What DeepSeek-V4 Changes​

According to the official DeepSeek announcement, the V4 preview brings several practical upgrades:

• 1M context support for long documents, larger knowledge bases, and longer multi-turn sessions
• stronger Agent performance for coding, document generation, and tool-driven workflows
• clearer model positioning: deepseek-v4-pro for higher-end reasoning and agent tasks, deepseek-v4-flash for lower latency and better cost efficiency

For customer service, knowledge retrieval, and AI workflow automation, those improvements matter because the model can keep more context in memory and handle longer task chains more consistently.

DeepSeek Models Already Available in Bytedesk​

Under the DeepSeek provider, Bytedesk now exposes the following model options:

That means existing DeepSeek users can move to the new generation by updating the selected model in the AI model configuration page.

Deprecation of Legacy Model Names​

DeepSeek has announced that these legacy model names will stop working on 2026-07-24:

• deepseek-chat
• deepseek-reasoner

During the transition period, they remain available for backward compatibility:

• deepseek-chat currently points to the non-thinking mode of deepseek-v4-flash
• deepseek-reasoner currently points to the thinking mode of deepseek-v4-flash

For any new robot, workspace, or tenant-level AI setup, it is better to use deepseek-v4-flash or deepseek-v4-pro directly instead of continuing with the legacy names.

Recommended Migration Path in Bytedesk​

If you already run DeepSeek in production, use this migration path:

1. Replace Legacy Model Names​

Update your default model selection as follows:

• deepseek-chat -> deepseek-v4-flash
• deepseek-reasoner -> deepseek-v4-pro or deepseek-v4-flash

Choose deepseek-v4-flash when you want faster responses and lower cost. Choose deepseek-v4-pro when you want stronger reasoning, more complex agent behavior, or better long-chain task quality.

2. Keep the Same Base URL​

The DeepSeek API endpoint does not change:

https://api.deepseek.com

The migration is mainly about updating the model value, not the endpoint.

3. Switch Models in the Admin Console​

In Bytedesk, the migration can be completed from the admin UI:

1. Sign in to the admin console
2. Open AI model settings
3. Choose DeepSeek as the provider
4. Change the default model to deepseek-v4-flash or deepseek-v4-pro
5. Save and verify the result with a chat test

This does not require reinserting chat code or changing the website integration path.

Practical Selection Advice​

• Use deepseek-v4-flash as the default choice for standard customer-service bots
• Use deepseek-v4-pro for harder knowledge tasks, AI assistants, and workflow copilots
• Keep deepseek-chat and deepseek-reasoner only as a short-term compatibility bridge, then migrate before the shutdown date

Summary​

Bytedesk already supports the latest DeepSeek-V4 models, including deepseek-v4-flash and deepseek-v4-pro. If your team depends on long context, stronger reasoning, and better agent execution, this is the right time to move off the old model names and standardize on the V4 identifiers.

Related Resources​

• DeepSeek-V4 Preview Announcement
• DeepSeek API Docs

Its answer is straightforward. Stop treating skill behavior as a loose prompt and start treating the agent skill as a versioned asset that can be created, evaluated, diagnosed, and refined over time.

That matters a lot for Bytedesk. Bytedesk already has the building blocks that many teams still lack: multi-model access, knowledge retrieval, bot routing, workflow settings, and human handoff. The next competitive gap will not come from “connecting more models.” It will come from building a customer-service system that can absorb failures, reuse domain experience, and evolve its skills with evidence.

The Core Problem the Paper Solves​

SkillForge is framed around enterprise cloud support, but the underlying problem maps closely to customer service systems.

The paper highlights two long-term issues:

• Initial skills are often not grounded enough in real business workflows. Generic skill creators do not understand private documentation, historical tickets, internal tools, or escalation logic.
• Once the skill is deployed, it usually does not improve in a systematic way. Teams collect bad cases every day, but many systems never trace those failures back to defects in the skill definition itself.

This is also why many AI customer-service demos look impressive early and then flatten out in production. The model may be strong, but answer quality is usually constrained by domain knowledge, clarification strategy, tool usage, response style, and whether those elements are refined from operational feedback.

The SkillForge Method in One Loop​

The paper treats the agent skill as a software-like artifact. Its core loop can be summarized in five steps.

