LLM Technical Service Description
1. Service overview
1.1 Service description
Agent Building Technical Service:
This service builds agents based on your business scenarios. It includes business scenario analysis, agent architecture design, engineering pipeline implementation, application integration, and the design and implementation of high availability and observability architectures.
Agent Optimization Technical Service:
This service optimizes agents based on your business scenarios. It includes agent architecture design, engineering pipeline optimization, and performance tuning.
LLM Post-training Technical Service:
This service uses post-training techniques to improve the performance of Large Language Models (LLMs) on vertical domain tasks. This service enhances metrics such as accuracy, recall rate, and response speed. It also reduces inference costs and resource consumption and improves the reliability, consistency, and security of model outputs.
LLM Computing Power Optimization Service:
This service provides computing power migration, environment adaptation, performance tuning, and deployment optimization for your model training and inference needs. This service helps you migrate algorithm tasks to the Alibaba Cloud platform and efficiently deploy them for training and inference using PAI and ACS products.
LLM Technical Supplementary Service:
This service supplements the preceding service packages with services not covered by a single package.
AI Product Training Service:
This service provides best practices training and workshops for the Alibaba Cloud AI products that you purchase. This helps you make better use of Alibaba Cloud products.
2. Scope of service
2.1 Scope of agent building technical service
The scope of this service includes:
Analyzing requirements for LLM-related application scenarios.
Designing, orchestrating, and implementing agent architectures.
Designing and implementing agent engineering pipelines.
Designing and implementing Context Engineering.
Building and implementing Retrieval-Augmented Generation (RAG) systems.
Integrating tools or skills.
Administering performance and stability.
Handling security and compliance.
Supporting project process management and the creation of implementation plans.
The scope of this service does not include:
Developing code for customer systems.
Daily O&M services for third-party software, such as installation, testing, troubleshooting, and optimization.
IT system architecture transformation or general digital transformation consulting not related to LLMs.
Note: For issues with third-party software not covered by Stabilization Services, you can log on to the Alibaba Cloud community for a free consultation or contact an Alibaba Cloud Marketplace vendor for assistance.
2.2 Scope of agent optimization technical service
The scope of this service includes:
Analyzing requirements for LLM-related application scenarios.
Performing in-depth optimization of Context Engineering.
Building and optimizing RAG systems.
Designing and orchestrating agent architectures.
Tuning agent engineering pipelines.
Performance tuning.
Handling security and compliance.
Supporting project process management and the creation of implementation plans.
The scope of this service does not include:
Developing code for customer systems.
Daily O&M services for third-party software, such as installation, testing, troubleshooting, and optimization.
IT system architecture transformation or general digital transformation consulting not related to LLMs.
2.3 Scope of LLM post-training technical service
The scope of this service includes:
Investigating business scenario requirements.
Designing post-training plans.
Dataset preparation, including training data and evaluation data.
Performance evaluation and continuous iteration.
Inference performance optimization.
Supporting project process management and the creation of implementation plans.
The scope of this service does not include:
Code development or feature modification for your existing business systems.
Installation, testing, troubleshooting, or optimization of third-party closed-source software.
Note: This service only supports optimization based on existing training and inference frameworks, such as Model Studio, PAI-LightLLM, vLLM, and sglang. It does not include framework-level modifications.
2.4 Scope of LLM computing power optimization technical service
The scope of this service includes:
Inference performance optimization, model image adaptation, and dependency environment compilation. This does not include modification of closed-source packages.
Deploying and stress testing inference services based on PAI or ACS.
Training environment setup, training efficiency optimization, and distributed strategy tuning.
Identifying abnormal issues and providing suggestions to improve computing power utilization.
Supporting project process management and the creation of implementation plans.
The scope of this service does not include:
Client-side code, model, and algorithm development.
Migration or modification of business systems not related to computing power.
Modification or optimization of third-party closed-source software or self-developed frameworks.
Underlying reconstruction of third-party training or inference frameworks, such as modifying the PyTorch kernel.
Note: This service only supports tuning based on existing primary stream training and inference frameworks, such as HuggingFace, DeepSpeed, and vLLM. It does not involve custom development at the framework level.
