Best AI Auto-Reply for Buyer Messages: Feishu + DeepSeek + Qianniu Full-Chain Automation

E-commerce customer service workload catches many sellers off guard. My Taobao store averages 200 orders a day, but daily inquiries exceed 300. Size questions: 35%. Shipping queries: 25%. Returns and exchanges: 20%. Product feature questions: 15%. Others: 5%. Without AI auto-reply, you need at least 2 full-time客服 to keep up.

There are plenty of auto-reply tools on the market, but most charge fees — $7-$28/month per store. If you have 5 sub-stores, that's $140/month. And many of these tools are just keyword matching, barely "AI" at all.

In 2026, with large language model APIs dropping dramatically in price, building a genuinely intelligent auto-reply system costs nearly nothing. DeepSeek's API costs just 2 RMB per million tokens. Processing 500 messages a day means monthly API costs under 20 RMB. Combined with Feishu and Qianniu's free tiers, it's essentially zero cost.

System Architecture

The system has three layers.

Message intake layer uses Taobao Qianniu Open Platform's message push interface. When a buyer sends a message on Wangwang, Qianniu pushes it in real-time to your server.

Intelligence processing layer uses DeepSeek's API for intent recognition and reply generation. DeepSeek offers incredible value — its Chinese understanding is top-tier among similar products. It判断user intent and generates replies based on your knowledge base.

Message sending layer pushes the AI-generated reply back to the buyer through Qianniu's send API. Total latency is 2-4 seconds — users barely notice they're chatting with AI.

If you're not technical, don't worry. I'll also cover no-code platforms like Qingliu or Mingdaoyun for basic auto-reply setup without writing code.

Knowledge Base: The Core of AI Reply Quality

90% of AI auto-reply quality depends on the knowledge base. This is what the AI uses to understand problems and generate answers. If your knowledge base is incomplete, AI replies will miss the mark.

Step 1: Categorize and organize高频 questions. Open Qianniu's chat history and export the last 3 months of conversations. Sort by topic and identify the top 20 most frequent questions. Document standard responses for each.

Typical categories: Size inquiries (standard and special size对照), Shipping queries (carrier, delivery time, shipping fee policy), Return policy (7-day no-reason returns, who pays return shipping, refund timeline), Product usage (instructions, precautions, common troubleshooting).

Step 2: Write your knowledge base document. Markdown format is recommended for easy AI reading and understanding. Each entry includes: question category, possible user phrasings (3-5 variations), standard reply, and notes.

Example size knowledge entry:

## Category: Size Inquiry
### User Phrasings
- I'm 180cm, what size should I get?
- Does this run big or small?
- I usually wear L, what should I pick for this?
### Standard Reply
Hi there! Based on your height and weight, I recommend size XL. This product has a slim fit — if you prefer a looser style, go up one size.
### Notes
If the user weighs over 90kg, recommend going up one size. Sports suits fit slightly looser than regular suits.

Step 3: Upload the knowledge base to Feishu Docs. Feishu supports document API access, so we can use Feishu docs作为the knowledge base storage location. When the AI receives a user message, it reads the relevant knowledge from Feishu docs before generating a reply. No separate database needed — just use Feishu's document management.

DeepSeek API Configuration

DeepSeek is currently the best value Chinese LLM. API pricing is just 1/50th of GPT-4, and its Chinese understanding accuracy exceeds many similarly priced models.

The key is writing a good System Prompt. Here's mine:

You are a professional Taobao store AI customer service assistant. Your task is to find the best matching answer from the knowledge base based on the user's question.
Rules:
1. Only use information from the knowledge base — never fabricate information
2. Be friendly and warm — start with "亲" or "您好"
3. If no matching answer is found in the knowledge base, reply: "亲抱歉这个问题我需要转给人工客服为您处理马上为您转接"
4. Don't ask for users' personal information
5. Keep replies under 50 characters — concise is key
6. For returns/exchanges/refunds, guide users to contact human customer service
7. Reply tone should be professional and warm, not mechanical

Parameter settings: Set temperature to 0.3 — balanced between creativity and accuracy. Higher temperature makes replies too随意, lower makes them too rigid. Set max_tokens to 200 to limit reply length and avoid overly long responses.

Feishu Bot Setup

Feishu's built-in bot功能is our infrastructure for the auto-reply system. Feishu bots support webhook triggers and auto-reply — and they're completely free.

Step 1: Create a Feishu app. Open Feishu Open Platform and create an enterprise self-built app. Enable "Message & Group" permissions and "Document" read permissions.

Step 2: Create a bot. Add bot functionality to the app. Configure message reception mode to "Receive Messages" so the bot can receive user messages.

Step 3: Build your knowledge base. Create a "CS Knowledge Base" document in Feishu Knowledge Base. Organize the高频 questions and standard replies by category.

Step 4: Configure auto-reply logic. Set up event callbacks in the Feishu app backend. When a user sends a message, trigger the callback and send the message content to your backend service. The backend calls DeepSeek API to generate a reply, then sends it back through Feishu API.

Step 5: Test and launch. First test with your own messages. Confirm all question types get correct replies before going live.

No-Code Option for Non-Technical Users

If you don't know programming, use no-code tools like Qingliu or Jijianyun for a simplified version. Less flexible than a custom build, but achieves about 80% of the效果.

In Qingliu, create a "Customer Service Auto-Reply" flow. Step 1: Set up a webhook trigger — fires when Qianniu receives a new message. Step 2: Add a "Condition Judgment" node. Step 3: Call DeepSeek API from the condition node output. Step 4: Send the reply back to Qianniu via an HTTP request node.