1. Build the Initial Skill from Domain Context​

Instead of generating a generic SKILL.md from a universal template, SkillForge first mines context from:

• historical tickets
• technical documentation and knowledge bases
• expert-used tools and recurring workflows

That context is then used to generate a better initial skill. The paper calls this the Domain-Contextualized Skill Creator.

2. Execute Online and Collect Bad Cases​

The agent runs production tasks using the current skill version. When its output diverges from the expert reference, or when humans do not adopt the response, the interaction is flagged as a bad case.

This is a key point. Self-evolution does not begin with “more prompt tuning.” It begins with a reliable definition of failure and a stable way to collect it.

3. Diagnose Failures Across Multiple Dimensions​

SkillForge does not reduce every failure to “the model answered badly.” It analyzes failures across four dimensions:

• Knowledge: missing, wrong, or conflicting knowledge
• Tool: missing tool invocation, wrong parameters, wrong interpretation of results
• Clarification: missing clarification, unnecessary clarification, or wrong clarification direction
• Style: robotic, cold, verbose, or otherwise misaligned response style

This is important because it transforms a vague “bad answer” into a structured defect.

4. Map the Failure Back to the Skill Definition​

The paper’s Skill Diagnostician reads the aggregated bad-case report and the current SKILL.md, then maps the problem back to the skill itself.

For example:

• a recurring FAQ failure may imply an incomplete troubleshooting section
• repeated tool misuse may indicate poor tool-call guidance
• consistently robotic answers may indicate weak style constraints

This turns online quality issues into concrete answers to a much more actionable question: which part of the skill should change?

5. Make Minimal Changes and Publish the Next Skill Version​

The Skill Optimizer updates SKILL.md and references, then produces the next skill version.

The paper emphasizes two engineering principles:

• only make the minimum necessary changes
• keep the skill asset versioned, traceable, and revertible

That is much closer to software engineering than to ad hoc prompt editing.

Why This Matters for Bytedesk​

Bytedesk is not a single chatbot. It spans visitor chat, agent workbench, knowledge base, tickets, workflows, voice/video, and enterprise integration scenarios. The more complex the system becomes, the less sustainable it is to rely on a single “LLM answer interface” as the center of AI quality.

Looking at the current codebase, Bytedesk already has several strong foundations.

1. Multi-Provider, Multi-Model Infrastructure Already Exists​

The provider configuration already includes OpenAI, Anthropic, Gemini, DeepSeek, Qwen-compatible providers, OpenRouter, Dify, n8n, Ragflow, and more. That means Bytedesk already has the abstraction needed to run skills on different model backends without redesigning the runtime from scratch.

2. Knowledge Retrieval and LLM Context Assembly Already Exist​

The current bot answering flow already aggregates knowledge-base search results and injects FAQ-derived context into the LLM pipeline. In other words, the core “domain context” layer from the paper is not missing in Bytedesk. What is missing is the next step: turning that context into a first-class, versioned skill asset instead of a one-off retrieval step.

3. Bot Routing and Human Fallback Already Exist​

Workgroup routing already supports decisions such as whether to transfer to bot mode, whether offline traffic should prefer backup human handling, and whether the visitor explicitly forces human service. This makes Bytedesk a natural fit for the paper’s human-in-the-loop model rather than an unsafe full-automation design.

4. Workflow and Service Settings Already Exist​

The service settings layer already exposes workflow-related and FAQ-related configuration, including workflow enabling, FAQ knowledge-base binding, and interaction options. That means Bytedesk does not need to invent an orchestration entry point. It needs to elevate “skill” into a first-class object alongside workflow and knowledge-base configuration.

5. A Feedback Entry Point Exists, but It Is Still Thin​

There is already a message-feedback entity in the service layer, but it is not yet rich enough to support structured failure analysis, automated diagnosis, or skill refinement. That makes it a good starting point, but not a finished self-evolution foundation.

The Most Valuable Upgrade Directions for Bytedesk​

If we translate the paper into an actual Bytedesk roadmap, five directions stand out.

