How a Multi‑Agent Large Model Transforms Ecological Big‑Data Analysis

Project Background and Goals

The ecological environment big‑data analysis project aims to digitize and precisely manage environmental governance. The primary research goal is to deeply explore the latest advances in multi‑agent large models, construct a flexible, efficient, and accurate multi‑agent framework, and devise domain‑specific evaluation methods for environmental large models.

Technical Solution Overview

The solution begins with a market survey of popular multi‑agent development frameworks, then proceeds to system design, front‑end and back‑end implementation, followed by comprehensive testing and scoring.

Core Agents

Planner : The main agent that generates planning instructions. Its prompt was refined to produce more stable output formats compared with the mid‑term version.

Coder : Receives Planner’s directives and generates Python code. Optimizations reduced internal loops for converting table data to JSON, improving execution speed.

Tooluser : Added several utility libraries, including city weather queries, Jiangsu VOC monthly reports, future air‑quality monitoring, and ground meteorological data tracking.

DataManager : Interfaces with enterprise databases, writing stable SQL queries to retrieve required data.

Informer : Queries domain‑specific knowledge bases, supporting both local repositories and external retrieval.

Asker : Handles ambiguous user requests by asking follow‑up questions. It introduces session management and front‑back‑end coordination, using a sectionID to track dialogue and maintain continuous logs for traceability.

Memory Mechanisms

Two memory layers were added: contextual memory based on LangGraph’s Memory Saver module, which preserves multi‑turn dialogue state, and long‑term memory built on the mem0 framework, employing trigger‑based writes and semantic retrieval stored in a MongoDB vector store.

Long‑Form Report Generation Workflow

The pipeline consists of data collection via SQL queries, data processing to extract key metrics into Excel, chart generation (bar, pie, line, rose), text generation from tabular data using a large model, and final assembly into a DOC document.

Backend and Frontend Stack

The backend uses FastAPI combined with LangGraph, providing a lightweight gateway, task‑pool management, and recoverable graph workflows. The frontend is built with React 18, designed for continuous readable dialogue streams, side‑by‑side delivery of artifacts, and a flat page structure.

Interactive Webpage Generation

After each conversation, multiple files of various formats are bundled into a shareable, visual HTML page for easy presentation.

Results and Evaluation

The team built a complete ecological environment multi‑agent system and a command‑line tool (LangGraphCLI) for cross‑platform operation. A dedicated test set evaluates the system using expert scoring and difficulty‑weighted metrics.

Documentation and Future Work

A technical documentation package was authored to guide future improvements. The authors plan to enhance long‑term memory stability, improve code‑generation reliability, and deepen collaborations with enterprises and governments to expand geographic functionalities.