Setup time: about 2 hours. Data flow isn't as smooth as a custom build, but it works fine for daily use. Maintenance and knowledge base updates are just a few clicks — lowest入门门槛.

Jijianyun's setup is similar. It has built-in connectors for Qianniu and Feishu — just drag and drop triggers and actions. No need to manually write API interface code, significantly lowering the barrier.

Real-World Implementation: Taobao Qianniu Integration

Qianniu Open Platform provides complete IM message APIs. The main ones are: message push callback interface and message sending interface.

First, register as a developer on Qianniu Open Platform and create an app. Select the "Message Service" API permission package. Configure the message callback URL with the address of the backend service you set up. Qianniu will then push all buyer messages to this address.

The message push data structure is simple: sender nickname, receiver nickname, message content, message type (text/image/emoji), and session ID. We only process text messages — images and emoji can't be handled by the AI. When receiving an image message, simply reply: "亲图片收到了建议文字描述您的问题哦."

Message sending requires two core parameters: the session ID (tells Qianniu which conversation to send to) and the message content (the AI-generated reply text). Rate limit: 10 calls per second — plenty for small to medium sellers.

Scaling for Million-Order Volume

When order volume exceeds 1,000/day, the basic architecture may struggle. Some optimizations needed:

First, caching layer. Cache replies to高频 questions in Redis. When a user asks the same question, the AI returns cached results without calling the API each time. Cache hit rate can reach 60%+, significantly reducing API costs.

Second, multi-model fallback. Use DeepSeek as the primary model, but configure GPT-4o-mini as backup. If DeepSeek API has errors or times out, automatically switch to the backup model to keep service running.

Third, human escalation circuit breaker. When the AI fails to answer a user's question 3 consecutive times, automatically trigger a "circuit break" — escalate to a human agent. This prevents the AI from repeatedly giving wrong answers and ruining the user experience.

Fourth, real-time monitoring alerts. Configure a real-time dashboard in Feishu. When the message queue backlogs over 100 messages or API success rate drops below 95%, automatically send alert notifications to your group.

Data Analysis and Continuous Optimization

Once AI auto-reply is live, continuous optimization matters more than the initial setup. Build a data-driven iteration loop.

Daily: Check an "AI Reply Quality Report." Metrics: total messages processed, AI successful replies, human escalations, average response time. Core metric: "AI resolution rate" — percentage where the AI replied successfully and the user didn't follow up. Above 80% means the system is running well.

Weekly: Review "Unresolved Case Analysis." Extract questions the AI couldn't answer. Analyze whether the issue is missing knowledge base entries or unclear user phrasing. Update the knowledge base accordingly. Each iteration cycle can improve the AI resolution rate by 2-3 percentage points every two weeks.

Monthly: Conduct a "User Satisfaction Sampling Survey." Randomly sample 100 conversations from AI-replied users. Score each on accuracy and tone on a 5-point scale. Investigate anything below 3 points.

Platform Differences

The above方案uses Qianniu as an example, but other platforms are similar:

Pinduoduo Merchant Edition has a similar message API to Qianniu. Create an app on Pinduoduo Open Platform and configure the message push URL. Same process, just different API names and parameters.

Douyin E-Commerce has its own IM system separate from Qianniu. Douyin Open Platform provides customer service message API and push callbacks. Note: Douyin requires AI reply messages to be sent within 30 seconds of the user's last message, or they'll fail. So Douyin requires faster system response.

Shopify doesn't need Qianniu integration. Customer messages come mainly through email and Chat plugins. Use platforms like Zendesk or Tidio, then connect their APIs to AI models for auto-reply.

Alibaba International Station requires English customer service. For this, use GPT-4o-mini instead of DeepSeek — GPT handles English understanding and reply quality better.

Cost Breakdown

DeepSeek API: Billed per token. Average user message: ~200 input tokens + ~100 output tokens. 500 messages/day = ~150,000 tokens/day. At DeepSeek's 2 RMB per million tokens, daily cost = ~0.3 RMB. Monthly = ~9 RMB.

Server costs: Minimum cloud server ≈ 50 RMB/month. With serverless architecture (Vercel, Cloudflare Workers), many plans have free tiers — zero cost to start.

Feishu and Qianniu API fees: Both free.

Total monthly cost: 9 RMB minimum, no more than 100 RMB. Compared to 8,000-15,000 RMB for human customer service, that's a 99%+ savings.

Launch Checklist

Before going live, check every item on this list to ensure stable operation:

Knowledge base: Confirm Q&A covers 80%+ of高频 questions. Every answer is human-reviewed for accuracy. Size charts and shipping templates are up to date.

Technical: Confirm API keys and permissions are correctly configured. Message push callback is working and test messages are received. AI replies can be sent to Qianniu. Temperature and token limits are properly set.

Emergency plan: Maintain the ability for human客服to manually take over. Configure an emergency kill switch for one-click shutdown. Add the person in charge's phone number to alert notifications.

Finally, for the first 3 days after launch, arrange for a human customer service agent to monitor AI reply quality in real-time. Intervene immediately if errors are found, and补充the knowledge base accordingly. After 3 days of verification, the system should run stably.

Summary

AI auto-reply isn't about replacing humans — it's about focusing limited customer service resources on what matters most. Daily size, shipping, and return inquiries — the AI handles these perfectly. Human客服get freed from repetitive work to focus on real complex issues and emotional customer complaints.

This Feishu + DeepSeek + Qianniu solution is low-cost, efficient, and easy to maintain. Spend half a day setting up your knowledge base, and the AI will handle 80% of your customer service workload. For budget-conscious small and medium sellers, this is currently the best value customer service solution available.

Is your store ready? Give yourself one day to build this system. The next day, you'll find your workload has dropped dramatically — while reply quality has actually improved.