Sage
Ask Sage about wildfire recovery, next steps, or community action.
Welcome to Sage
Sage is a civically minded chatbot that helps you think through wildfire recovery, rebuilding, insurance, housing, and community action.
How to use Sage
- Start with a suggested question, or type something on your mind.
- Answer any follow-up questions so Sage understands your situation.
- Use the response to plan next steps, compare options, or draft materials.
Try: “Help me draft an outreach message for a community meeting.”
Response mode
Response mode controls what kind of information Sage uses when answering.
How to choose
- Expert-informed: uses curated research and expert resources to provide more grounded, evidence-based guidance.
- General response: generates flexible, general-purpose answers that are useful for brainstorming, drafting, or exploring ideas.
Local news context
Local news helps Sage tailor answers to what’s happening in your area.
How to use it
- Turn on “Include local news context.”
- Select a state.
- Choose a community, or leave it as “All communities.”
About this project
CS 150: Generative AI for Social Impact · Tufts University · Spring 2026
Authors: Miriam Shamash, Alisalee Aaron, Ning Ni
This project brought together our civic interests and our computer science backgrounds, and we had a great time building it. Sage was one of two projects we built as a team. The second was the , which we designed to evaluate Sage and other civic AI systems.
We are grateful to Peter Levine for helping us understand critical civic studies viewpoints and for generously dedicating his time to guide us, and to Professor Fahad Dogar for advising the technical implementation of the project and for making this all happen.
To contact us:
Miriam Shamash (author) - miriam.shamash@gmail.com
Wendan Jiang (project inquiries)- wendan.jiang@tufts.edu
CivicBench: a Civic AI benchmark
CivicBench checks whether an AI system gives civic guidance that is accurate, actionable, and context-aware. It was built to evaluate Sage and is designed so other civic AI systems can be tested the same way.
What it covers
- Seven civic categories, from civic knowledge and information access to formal participation, organizing, community problem-solving, and relationships and trust.
- 140 scenario-based questions, four for each of 35 subcategories.
- Rubrics that score each response by category and difficulty, so results show where a system is strong or weak.
How it scores
- LLM as a jury: several LLM judges score each response against the same rubric, so no single judge decides the result.
- Simulated user: Orchid, a persona agent, asks follow-up questions for up to five turns, so bots that need a conversation are tested fairly.
- Baseline: a simpler “raw bot” gives a comparison point for the components that make Sage different.