for information extraction, offering an efficient way to extract structured data from unstructured text without the overhead of training bespoke models or relying solely on expensive, large generative LLMs. It enables quick prototyping and can be readily deployed in production environments.

Graph View