1. Turn Skills into Explicit Assets Instead of Hidden Prompt Fragments​

In many customer-service systems, “prompting” is scattered across robot settings, workgroup settings, knowledge bases, default replies, and workflow nodes. That makes capability hard to reuse and harder to improve.

A more durable model is to define a skill as a first-class object with at least these layers:

• instruction layer: scope, goals, boundaries, clarification strategy, response style
• knowledge layer: FAQs, document chunks, terminology, sample tickets, fault trees
• tool layer: allowed tools, when to call them, input/output constraints
• process layer: handling order, escalation conditions, transfer-to-human conditions
• evaluation layer: success definitions, failure categories, feedback mapping rules

Once that exists, robots, workgroups, assistants, and workflows can reuse the same skill assets instead of each owning isolated fragments.

2. Upgrade Feedback Logs into Structured Failure Records​

Without structured failure records, there is no self-evolution loop. Bytedesk should start collecting and normalizing signals such as:

• whether the user keeps asking the same unresolved question
• whether the user triggers transfer to a human
• whether the agent rewrites the AI suggestion
• whether the AI answer is adopted, partially adopted, or discarded
• whether the user downvotes, complains, or gives low satisfaction
• whether tool calls fail, time out, or miss expected results

Those signals should then be classified using the same four primary dimensions from the paper: Knowledge, Tool, Clarification, and Style.

3. Evolve the Knowledge Base into a Skill Knowledge Base​

Bytedesk already has FAQs, vector retrieval, and source citation. That is a solid base. But the paper makes a more subtle point: enterprise agent knowledge is not only about documents. It is also about how experts solve problems.

So the knowledge layer should eventually separate two categories:

• static knowledge: FAQs, product docs, API docs, rules, policies
• dynamic experience: high-quality historical ticket paths, clarification patterns, escalation judgment, frequent tool combinations

If Bytedesk only does static RAG, it will answer what it knows. If it also absorbs high-quality ticket trajectories, it starts behaving more like an experienced support engineer.

4. Add a Skill Diagnostician and a Skill Optimizer​

This is the most inspiring part of the paper and also the biggest current gap in Bytedesk.

Bytedesk could introduce two background capabilities on top of the existing AI module:

• Skill Diagnostician: periodically reads failed cases and outputs a report describing which skill sections are defective
• Skill Optimizer: generates revised skill drafts or updated references based on that report

This does not need to be fully autonomous on day one. A pragmatic first-stage path is:

1. auto-generate the diagnostic report
2. auto-generate the proposed skill revision
3. let operators, QA, or admins review it
4. publish the new version after approval

That would already be a major step forward from manually reading chat logs and editing prompts by hand.

5. Make Skill Operations Visible in the Admin Console​

If skills only live in files, most enterprises will never operate them well. Bytedesk is better positioned to expose them as admin-facing operational assets:

• view skill version history
• compare differences across versions
• inspect how each version affects hit rate, transfer-to-human rate, and satisfaction
• run A/B tests by tenant, workgroup, or robot
• import industry skill templates

Once that happens, the platform is no longer just “a customer-service system with AI.” It becomes an enterprise agent-skill operating system.

A Practical Rollout Sequence for Bytedesk​

If we prioritize by implementation leverage, a three-stage roadmap is more realistic than trying to automate everything at once.

Phase 1: Build the Data Loop First​

• enrich the message feedback model and failure record schema
• unify AI suggestion, human adoption, transfer-to-human, user follow-up, and user rating events
• build a base failure-analysis dashboard

Phase 2: Promote Skill to a First-Class Config Object​

• add skill binding in robot settings, workgroup settings, and agent settings
• support skill templates, versioning, publishing, and rollback
• store high-quality FAQ content, ticket summaries, and workflow guidance as skill references

Phase 3: Introduce Semi-Automated Evolution​

• run scheduled skill diagnosis jobs
• generate optimization suggestions and revised skill drafts
• release via review and gray rollout
• compare key metrics before and after publication

The main advantage of this roadmap is that it does not require blind trust in “AI automatically rewriting skills.” It lets Bytedesk build the evidence chain, diagnosis chain, and approval chain first, then gradually raise the degree of automation.

The Real Lesson SkillForge Offers​

The most valuable takeaway from this paper is not the framework name. It is the shift in focus: in enterprise agents, the core asset is moving away from the model alone and toward the skill system around the model.

