How a Large‑Model Powered Bot Boosts Logistics Ops with Smart Q&A and Data Insights

Background & Problems

Functional staff (operations managers) need to access scattered SOPs, notices, and manuals, often via email or chat, leading to high time cost and poor experience. Frontline workers repeatedly ask common system questions, causing duplicated effort and delayed responses. Managers lack a unified query portal for reports, facing long lookup chains and missing proactive alerts. Collaboration suffers from missing channels for temporary information, and most tasks are PC‑only, lacking mobile access.

Measures & Goals

Based on a large model, the "Yunli Xiao Zhi" smart robot was built as a portable knowledge‑base platform that combines knowledge Q&A and data analysis. It covers SOPs, system FAQs, manuals, real‑time info (weather, safety), report queries, and analytics, aiming to provide efficient, personalized answers and reduce knowledge‑acquisition cost.

Key Functions

Intelligent Q&A: multi‑turn dialogue that answers common operational questions and performs data queries.

Proactive Alerts: sends scheduled or anomaly‑driven messages to individuals or groups on PC and mobile.

Implementation Details

1. Knowledge Q&A

The system uses the open‑source LangChain framework and the company’s LLM API to implement a Retrieval‑Augmented Generation (RAG) chatbot. Two knowledge bases are built:

QA knowledge base: generated from existing QA pairs, model‑extracted document snippets, and annotated logs.

Document knowledge base: stores PDFs, DOCX, PPTX as raw text, with structured table extraction via a custom PDF parser.

Both bases are vectorized and stored in JD’s Vearch vector database. Example of PDF parsing result is shown below.

{
    "metadata": {
        "footers": [],
        "headers": [],
        "catalogs": []
    },
    "chapters": {
        "1": "[CHAPTER_ROOT]",
        "1.1": "第一条 xxx",
        "1.2": "第二条 xxxx",
        "1.3": "第三条 xxxx"
    },
    "context": [
        {
            "text": "JDLxxxx规定",
            "type": "text",
            "pid": 1,
            "sid": 1,
            "metadata": {"section_range": []},
            "cid": "1"
        },
        ...
    ]
}

Table‑type text blocks are also stored with cell coordinates, enabling conversion to markdown or JSON.

{
    "text": [
        [[0,0,1,1], "名称"],
        [[0,1,1,2], "尺寸"],
        [[0,2,1,3], "三层加强材质"],
        ...
    ],
    "type": "table",
    "pid": 89,
    "sid": 111,
    "metadata": {"section_range": []},
    "cid": "1.8"
}

During answer generation, the system first retrieves relevant knowledge, then constructs a prompt for the LLM. An additional question‑condensing step improves similarity matching:

from langchain import PromptTemplate
from langchain.chains import LLMChain
from langchain.chat_models import ChatOpenAI

def get_condense_question_chain(self):
    """Condense multi‑turn question"""
    CONDENSE_QUESTION_PROMPT = PromptTemplate.from_template(
        """Given chat_history and a follow‑up question, rewrite the follow‑up into a standalone question using the original language, avoiding ambiguous pronouns.
        Chat History:
        {chat_history}
        Follow‑up Input: {question}
        Standalone Question:"""
    )
    condense_question_chain = LLMChain(
        llm=ChatOpenAI(
            model="",
            temperature="",
            openai_api_key="",
            openai_api_base="",
        ),
        prompt=CONDENSE_QUESTION_PROMPT,
    )
    return condense_question_chain

2. Data Analysis

To handle diverse reporting needs, a NoETL derived logical model asset framework was created. It automatically generates logical models for any time grain and business dimension without extra manpower, based on metadata of indicators.

A semantic knowledge graph is built from model metadata, allowing natural‑language queries to be mapped to technical fields. The AI‑enhanced SQL generation combines accurate element selection with prompts, producing precise queries that outperform most manual analysts.

Key agent capabilities include:

Natural‑language to OLAP translation.

Metric drift attribution.

AI‑augmented analysis and interpretation.

Capability Demonstrations

Feature 1: Indicator Query

Users can ask for metrics such as on‑time arrival rates, load factors, or efficiency indicators using a concise format (time + dimension + metric + chart type). The bot returns data instantly via a unified entry point, shortening the analysis chain.

Feature 2: Knowledge Q&A

Provides quick answers to SOPs, system FAQs, daily reports, and TMS guides, dramatically reducing manual lookup time.

Feature 3: Trajectory Query

By entering a dispatch order number (TW), users obtain vehicle trajectory information in one step, cutting query time from 2‑3 minutes to under 1 minute.

Feature 4: Driving‑License Image Query

Enter a license‑plate number to retrieve front and back images of the vehicle’s driving license, reducing verification time from ~10 minutes to under 1 minute.

Feature 5: Report Push

Supports scheduled and alert‑based report pushes to groups in JD ME, enabling automated delivery of key metrics.

Feature 6: Information Push

Weekly announcements, surveys, and other notices can be proactively pushed to users.

Effect Evaluation

Usage grew to 50‑100 active weekly users with over 500 queries, covering 154 organizations and 347 users. Trajectory queries dropped from 2‑3 minutes to under 1 minute, with 1,230 total queries. Driving‑license lookups reduced from ~10 minutes to 1 minute. Success rates: data‑analysis queries ~70 %, knowledge Q&A improved from 20‑30 % to ~50 % after user guidance.

Summary & Future Plan

"Yunli Xiao Zhi" integrates knowledge Q&A, data analysis, and report pushing to improve efficiency for logistics staff. Q2 focuses on delivering these core functions, refining retrieval stability, and scaling adoption. Future roadmap includes expanding AI capabilities, enhancing semantic graph coverage, and broader enterprise rollout.