Deploying an AI‑Powered Shopping Guide on the Vivo Official Site

Introduction

With the rapid advancement of large language models, AI can understand natural language queries and infer user intent. The team explored applying an AI shopping guide to the Vivo official website to assist users in product browsing, comparison, and recommendation through multi‑turn dialogue.

Technical Selection

Before building the solution, the team compared a traditional database‑retrieval approach with a large‑model‑centric approach. Considering cost, risk, and business fit, they chose an Agent workflow that combines information engineering, prompt engineering, a nine‑question platform (knowledge base + function call + RAG), and a large language model.

Overall Architecture

The system is organized into four layers:

Application layer : user entry point, streaming LLM responses, and structured product card display.

Strategy layer : intent classification and routing to appropriate downstream services.

Nine‑question platform : creates agents, manages prompts, knowledge bases, and RAG retrieval.

Core model layer : provides compute resources, QPS allocation, and safety auditing for model inputs and outputs.

Detailed Solution

Intent‑recognition small model

The system uses a fast, lightweight model (FastText) to classify user queries into two core capabilities: phone recommendation or phone‑parameter explanation. Two options were evaluated: a large model as a front‑end (high latency) and a small model (FastText) with fast inference. The small model was selected and trained on synthesized data.

Data synthesis

输入：用户咨询“iQOO12对比X100怎么样？”
输出：_label_compare iQOO12对比X100怎么样

输入：x100的长焦效果
输出：_label_feature x100的长焦效果

FastText training code

import fasttext
# 超参数调优结果
model = fasttext.train_supervised(
    input=temp_file,
    lr=0.5,
    dim=200,
    ws=5,
    epoch=50,
    wordNgrams=3,
    minn=3,
    maxn=6,
    loss="softmax",
)

Hyper‑parameters such as wordNgrams, minn, and maxn are explained in the article. The trained model ( model.bin) runs on CPU with ~10 ms inference latency.

Model deployment

The model is containerized and deployed on CPU machines, providing near‑real‑time responses without user‑perceivable delay.

Intelligent Agent Design

Two specialized agents handle the core capabilities:

Phone‑parameter interpretation bot : extracts brand, model, and specific parameter queries, retrieves structured product information from the backend, and feeds it to the LLM.

Phone‑recommendation bot : parses user requirements (e.g., photography focus, budget ~4000 CNY, latest flagship), matches them against a curated recommendation list, and lets the LLM generate a concise recommendation with reasons.

Prompt engineering examples (truncated for brevity):

你是一位热情友好的手机导购助手，拥有丰富的手机知识，擅长帮助顾客选择最适合他们的手机。请按照以下指南与用户互动：
1. 仅回答系统提示词中补充的机型，主要为vivo和iqoo手机，其他品牌手机不予回答。
2. 控制回答的字数，不要超过200字。若回答不上来，则提示用户退出ai手机导购，转人工咨询。
...

Knowledge Base and RAG

Two vertical knowledge bases are built:

Recommendation KB : core selling points and evaluation excerpts for each phone.

Evaluation KB : detailed technical specs, performance analysis, and user‑facing descriptions.

During a query, relevant segments are retrieved via a Retrieval‑Augmented Generation (RAG) pipeline and combined with the LLM prompt.

Structured Output

The final response to the user consists of three parts:

Product card (price, image, name, user reviews) fetched from the product API.

LLM‑generated answer (streamed via SSE, with client‑side buffering to smooth display).

Related community posts (top‑5 most relevant) to provide additional context.

Evaluation

Safety testing : 16 k safety test cases covering ethics, low‑quality content, and disallowed topics were prepared. The model’s responses were audited, with policies to refuse or hand over to human agents when risks were detected.

Performance : First‑character latency is kept under 2.5 s, providing a smooth user experience.

Business results : After full rollout, A/B testing showed a positive impact on GMV and issue‑resolution rate, exceeding expectations.

Conclusion and Outlook

The AI shopping guide marks the first AI‑driven commerce interaction on the Vivo website, demonstrating technical feasibility, architectural robustness, and measurable business value. Future work includes expanding the capability boundary to accessories, after‑sale services, voice interaction, and digital‑human avatars.