The teams that can turn domain knowledge, tool usage norms, workflows, failure feedback, and expert experience into a living skill system will be the ones that move beyond agents that merely “talk” and into agents that reliably solve problems.

Bytedesk already has several of the foundational modules needed for that transition. If the platform can connect knowledge base, routing, workflow, feedback, QA, and admin configuration into a SkillForge-like loop, it will no longer just be an AI-enabled support product. It will become a self-improving enterprise agent platform.

References​

• SkillForge: Forging Domain-Specific, Self-Evolving Agent Skills in Cloud Technical Support
• Bytedesk AI and Big Data Customer Service Overview

Core Applications: AI & Big Data Redefining the Service Foundation​

The transformation AI and Big Data bring to customer service is essentially a shift from "labor-intensive" to "technology-intensive" operations, driven by data-driven decision-making + intelligent automation, covering every critical node across the service lifecycle.

1. Big Data: The "Insight Engine" for Customer Service​

The core value of Big Data lies in breaking down data silos and enabling precise anticipation of user needs throughout their lifecycle. Weiyu integrates data across multiple channels — phone, web, app, and social — to build a three-dimensional customer profile:

• Basic Attributes: Age, region, purchasing power
• Behavioral Preferences: Browsing history, consultation records, purchase frequency
• Demand Characteristics: High-frequency questions, latent needs, emotional tendencies

This panoramic insight transforms customer service from "passively waiting for inquiries" to "proactively anticipating needs." When a user repeatedly browses a product's after-sales policy, the system can proactively push operation guides; when a high-value customer's service frequency drops, the system automatically triggers a dedicated follow-up — enabling churn prediction and prevention.

Real-time data analytics also optimizes resource allocation. By statistically modeling historical call volumes and consultation trends, Weiyu can precisely forecast peak-period demand (e.g., during e-commerce promotions or policy changes) and dynamically adjust the ratio of AI bots to human agents, eliminating wait times.

2. AI: The "Intelligent Execution Hub" for Customer Service​

If Big Data is the "eyes," AI is the "brain and hands" of customer service. Weiyu's AI capabilities penetrate the entire service lifecycle:

Intelligent Response Layer​

AI customer service powered by Large Language Models (LLM) delivers 7×24-hour instant responses, supporting multimodal interactions (voice, text, image), solving the pain point of "asking the same simple question repeatedly." Context-aware multi-turn dialogue eliminates mechanical repetitive questioning.

Human-AI Collaboration Layer​

When issues exceed AI's scope, the system automatically generates conversation summaries and pre-fills ticket fields, enabling human agents to "take over and act immediately." The intelligent assistant module pushes real-time script recommendations and compliance reminders during live sessions, helping new agents ramp up quickly and significantly reducing training periods.

Intelligent Operations Layer​

Weiyu leverages knowledge graph technology for authoritative knowledge updates and dynamic learning. When a company updates its after-sales policy, administrators simply upload the new document — AI automatically parses the changes and synchronizes response strategies. Unresolved issues are continuously distilled into new knowledge points, forming a self-reinforcing "serve–optimize" cycle.

Core Value: Three Breakthroughs​

Weiyu's integration of AI and Big Data delivers three core value breakthroughs:

Efficiency Revolution​

AI bots handle 70%+ of high-frequency, simple inquiries, freeing human agents to focus on complex issues and emotional support. In practice, first-contact resolution rates increase by up to 50%, and customer service operating costs drop significantly.

Experience Upgrade​

Omnichannel seamless connection, personalized responses, and instant replies free users from the frustrations of "waiting in queues" and "repeating themselves," driving substantial improvements in customer satisfaction scores.

Value Reconstruction​

Customer service is no longer a pure after-sales function — it becomes a critical touchpoint for demand discovery, business conversion, and product optimization. User conversation data feeds back into front-end product design and marketing decisions, truly transforming the "cost center" into a "growth engine."

