How LLM‑Powered Multi‑Agent AI Boosts Vehicle Selection in HuoLala’s Customer Service

Background

AI customer service (AI客服) aims to improve service efficiency and quality while reducing operational costs, and the vehicle‑selection assistant is a key sub‑scenario.

Problem Overview

Existing rule‑based solutions require extensive configuration of complex flowcharts and rules, resulting in poor user experience, high hand‑off rates, and increased support costs.

Technical Background

Rapid advances in large language models (LLMs) have opened new possibilities for general AI capabilities. HuoLala has internally developed a freight‑specific LLM to power its services.

System Framework

The architecture is divided into four layers:

Interaction Layer – web, app, mini‑program channels.

Application Layer – IM backend and websocket gateway.

Logic Layer – dialogue management controlled by a DM system.

Algorithm Layer – core services built on a Multi‑Agent framework and an internal machine‑learning platform for model management and deployment.

Algorithm Solution

To leverage agent capabilities while keeping behavior controllable, the project uses SOP‑based task decomposition, enabling fine‑grained control of LLM agents and easier deployment.

NLG Agent

The NLG Agent receives the user query, determines its type, and executes the appropriate action. It interacts with Memory, Tools, and the NLU Agent to answer FAQs, collect cargo information, or recommend vehicles.

Memory

Memory stores prior knowledge, including a general knowledge base, FAQs, and short‑term/long‑term dialogue memory.

Tools

Tools provide external integrations such as vehicle‑recommendation APIs, sensitive‑word filtering, business SOP execution, and flow‑guidance modules.

NLU Agent

The NLU Agent extracts cargo dimensions, weight, and transport method from user utterances. Traditional NER requires large labeled datasets and struggles with nested entities. The in‑house freight LLM achieved only 50% accuracy initially, 60% after prompt tuning, and, after incorporating retrieval‑augmented generation (RAG) and rule‑based post‑processing, reached 90% accuracy on internal test sets.

RAG Technique

Multi‑path retrieval combines vector similarity and product‑name recall to generate candidate contexts; the top‑10 results are fed to the freight LLM, raising extraction accuracy to 80%, and with contextual cues and rules to 90%.

Effect Demonstration

Examples illustrate the assistant handling single‑item and multi‑item queries, correcting user‑provided cargo information, and maintaining dialogue flow without unnecessary hand‑offs.

Effect Analysis

Flow guidance returns to the main process after answering FAQs.

Human‑like concise replies improve user perception.

Contextual understanding corrects vehicle recommendations when users amend cargo details.

Conclusion and Outlook

The LLM‑driven multi‑agent AI客服 vehicle‑selection assistant significantly improves efficiency and user experience. Future work will continue to optimize the model and extend large‑model capabilities to additional business scenarios.