DeepSeek Enterprise Application Implementation: How Intelligent Customer Service Empowers Tech Companies

ZStack, a company with deep technical roots, has embraced the AI era. Beyond offering clients AI platforms and solutions, it continually innovates internally with AI technology. SupportAI emerged in this context as an innovation in customer service at ZStack. It has been successfully operational for over two years.

SupportAI’s key capabilities include:

• Extracting internal data (website, tickets, forums, Jira, Confluence) to enhance RAG capabilities
• Seamlessly integrating with DingTalk, WeChat Work, WeChat customer service, and other systems
• Providing exceptional support to customers, sales, pre-sales, and technical teams
• Ensuring smooth information flow and efficient collaboration via intelligent agents and ChatRobot

The release of DeepSeek-R1 has brought new opportunities for SupportAI, equipping it with a “smarter” brain. With joint efforts from the technical and business teams, SupportAI has now integrated the latest DeepSeek-R1 model.

This article will share the journey from SupportAI’s design in 2022 to its recent integration with DeepSeek. We hope our experience can assist more companies in their AI-driven digital transformation.

I. Why Build SupportAI?

Since 2015, ZStack has deployed over 600 trusted clouds, aiding more than 4,000 companies in digital transformation. This journey has amassed rich practical experience and best practices. As the business grew, the technical support department faced increasing pressure, with the existing customer service system encountering new challenges.

In 2022, the rise of large models like OpenAI’s sparked an opportunity for innovation in technical services. ZStack quickly mobilized its technical team to develop SupportAI. By integrating data from the website, enterprise tickets, internal forums, Jira, and Confluence, SupportAI processed and fine-tuned this data for a large model. This brought AI customer service capabilities to the technical support team. After initial exploration and patchwork efforts, it transitioned to ZStack’s AIOS Helix platform. Continuous optimization of model deployment and evaluation improved support effectiveness and response accuracy. Today, SupportAI is a core asset for the technical service department.

While addressing ZStack’s own business needs, SupportAI also serves as a practical case study in applying large models to enterprise AI solutions. This article details SupportAI’s design and steps, hoping to benefit other companies facing similar challenges.

II. Version Evolution: Technical Breakthroughs Behind SupportAI

(1) 2022 MVP Version

Before the 2022 large model boom, ZStack used automated chatbots to help engineers retrieve issues efficiently, linking relevant documents via a recommendation system. With ChatGPT and domestic models emerging, the technical team began testing large model-based solutions. After evaluating multiple models, including overseas options, Qwen2 was chosen as the foundation for the first SupportAI version due to its superior handling of Chinese semantics.

The middle layer used FastGPT to orchestrate workflows and load knowledge base documents. The base layer relied on Qwen for Q&A capabilities. This was the initial architecture. The MVP version offered limited efficiency gains for engineers. It primarily validated technical feasibility and closed the workflow loop. However, it occasionally delivered irrelevant or low-quality responses.

(2) 2024 AIOS Helix-Based Version

In 2024, ZStack launched AIOS Helix, enabling easy resource scheduling, model loading, evaluation, fine-tuning, and inference service deployment. After internal research, SupportAI was rebuilt on Helix as its AI foundation.

With MVP experience and collaboration between developers and service staff, the migration, data fine-tuning, and deployment were completed within a week. Since going live, it has been continuously optimized. Thanks to AIOS’s model management and high availability, the application remains operational during upgrades.

Helix provides modules for model management, selection, data engineering, hardware computing, operations, and API control. Before Helix, these required manual development and maintenance. With Helix, SupportAI frees up manpower to focus on improving response quality, accuracy, and business applications.

Built-in components like FastGPT and Dify in AIOS Helix offer web interfaces for workflow orchestration, knowledge base uploads, release management, and model switching. This simplifies SupportAI’s development and maintenance.

III. Pathways to Precise Responses

The MVP version’s average response quality stemmed largely from poor raw data quality. ZStack engineers tackled this by cleaning data and refining question categorization to boost knowledge base quality and response effectiveness.

(1) Improving Database Quality Through Data Cleaning

• Data Collection: Sources were diverse, with internal documents in multiple formats across platforms—PDFs, Word files, DingTalk docs, Confluence, internal BBS, ticketing systems, R&D knowledge bases, Jira, and the website. Local files were converted to structured data using tools. Some were shifted online. Online docs were processed with custom scripts, unifying formats and access interfaces.
• Data Cleaning: Custom scripts cleaned data, fixing typos, casual phrasing, redundant content, and off-topic material, yielding a high-quality knowledge base.
• Embedding Choices: Direct token-based chunking preserved details but hurt retrieval precision. QA extraction let the model pull questions from docs, improving search accuracy but losing some detail. Best practice combined both approaches.