Industry Application Scenarios​

Weiyu's AI and Big Data capabilities apply across a wide range of industries:

Future Trends​

As technology continues to evolve, Weiyu Customer Service will advance along three major trends:

1. Multimodal Interaction Goes Mainstream: AI enables "visual diagnosis" and "remote guidance" through images and video, further reducing user communication costs.
2. Privacy Compliance Meets Data Security: Permission tiering and data encryption balance service experience with user trust, meeting enterprise compliance requirements.
3. From "Demand Response" to "Demand Foresight": Through deeper data analysis, solutions are proactively delivered before users even raise issues — "service arrives before the call."

AI and Big Data are reshaping the core logic of customer service. Weiyu not only makes service more efficient and empathetic, but also transforms customer service from an enterprise "cost burden" into a "hidden engine" driving growth — injecting lasting momentum into digital transformation across industries.

Experience Weiyu Customer Service Platform and start your intelligent customer service upgrade today.

概述​

多模态集成是指系统能够处理文本、图像、视频、音频等多种形式的信息输入，并将其转化为统一的知识表示，从而实现跨模态的信息理解与响应。微语系统集成了先进的多模态模型，使客服机器人能够：

• 读取并理解用户上传的图片内容
• 提取视频中的关键信息和场景
• 转录并理解音频内容
• 结合企业知识库，对多模态内容进行专业解答

视觉理解能力​

微语系统的视觉理解模块可以处理多种类型的图像内容，为用户提供智能分析和解答。

图像处理场景​

应用场景示例​

• 智能客服场景：用户上传产品图片，系统自动识别产品型号并提供相关信息
• 文档处理：将图像类文档解析为结构化文本，精准识别文字并提取表格信息
• 图像问答：识别图像中的人物、物体、场景等，并进行分类标记
• 数学题解答：识别并解答用户拍摄的数学题目，适用于各教育阶段
• 物体定位：在图像中准确定位特定物体，返回坐标信息
• 表单信息提取：从票据、证件、表单中提取关键信息并格式化输出

微语系统支持多语言文字识别，包括：中文、英语、日语、韩语、阿拉伯语、越南语、法语、德语、意大利语、西班牙语、俄语和葡萄牙语。

视频理解能力​

微语系统能够分析视频内容，提取关键信息，为用户提供更全面的服务支持。

视频处理功能​

• 场景识别：自动识别视频中的关键场景和内容
• 事件定位：定位视频中的特定事件并生成时间戳
• 内容摘要：生成视频关键时间段的文字摘要
• 视频问答：针对视频内容回答用户提问

视频应用场景示例​

• 教学视频分析：从教学视频中提取关键知识点
• 产品演示理解：分析产品演示视频，提取操作步骤和要点
• 视频故障诊断：识别设备故障视频中的异常状况

音频理解能力​

微语系统集成了先进的音频语言模型，能够处理多种音频输入并提供智能理解和分析。

音频处理功能​

• 语音转文字：将用户语音准确转录为文本
• 音频语义理解：理解语音内容的深层含义
• 情感分析：分析语音中的情感色彩和语气
• 音频事件检测：识别特定音频事件和场景
• 多语言支持：支持多种语言的语音识别和理解

音频应用场景示例​

• 客服语音交互：理解用户语音问题并给出专业回答
• 语音指令处理：执行用户通过语音发出的各类指令
• 会议记录整理：自动转录会议内容并提取关键信息
• 情感分析：分析客户语音反馈中的情感倾向

与知识库结合​

微语系统的多模态能力与企业知识库深度结合，实现了更加智能的用户服务体验：

1. 多模态输入理解：系统首先理解用户上传的图片、视频或音频内容
2. 知识库联动查询：将理解的内容与企业知识库进行关联查询
3. 专业解答生成：结合多模态理解与知识库信息，生成专业、准确的回答

这种结合使客服系统能够：

• 对用户上传的产品照片进行型号识别并提供相应的使用指南
• 分析用户提交的故障视频并给出针对性的解决方案
• 理解用户的语音描述并匹配知识库中的相关信息

总结​

微语系统的多模态集成能力大大拓展了智能客服的服务边界，使系统能够处理更加丰富的用户输入形式，提供更加全面、精准的服务。通过结合企业知识库，微语系统不仅能够"看懂"和"听懂"用户问题，还能给出专业的解答，真正实现智能化的客户服务体验。

Pellentesque elementum dignissim ultricies.

