CARE-iESO-Agent is a structured repository for documenting and applying the CARE (Collaborative Agent Reasoning Engineering) methodology to design a reliable IESO-style LLM chatbot agent for NASA Worldview: https://worldview.earthdata.nasa.gov/
This repository captures structured design documentation and reusable artifacts for building an agent that helps users interpret, explore, and validate Earth observation imagery and related dataset context inside NASA Worldview—while staying grounded in authoritative sources and SME-reviewed behaviors.
This work is specifically tailored to an iESO/Worldview design pattern, where the agent operates through a consistent engineering loop. The goal is to produce a reviewable, testable agent design that supports both new and expert users in performing correct scientific and operational interpretation of what they see in NASA Worldview.
This project will be developed collaboratively with participating subject-matter experts:
- Rousseaux, Cecile S. (GSFC-6160) — [email protected]
- Brosnan, Ian G. (ARC-SG) — [email protected]
- Hain, Christopher R. (MSFC-ST11) — [email protected]
- Cetinic, Ivona (GSFC-616.0) [MORGAN STATE UNIV.] — [email protected]
- Felikson, Denis (GSFC-6150) — [email protected]
- Kumar, Sujay V (GSFC-6170) — [email protected]
CARE (Collaborative Agent Reasoning Engineering) is a staged, artifact-driven methodology for engineering LLM agents that are:
- Explicitly specified rather than prompt-tuned by trial and error
- Grounded in authoritative Earth science and NASA/Earthdata context
- Designed with transparent reasoning policies and guardrails
- Evaluated with realistic queries and structured benchmarks
CARE emphasizes repeatability, reviewability, and maintainability of agent behavior over time, enabling safe iteration as datasets, UI behavior, and science expectations evolve.
This repository exists to:
- Organize CARE artifacts specifically for the NASA Worldview chatbot agent
- Capture IESO phase prompts, specifications, and outputs in a reviewable format
- Support collaboration between SMEs, developers, and LLM-based helper agents
- Track design evolution, decisions, and benchmarking results over time
The focus is primarily on agent design documentation and evaluation artifacts, not runtime application code.
The repository is organized into five CARE phases, mapped to IESO design needs:
Define user goals, scientific intent classes, constraints, and operational boundaries in Worldview.
Document datasets, layer metadata, authoritative references, known pitfalls, and required context.
Define verification rules, uncertainty handling, citation expectations, and safety/failure behaviors.
Translate the design into structured prompts, tool routing, UI action steps, and response templates.
Build scenario-based test suites (scientific + operational), rubrics, and regression evaluation outputs.
AKD-CARE/
├── phase_1_scope_and_decompose/
├── phase_2_key_information_elicitation/
├── phase_3_reasoning_policy_and_guardrails/
├── phase_4_prompt_architecture_and_tool_orchestration/
├── phase_5_benchmarking/
├── examples/
├── docs/
└── changelog.md
The deliverable of this repo is a reusable, reviewable IESO agent design package for NASA Worldview—supporting:
- user intent understanding (scientific + operational)
- grounded assistance for selecting imagery/layers/time ranges
- careful interpretation aligned with Earth science best practices
- actionable guidance for using Worldview correctly
- evaluation that goes beyond demos into repeatable benchmarks