WebGLM is an efficient web-enhanced question-answering system designed for researchers and developers seeking to integrate real-time web search and retrieval into large language models. It leverages the 10-billion-parameter General Language Model (GLM) to provide accurate and human-like answers by augmenting it with web search capabilities.

How It Works

WebGLM employs a three-pronged approach: an LLM-augmented Retriever to fetch relevant web content, a bootstrapped Generator that uses the GLM to formulate answers based on retrieved information, and a Human Preference-aware Scorer to evaluate and prioritize response quality. This architecture aims to improve the efficiency and cost-effectiveness of deploying QA systems in real-world scenarios.

Quick Start & Requirements

• Install: pip install -r requirements.txt, playwright install
• Prerequisites: Python, Node.js, Playwright dependencies, SerpAPI key (or use --searcher bing with Playwright). Retriever checkpoint download required.
• Setup: Requires downloading checkpoints and potentially training data.
• Links:
  • Paper: KDD'23
  • HF Repo: WebGLM-10B, WebGLM-2B
  • Dataset: WebGLM-QA
  • Demo: Link

Highlighted Details

• Supports both WebGLM-10B and WebGLM-2B models.
• Integrates with SerpAPI for web search or can use Bing via Playwright.
• Provides CLI and web service interfaces for easy deployment.
• Includes scripts for training the generator and retriever components.
• Offers evaluation scripts for reproducing benchmark results.

Maintenance & Community

• Developed by THUDM.
• Recent updates include Bing search integration and release of ChatGLM2-6B (though not directly part of WebGLM, it indicates active development in the GLM ecosystem).

Licensing & Compatibility

• Repository licensed under Apache-2.0.
• Model weights subject to Model_License.
• Open-sourced data is for research purposes only.

Limitations & Caveats

The README mentions that the current version of ChatGLM2-6B (related to the GLM family) has limited understanding of single-round ultra-long documents, which is a focus for future optimization. The use of open-sourced data is restricted to research purposes.