Pellentesque elementum dignissim ultricies.

Qwen3大模型介绍​

通义千问Qwen3是阿里云推出的大型语言模型，具有强大的理解能力和生成能力。Qwen3系列模型在多轮对话、文本生成、问答解析等方面表现出色，非常适合客服场景应用。

Qwen3是千问系列大语言模型的最新一代，提供了全面的稠密模型和混合专家模型（Mixture-of-Experts，MoE）。您可以访问Ollama官方库获取更多详细信息。

目前Qwen3系列提供多种不同参数规模的模型版本，以适应不同的应用场景和硬件环境：

• Qwen3-tools: 针对工具使用进行特别优化的版本
• Qwen3-0.6b: 超轻量版本，适合资源受限场景
• Qwen3-1.8b: 轻量级模型，平衡性能和资源消耗
• Qwen3-4b: 中小型模型，提供更好的理解力
• Qwen3-8b: 中型模型，具有较强的推理能力
• Qwen3-14b: 较大模型，提供优秀的理解和生成能力
• Qwen3-30b: 大型模型，适合复杂任务处理
• Qwen3-32b: 高性能大模型，强大的多任务能力
• Qwen3-234b: 超大规模模型，顶级性能表现

在本指南中，我们将使用Qwen3-4b版本进行演示，这是一个非常平衡的选择，既能提供良好的对话质量，又不会对普通设备造成过大负担。

一、Ollama安装Qwen3​

Ollama是一个开源的大模型运行框架，可以在本地部署运行多种大型语言模型，包括Qwen3。下面是安装和配置步骤：

1. 安装Ollama​

根据您的操作系统，选择相应的安装方法：

MacOS:

curl -fsSL https://ollama.ai/install.sh | sh

Linux:

curl -fsSL https://ollama.ai/install.sh | sh

Windows: 从Ollama官网下载并安装Windows版本。

2. 拉取Qwen3模型​

安装完成后，通过命令行拉取Qwen3模型：

# 拉取Qwen3 4b模型
ollama pull qwen3:4b

# 如果需要更大参数的模型，也可以选择其他版本
# ollama pull qwen3:8b
# ollama pull qwen3:14b

3. 验证模型安装​

通过以下命令验证Qwen3模型是否安装成功：

ollama list

您应该能看到已下载的qwen3模型列表。

4. 启动Ollama服务​

确保Ollama服务正在运行：

# 在某些系统上，安装后会自动启动服务
# 如果没有自动启动，请使用以下命令
ollama serve

默认情况下，Ollama服务会在http://localhost:11434端口运行。

二、在微语管理后台设置Qwen3对话模型​

完成Ollama和Qwen3模型的安装后，我们需要在微语管理后台进行配置：

1. 登录微语管理后台​

访问您的微语管理后台，输入账号和密码登录系统。

2. 导航到AI设置​

在左侧导航栏中，找到并点击"AI助手"->"机器人"选项。

图1：导航到AI设置

3. 切换Qwen3模型配置​

在AI设置页面中：

图2：切换Qwen3模型配置

1. 点击"AI模型选择"按钮
2. 选择模型类型为"Ollama->Qwen3"
3. 确认无误后，点击"确定"按钮

三、开始使用Qwen3进行智能对话​

配置完成后，您可以开始体验Qwen3赋能的智能客服功能：

1. 创建知识库（可选）​

为了让AI回答更加准确，您可以创建和维护特定领域的知识库：

1. 导航到"知识库"或"AI训练"模块
2. 点击"新建知识库"，输入名称和描述
3. 上传文档或手动添加Q&A对，丰富AI的专业知识

2. 测试对话效果​

您可以通过以下方式测试Qwen3的对话能力：

1. 在管理后台的"对话测试"功能中，输入问题进行测试
2. 通过客服端应用，模拟用户提问，验证AI回复效果
3. 通过访客端，体验实际用户视角下的AI交互

3. 对话效果展示​

以下是一些使用Qwen3进行智能对话的演示截图：

图3：Qwen3能够根据上下文提供连贯的多轮对话

四、优化和调整​

为了获得最佳的Qwen3对话效果，您可以进行以下优化：

1. 调整模型参数​

根据实际需求调整模型参数，如温度值、最大token数等，以平衡回答的创造性和精确性。

2. 优化系统提示词​

系统提示词对AI的行为有重要影响，您可以根据业务场景定制专业的提示词，引导AI表现出理想的对话风格。

3. 结合人工审核​

设置人工干预机制，对AI无法准确回答的问题进行人工接管，并将这些案例记录下来用于进一步训练和优化。

总结​

通过将微语客服系统与通义千问Qwen3大模型对接，您可以显著提升客服自动化水平和用户体验。本指南详细介绍了从安装Ollama、配置Qwen3模型到实际应用的完整流程。

