expectedparrot/edsl

What it solves

EDSL simplifies the process of conducting computational social science and market research using AI. It allows researchers to design and run surveys and experiments with multiple AI agents and large language models (LLMs) simultaneously, as well as perform complex data labeling tasks.

How it works

EDSL provides a declarative design for survey questions (like multiple-choice or linear scales) to ensure consistent results without needing JSON schemas. It uses "scenarios" to parameterize prompts with data imported from sources like CSV, PDF, or PNG. Users can define AI agent personas with specific traits to simulate diverse responses and can run surveys across multiple LLMs to compare their responses.

The system includes piping and skip-logic to create complex data labeling flows. To ensure reproducibility, API calls are cached automatically, and remote results are stored on the Expected Parrot server with verified prompts and timestamps.

Who it’s for

Researchers in computational social science and market research, as well as data labelers who need to conduct experiments with AI agents and LLMs.

Highlights

• Declarative Question Design: Consistent results through predefined question types.
• Parameterized Prompts: Automatic data import for prompt control via scenarios.
• AI Agent Personas: Ability to construct agents with specific traits for prototyping and pre-testing.
• Multi-Model Comparison: Simplified access to run surveys across various LLMs simultaneously.
• Reproducibility: Automatic caching and a universal remote cache for sharing and replicating results at no cost.
• Integrated Collaboration: Integration with Coop, a platform for sharing AI-based research workflows and projects.