Building a High‑Accuracy Automotive Maintenance Q&A System with Multi‑Agent LLMs

1. Introduction

Since ChatGPT’s debut in November 2022, the "large‑model+" era has opened unprecedented opportunities for AI software. Forecasts predict the AI software market could reach $13 trillion by 2030, with generative AI accounting for $1.3 trillion. Early‑stage large‑model applications, especially knowledge‑base retrieval‑augmented generation (RAG), are still in an expectation‑inflation phase.

2. Background

Automotive manufacturers generate massive fragmented knowledge in spreadsheets (xlsx) covering warranty, maintenance, pricing, and policies. These tables mix structured and unstructured data, making manual query inefficient. Integrating these tables into an intelligent Q&A system is therefore critical.

3. Scenario Analysis

Smart Q&A for automotive maintenance is a classic RAG use case. Using Alibaba Cloud’s Bailei platform to build a knowledge base, select a large model, and fine‑tune prompts yields acceptable performance, but often only 60 % logic quality because deeper system design is ignored.

4. Core Element 1: Large Model – Technology‑Driven Product

LLMs belong to the connectionist stream of AI and typically adopt a Decoder‑Only Transformer architecture, which offers simpler design, higher inference efficiency, and better scalability than Encoder‑Only or Encoder‑Decoder variants. The inference process consists of a compute‑heavy Prefill stage and a memory‑intensive Decode stage.

5. Core Element 2: Bailei – Assembable Platform

Bailei provides three construction modes (agent, workflow, multi‑agent) that enable rapid RAG development with minimal trial‑and‑error cost. Its modular, assembable nature accelerates feature delivery by up to 80 % compared with traditional pipelines.

6. Core Element 3: Agent – Key Carrier for LLM Deployment

Agents combine planning, memory, tools, and execution. Single‑agent designs struggle with complex tasks, while multi‑agent collaboration (as shown in AutoGen research) yields superior performance for heterogeneous queries.

7. Core Element 4: Prompt – Bridge Between LLM and Application

Prompt engineering (PE) adjusts model behavior without changing parameters, offering high cost‑effectiveness and flexibility. Common techniques include few‑shot CoT, role specification, and output formatting.

8. Feasibility Study – Three Solution Paths

Solution 1: Standard Bailei RAG – failed because complex xlsx files contain both structured and unstructured data, causing parsing errors.

Solution 2: Multimodal RAG with Qwen‑VL‑Max – converts spreadsheets to images for visual extraction; technically feasible but incurs high manual preprocessing cost, expensive inference, and scalability limits.

Solution 3: Document‑Parsing (LLM‑based) + Bailei – leverages Alibaba Cloud’s document‑intelligence API to extract markdown from xlsx, then builds a Bailei knowledge base. This approach offers the highest standardization and best initial results.

9. Final Implementation – Multi‑Agent Architecture

Key challenges identified were data quality, query diversity, and precision for price calculations. The solution combines:

Data cleaning and classification to improve knowledge‑base quality.

Vertical agents for distinct intents (price, content, policy) following the single‑responsibility principle.

Serverless price‑query plugins (Function Compute) to guarantee exact numeric answers.

The overall pipeline follows a three‑layer design: model core → platform services → assembled APIs, with a Router Agent directing traffic.

Data Engineering – Better Input, Better Output

We clean duplicate, erroneous, and merged cells, then classify maintenance content into structured knowledge bases. Policy documents are parsed to markdown; price tables are persisted in RDS and exposed via Serverless functions.

Vehicle Model Rewriting Agent

# Role
You are a question‑rewriting expert who uses the Vehicle‑Match plugin to normalize vehicle mentions.
# Steps
1. Rewrite the user question based on dialogue history.
2. Extract vehicle and time entities.
3. Call the Vehicle‑Match plugin if a vehicle is present.
4. Use the plugin output to rewrite the question again.
5. Classify the intent (price, content, policy, other).
# Constraints
Follow the steps strictly; output JSON.

Intent Recognition

After rewriting, the system identifies the user’s true intent to route the query to the appropriate agent (e.g., price‑calculation agent).

Document Parsing Workflow

def process_file(file_path):
    client = get_cred_client()
    print(f"file_path:{file_path}")
    id = submit_file(client, file_path)
    while True:
        status = query(client, id)
        if status == "success":
            break
        time.sleep(1)
    get_parser_result(client, id)

10. Evaluation

Initial feasibility testing of Solution 3 yielded 51.2 % accuracy, insufficient for production. After deploying the multi‑agent + price plugin architecture, batch testing achieved 93.8 % accuracy, surpassing expectations.

11. Conclusion

The article demonstrates how to:

Apply the "who, where from, where to" philosophical lens to LLM design.

Define the "two cores, four elements" framework (model, platform, agents, prompts) and the principle of maximizing standards while minimizing custom development.

Implement a multi‑agent, serverless‑enhanced RAG solution that generalizes beyond automotive maintenance to other industries.

Team: Public Cloud Business – Technical Services Department