Shipment Intelligence：多智能体物流智能系统与混合RAG架构实践

This post introduces Shipment Intelligence, an open-source autonomous logistics intelligence system built by SURYA-S-A. It combines LangGraph-based multi-agent architecture and hybrid RAG retrieval technology to enable autonomous query, analysis, and decision support for freight information, addressing key challenges in the logistics industry.

Global logistics handles massive freight documents daily, but traditional manual retrieval across multiple systems is inefficient and error-prone. While large language models (LLMs) offer new possibilities, direct application in logistics faces issues like domain knowledge accuracy requirements, multi-source data integration needs, and inability to handle complex queries with simple single-round QA.

Shipment Intelligence is an open-source system targeting intelligent information query, analysis, and decision support for the logistics and supply chain field. Its key 'autonomous' feature allows it to not only answer direct user questions but also proactively analyze context, identify information gaps, and plan query strategies (e.g., auto-checking shipment status, customs state, and estimated arrival time from multiple sources to generate a comprehensive report).

The system uses specialized agent division: information retrieval agents search data sources, analysis agents extract key insights, validation agents check accuracy/consistency, and report generation agents organize results. LangGraph framework enables workflow orchestration with loops (iterative query optimization), conditional branches (action based on results), parallel execution (multi-source queries), and visualizable execution paths for debugging.

Hybrid RAG combines vector retrieval (semantic matching), keyword retrieval (exact matches like waybill numbers/port codes), and graph retrieval (entity relationships like carrier-route links). It integrates data from transport management systems (TMS), warehouse management systems (WMS), and customs systems, converting diverse formats (PDF, JSON, DB records) into a unified internal representation.

For complex requests (e.g., analyzing optimal Shanghai-Los Angeles shipping plans), the system decomposes into sub-tasks (route query, rate comparison, port congestion check, risk assessment). It uses task templates and dynamic adjustment. Validation includes confidence marking for single-source info, cross-source confirmation to boost confidence, and conflict prompts with difference comparisons when info conflicts.

Business value: Reduces manual query time for freight agents, unifies info entry for importers/exporters, and assists analysts in identifying optimization opportunities/risk. Technical highlights: Modular tool design (independent external system interfaces), fine-grained state management (breakpoint resumption/error recovery), and systematic prompt engineering (agent role definitions/collaboration protocols).

Shipment Intelligence is a typical case of deep AI application in vertical industries. It is not just an LLM chat interface but a complete system covering task planning, info retrieval, multi-source integration, and result validation. The hybrid RAG + multi-agent workflow provides a reusable paradigm for complex info query applications, and such autonomous systems will become key tools for logistics digital transformation.