(2) Enhancing Response Quality Through Question Categorization

Post-processing yielded a high-quality knowledge base, markedly improving Q&A. User questions spanned company intros, private cloud solutions, CPU overload fixes, ZStack Cloud features, and project integrations. Engineers pre-labeled knowledge base categories—often aided by the model. Broad queries like “How’s ZStack?” were enriched with context (if available) for a fuller prompt, then routed to relevant category-specific bases.

For example, after a technical query, the prompt might become: “I’m a technician who asked about CPU overload fixes. Now I want to know ‘How’s ZStack?’” The model then focuses on product and technical details. For investor queries, it highlights investment aspects; for enterprise users, it emphasizes solutions and case studies.

This demanded a platform to manage multiple models, evaluate performance, and validate service quality. The MVP ran on internal servers with manual management, meeting initial needs but lacking scalability. This drove the shift to AIOS Helix.

(3) Ensuring Generated Content Safety and Accuracy

Safety and accuracy are critical. Model “hallucinations” can produce irrelevant, inaccurate, or harmful content—especially problematic for external users. For enterprises, this is a key concern. Factors like poor document quality, ambiguous user inputs, or insufficient computing power can disrupt generation.

Generated Content Safety & Accuracy Check Steps

ZStack implemented these safety strategies in SupportAI:

• Prompt Restrictions: Optimized prompts explicitly block sensitive info, using a sensitive word/keyword library for the model to check outputs.
• Content Safety Filtering: SupportAI uses AIOS to fine-tune a model for safety checks. It also partners with third-party safety firms via API for enhanced detection.
• Manager Forwarding: Early on, quality improvement was gradual. Initially, internal staff manually reviewed and forwarded RAG-processed content to users. Now stable, SupportAI is gradually rolling out to external users.

IV. The Arrival of DeepSeek-R1

SupportAI needed a reasoning-capable model for intent analysis and accurate responses—a gap DeepSeek-R1 fills with industry-leading reasoning. Applying it to customer service was a bold challenge. The technical team promptly studied DeepSeek-R1, adding it to AIOS’s model repository. SupportAI tested versions in ZStack’s private environment. With the shift from physical servers to Helix for model management, the evaluation and replacement process was seamless.

(1) Model Testing and Selection

DeepSeek’s main versions were deployed and tested. Its reasoning chain suited answer generation but slowed response time, unfit for question completion. Helix’s evaluation tools assessed DeepSeek’s capabilities. Users selected active inference services, datasets, and sample sizes, configuring resources as needed.

(2) Model Fine-Tuning

After selecting the cost-effective DeepSeek 32B for general enterprise Q&A, existing datasets were loaded onto the base model for fine-tuning into an industry-specific model. This process, involving resource allocation and data loading, was streamlined via Helix’s web interface, freeing users from infrastructure concerns.

(3) Model Management and Selection

Post-tuning results are a “blind box” needing validation. A platform to manage running, testing, and fine-tuning models—plus future new releases—was critical. Helix’s model repository offers built-in models, local imports, and integrations with Hugging Face and ModelScope.

(4) Model Usage

Optimal scenarios vary: Qwen excels at question completion, DeepSeek at generation. A multi-model approach maximizes strengths. Front-end Q&A decoupled from back-end models allows API updates or load-balanced switches. Post-tuning, Helix deploys inference services, selecting models, templates, and deployment options (containers or VMs) with resource specs. Once running, interaction is via API or web.

V. Enterprise AI Application Implementation Insights

(1) Multi-Model Management is Essential—Recommend Helix

ZStack Helix, a next-gen AI infrastructure OS, accelerates enterprise AI adoption by restructuring compute, model, and operations layers. It supports upgrades from traditional to intelligent clouds, compatible with NVIDIA, Hygon DCU, and other GPUs. It enables fine-tuning of thousands of open-source models with data privacy and multi-tenant isolation. With low costs, dynamic resource allocation, and full lifecycle services, it’s scaled in GPU vendors, universities, and media. On February 2, Helix supported DeepSeek V3/R1/Janus Pro, deployable privately on Hygon, Ascend, NVIDIA, and Intel CPU/GPUs.

SupportAI’s journey began with internal servers and Qwen for knowledge Q&A. As scenarios diversified, managing active and testing models grew complex. A platform like Helix became vital to streamline model management, hide resource scheduling, and provide monitoring data—forming the backbone of enterprise AI solutions.

(2) Achieving 4 “Readys” for AI Adoption

Many firms have dormant data—website content, internal knowledge bases—that DeepSeek can unlock. Enterprise AI will deeply integrate into existing systems. To digitally transform, firms need four “Readys”: team, resources, data, and capabilities.

Focus should be on efficiency, customer satisfaction, and business reach—not infrastructure details, which expert teams can handle. A flexible, easy-deployment platform like Helix, leveraging heterogeneous resources, is key. Applications can use ModelScope or Hugging Face models, managed and deployed via Helix, using existing or new servers.

Enterprises must seize the AI wave. Missing it isn’t an option. Challenges abound in AI adoption, but ZStack—experts in DeepSeek privatization and enterprise AI infrastructure—offers robust support.