随着您不断优化提示词和积累领域知识库，AI助手的表现会越来越符合您的业务需求，为客户提供更加专业、高效的服务体验。

如有任何问题或需要进一步的技术支持，请随时联系我们的技术团队。

希望本指南对您成功部署和使用微语+Qwen3智能客服系统有所帮助！

QR Code Login Implementation Process

Here are a few tips you might find useful.

Simply add Markdown files (or folders) to the blog directory.

Regular blog authors can be added to authors.yml.

The blog post date can be extracted from filenames, such as:

• 2019-05-30-welcome.md
• 2019-05-30-welcome/index.md

A blog post folder can be convenient to co-locate blog post images:

The blog supports tags as well!

And if you don't want a blog: just delete this directory, and use blog: false in your Docusaurus config.

For example, use JSX to create an interactive button:

<button onClick={() => alert('button clicked!')}>Click me!</button>

Use a <!-- truncate --> comment to limit blog post size in the list view.

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Pellentesque elementum dignissim ultricies. Fusce rhoncus ipsum tempor eros aliquam consequat. Lorem ipsum dolor sit amet

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Pellentesque elementum dignissim ultricies. Fusce rhoncus ipsum tempor eros aliquam consequat. Lorem ipsum dolor sit amet

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Pellentesque elementum dignissim ultricies. Fusce rhoncus ipsum tempor eros aliquam consequat. Lorem ipsum dolor sit amet

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Pellentesque elementum dignissim ultricies. Fusce rhoncus ipsum tempor eros aliquam consequat. Lorem ipsum dolor sit amet

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Pellentesque elementum dignissim ultricies. Fusce rhoncus ipsum tempor eros aliquam consequat. Lorem ipsum dolor sit amet

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Pellentesque elementum dignissim ultricies. Fusce rhoncus ipsum tempor eros aliquam consequat. Lorem ipsum dolor sit amet

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Pellentesque elementum dignissim ultricies. Fusce rhoncus ipsum tempor eros aliquam consequat. Lorem ipsum dolor sit amet

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Pellentesque elementum dignissim ultricies. Fusce rhoncus ipsum tempor eros aliquam consequat. Lorem ipsum dolor sit amet

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Pellentesque elementum dignissim ultricies. Fusce rhoncus ipsum tempor eros aliquam consequat. Lorem ipsum dolor sit amet

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Pellentesque elementum dignissim ultricies. Fusce rhoncus ipsum tempor eros aliquam consequat. Lorem ipsum dolor sit amet

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Pellentesque elementum dignissim ultricies. Fusce rhoncus ipsum tempor eros aliquam consequat. Lorem ipsum dolor sit amet

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Pellentesque elementum dignissim ultricies. Fusce rhoncus ipsum tempor eros aliquam consequat. Lorem ipsum dolor sit amet

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Pellentesque elementum dignissim ultricies. Fusce rhoncus ipsum tempor eros aliquam consequat. Lorem ipsum dolor sit amet

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Pellentesque elementum dignissim ultricies. Fusce rhoncus ipsum tempor eros aliquam consequat. Lorem ipsum dolor sit amet

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Pellentesque elementum dignissim ultricies. Fusce rhoncus ipsum tempor eros aliquam consequat. Lorem ipsum dolor sit amet

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Pellentesque elementum dignissim ultricies. Fusce rhoncus ipsum tempor eros aliquam consequat. Lorem ipsum dolor sit amet

...consectetur adipiscing elit. Pellentesque elementum dignissim ultricies. Fusce rhoncus ipsum tempor eros aliquam consequat. Lorem ipsum dolor sit amet