MAESTRO: Agentic RAG for supply-chain optimization

Corteva · Industry-sponsored research · August–December 2025

Project overview

MAESTRO stands for Multi-Agent Expert System for Transparent Response Orchestration. It helps non-technical stakeholders understand supply-chain optimization models and their outputs by combining retrieved context, computed spreadsheet facts, and explanations from specialized agents.

The project covers two settings: facility location, where questions concern the placement of warehouses or other facilities, and multi-commodity flow, where users need to understand how products move through a logistics network. The goal is to turn model documentation and solver output into explanations with clear site names, relevant numbers, and supporting sources.

MAESTRO pipeline: text and spreadsheet inputs, ChromaDB retrieval, optional Pandas analysis, expert routing, and aggregation into a business answer with sources. — MAESTRO combines retrieved document context and computed spreadsheet facts before expert routing and response aggregation.

How it works

Retrieve relevant context. Text documents and spreadsheet content are indexed in ChromaDB using sentence-transformer embeddings. Semantic search supplies the relevant passages, together with their source filenames and chunk identifiers.
Compute facts from solver output. When matching spreadsheets are present, a Pandas analysis node computes product movements over time, throughput by location, and capacity utilization. Those facts are added to the context used by the agents.
Route to specialized experts. Keyword-based routing selects mathematical, optimization, supply-chain, or implementation expertise. If no routing keywords match, all four roles are consulted.
Synthesize a stakeholder-facing answer. An aggregator reconciles the expert responses into a direct answer, supporting findings, an explanation, and relevant recommendations. The command-line interface also reports the retrieved sources, chunk relevance, and expert domains used.

The implementation uses Python, LangGraph, ChromaDB, sentence-transformers, Pandas, and Gemini. LangGraph carries the retrieved context, computed facts, selected experts, and responses through the workflow.

Two supply-chain workflows

Comparison of facility-location questions answered from text context and multi-commodity-flow questions answered using text, solver output, and computed spreadsheet facts. — Facility-location questions primarily use document context; multi-commodity-flow questions additionally use spreadsheet calculations.

For facility location, the knowledge base contains the problem description, assumptions, formulation, solution approach, and implementation details. Retrieved context helps explain what the model does and why its assumptions matter.

For multi-commodity flow, the system also reads solver-output and capacity spreadsheets. This supports questions such as:

Which locations processed the most product, and which processing steps contributed?
Where is utilization highest relative to available capacity?
How does a particular product move through the network over time?

The repository includes three implementations: a multi-expert baseline without retrieval, traditional RAG with vector retrieval, and the LangGraph workflow with spreadsheet analysis. These provide progressively richer ways to ground the experts’ answers.

My contribution

I built a multi-agent chatbot to explain supply-chain optimization outputs to non-technical stakeholders. My work focused on:

Agentic RAG orchestration: Implemented the LangGraph workflow connecting retrieval, tool use, and response generation for source-grounded explanations.
Expert routing and aggregation: Routed questions to role-based expert agents and synthesized their responses into an answer that stakeholders could understand.

Collaborators: Manav Vora, Mo Phoompho, and Melkior Ornik.

The diagrams are from the MAESTRO repository, released under the MIT License.