GABRIEL is an official OpenAI toolkit designed to empower social scientists and data scientists by transforming unstructured qualitative data—text, images, or audio—into quantifiable, analysis-ready datasets using the GPT API. It addresses the challenge of building robust LLM-driven analysis pipelines by abstracting complexities like prompt engineering, batching, retries, and checkpointing, allowing users to treat AI-assisted measurement as a reproducible scientific instrument.

How It Works

GABRIEL operationalizes large language models for attribute measurement and data manipulation by providing a structured framework for interacting with GPT APIs. Its core approach involves packaging user-defined attributes and context into effective prompts, managing API calls with built-in parallelism, retries, and checkpointing for scalability, and returning structured outputs like ratings, classifications, or extracted facts in tidy DataFrames. This design allows for human-level comprehension on demand, treating LLM analysis as a reliable measurement tool rather than ad-hoc scripting.

Quick Start & Requirements

• Installation: pip install openai-gabriel or directly from GitHub (pip install git+https://github.com/openai/GABRIEL.git@main).
• Prerequisites: An OpenAI API key must be set as the OPENAI_API_KEY environment variable.
• Resources: Requires API access and sufficient compute for processing.
• Documentation: A comprehensive tutorial notebook (available via Colab or locally) serves as the primary starting point. Links to a blog post and research paper are also provided.

Highlighted Details

• Measurement Primitives: Offers functions for rate (0-100 scores), rank (pairwise comparisons), classify (labeling), and extract (structured facts).
• Multimodal Capabilities: Supports analysis across text, images, audio, PDFs, and can leverage web search for context.
• Operational Robustness: Includes automatic parallelism, resumable runs, detailed logging, and checkpointing for large-scale data processing.
• Data Wrangling Tools: Provides utilities for merging datasets, deduplication, filtering based on natural language conditions, and PII de-identification.

Maintenance & Community

The project is hosted on GitHub, serving as the primary channel for feedback, bug reports, and feature requests. While specific community channels like Discord or Slack are not mentioned, the project is associated with OpenAI and NBER working papers, indicating a research-oriented development context.

Licensing & Compatibility

GABRIEL is released under the permissive Apache 2.0 License. This license generally allows for commercial use and integration into proprietary software without significant restrictions, provided attribution and license terms are followed.

Limitations & Caveats

The toolkit's functionality is contingent on access to and cost of the OpenAI GPT API. Effectiveness is tied to the underlying LLM's performance and the clarity of user-defined prompts. While designed for robustness, LLM outputs can exhibit variability, and users should be mindful of potential biases or inaccuracies inherent in the models.