How Large Language Models Boost Courier Efficiency: From Voice Commands to Smart QA

Introduction

In 2022 OpenAI released ChatGPT, a conversational model that demonstrates impressive language understanding, content generation, and knowledge reasoning. It can comprehend human intent and produce clear, complete answers that are often indistinguishable from those of a human.

Courier Operations Analysis

The courier terminal system is used daily by couriers, drivers, and site managers. It is the most widely used and diverse logistics application. By analyzing 143 courier actions—grouped into collection, delivery, on‑site, auxiliary, and customer service—research identified 69 actions that could benefit from large‑model integration, such as information entry, outbound calls, SMS sending, waybill queries, aggregation queries, knowledge Q&A, and precise prompts.

Smart Operations

Couriers frequently make phone calls or send SMS messages, but phone numbers are hidden for privacy, requiring multiple UI steps. A large model can parse spoken intent, extract waybill numbers or addresses, and trigger the appropriate API call without manual navigation. Unlike traditional pipelines that chain several small models and suffer error propagation, a single large model eliminates the need for labeling and training, reducing development time.

The workflow is:

ASR converts speech to text.

The large model performs intent recognition and information extraction.

The generated command calls the system API to complete the task.

Smart Question Answering

Couriers must follow hundreds of operational documents (e.g., handling procedures, safety standards). Remembering all rules is unrealistic, leading to errors and inefficiency. By combining prompts with Retrieval‑Augmented Generation (RAG), the system first retrieves relevant knowledge from an external knowledge base to mitigate hallucinations and outdated information, then generates concise answers.

Content Extraction & Splitting

Documents contain diverse formats, including tables that plain text extraction cannot handle. The pipeline extracts tables, converts them to semantic text, and replaces placeholders in the original document. Because large models have token limits, the text is split into chunks. Initial splitting used a fixed 300‑character window, which broke paragraph integrity. The improved method splits by paragraph with a 500‑character limit and a 10% overlap.

from src.document_loader.document_loader import DocumentLoaderUtil
processor = DocumentLoaderUtil(file_path=path_ori, pic_save_dir=dir_save_picture)
texts = processor.load()
texts = json.dumps(texts, ensure_ascii=False, indent=4)
with open(os.path.join(dir_save_text, f"{os.path.basename(path_ori)}.txt"), "w") as f:
    f.write(texts)

# Example of paragraph‑based splitting
from src.text_splitter.text_splitter import TextSplitterUtil
splitter_name = "RecursiveCharacterTextSplitter"
splitter_args = {"chunk_size": 500, "chunk_overlap": 50, "length_function": len}
splitter = TextSplitterUtil(splitter_name, splitter_args)
texts_splitted = splitter.create_documents(
    texts=[t["page_content"] for t in texts],
    metadatas=[{"source": f"{path_ori}_{i}"} for i, t in enumerate(texts)],
)

Embedding & Vector Store

After splitting, each chunk is embedded using OpenAI's text‑embedding‑ada‑002 model and stored in a Vearch vector database. The top‑k most similar vectors are retrieved and fed to the large model via a prompt to generate answers.

from src.embedding.get_embedding import get_openai_embedding
model_name = "text-embedding-ada-002-2"
texts_embedding = [
    get_openai_embedding(text=t.page_content, model_name=model_name, model_key="xxxx")
    for t in texts_splitted
]

Smart Prompts

Complex business documents and process standards are transformed into concise, courier‑friendly prompts. For example, customized packaging requirements for key accounts are summarized and delivered via text‑to‑speech (TTS) to guide couriers, improving compliance and service quality.

Intelligent Agents

Large‑model agents (e.g., GPT‑based agents) can perceive the environment through sensors and act via actuators. While current capabilities are promising, further work is needed in model orchestration, domain‑specific fine‑tuning, and safety. The team continues to explore agents for anomaly detection, root‑cause analysis, and automated resolution in the courier domain.