Skip to content

Structured documentation and reusable artifacts for engineering reliable LLM agents using the CARE (Collaborative Agent Reasoning Engineering) methodology—focused on the IESO design pattern for consistent task decomposition, execution, self-evaluation, and iteration.

Notifications You must be signed in to change notification settings

NASA-IMPACT/CARE-iESO-Agent

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

2 Commits
 
 

Repository files navigation

CARE-iESO-Agent

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.


Project Focus: IESO Agent for NASA Worldview

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.


Participating SMEs

This project will be developed collaboratively with participating subject-matter experts:


What is CARE?

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.


Repository Purpose

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.


CARE Phases (IESO-Aligned)

The repository is organized into five CARE phases, mapped to IESO design needs:

Phase 1: Scope and Decompose (Interpret)

Define user goals, scientific intent classes, constraints, and operational boundaries in Worldview.

Phase 2: Key Information Elicitation (Explore)

Document datasets, layer metadata, authoritative references, known pitfalls, and required context.

Phase 3: Reasoning Policy and Guardrails (Interpret + Synthesize)

Define verification rules, uncertainty handling, citation expectations, and safety/failure behaviors.

Phase 4: Prompt Architecture and Tool Orchestration (Explore + Operationalize)

Translate the design into structured prompts, tool routing, UI action steps, and response templates.

Phase 5: Benchmarking (Synthesize + Operationalize)

Build scenario-based test suites (scientific + operational), rubrics, and regression evaluation outputs.


Repository Structure

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

Outcome

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

About

Structured documentation and reusable artifacts for engineering reliable LLM agents using the CARE (Collaborative Agent Reasoning Engineering) methodology—focused on the IESO design pattern for consistent task decomposition, execution, self-evaluation, and iteration.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published