2.5 Scope of LLM technical supplementary service
The scope of this service includes:
Services related to agent building, including the following:
Analyzing requirements for LLM-related application scenarios.
Designing and implementing in-depth plans for Context Engineering.
Implementing RAG systems.
Designing agent architectures.
Implementing agent engineering pipelines.
Handling security and compliance.
Supporting project process management and the creation of implementation plans.
Services related to agent optimization, including the following:
Analyzing requirements for LLM-related application scenarios.
Performing in-depth optimization of Context Engineering.
Building and optimizing RAG systems.
Designing and orchestrating agent architectures.
Tuning agent engineering pipelines.
Performance tuning.
Handling security and compliance.
Supporting project process management and the creation of implementation plans.
Post-training technical services, including the following:
Investigating business scenario requirements.
Designing post-training plans.
Dataset preparation, including training data and evaluation data.
Performance evaluation and continuous iteration.
Supporting project process management and the creation of implementation plans.
Services related to computing power optimization, including the following:
Model image adaptation and dependency environment compilation. This does not include modification of closed-source packages.
Deploying and stress testing inference services based on PAI or ACS.
Training environment setup, training efficiency optimization, and distributed strategy tuning.
Identifying abnormal issues and providing suggestions to improve computing power utilization.
Supporting project process management and the creation of implementation plans.
The scope of this service does not include:
Developing code for customer systems.
Daily O&M services for third-party software, such as installation, testing, troubleshooting, and optimization.
IT system architecture transformation or general digital transformation consulting not related to LLMs.
Code development or feature modification for your existing business systems.
Client-side code development or model algorithm R&D.
Modifying, installing, testing, troubleshooting, or optimizing third-party closed-source software or your self-developed frameworks.
Note 1: This service only supports optimization based on existing training and inference frameworks, such as Model Studio, PAI-LightLLM, vLLM, sglang, HuggingFace, and DeepSpeed. It does not involve framework-level modifications or custom development.
Note 2: For issues with third-party software not covered by Stabilization Services, you can log on to the Alibaba Cloud community for a free consultation or contact an Alibaba Cloud Marketplace vendor for assistance.
2.6 Scope of AI product training service
The scope of this service includes:
Best practices training and workshops for Alibaba Cloud AI products.
The scope of this service does not include:
Pre-sales consulting or proof of concept (POC) for Alibaba Cloud AI products.
Integration of Alibaba Cloud AI products with your internal systems.
Daily technical support, Q&A, or fault response for Alibaba Cloud AI products.
Consultation, Q&A, and training for non-AI Alibaba Cloud products.
Note: For issues with third-party software not covered by Stabilization Services, you can log on to the Alibaba Cloud community for a free consultation or contact an Alibaba Cloud Marketplace vendor for assistance.
3. Prerequisites
3.1 Prerequisites for agent building technical service
You must submit a service request at least 15 calendar days in advance to allow for a feasibility assessment and resource preparation.
You must provide the necessary access permissions to non-production environments, remote access channels, and contact information for key stakeholders.
You must appoint a project manager with decision-making authority to act as the primary contact.
You must purchase Alibaba Cloud computing power or use Alibaba Cloud MaaS services.
After Alibaba Cloud submits the implementation plan, you must confirm the plan in writing (including by email) within five working days.
3.2 Prerequisites for agent optimization technical service
You must submit a service request at least 15 calendar days in advance to allow for a feasibility assessment and resource preparation.
You must provide the necessary access permissions to non-production environments, remote access channels, and contact information for key stakeholders.
You must appoint a project manager with decision-making authority to act as the primary contact.
After Alibaba Cloud submits the implementation plan, you must confirm the plan in writing (including by email) within five working days.
3.3 Prerequisites for LLM post-training technical service
You must submit a service request at least 15 calendar days in advance. If large-scale data processing or computing power migration is involved, you must submit the request 30 days in advance.
You must provide the necessary non-production environments, data samples, API documentation, and remote access permissions.
You must appoint a project manager to participate in requirement alignment, plan reviews, and progress coordination throughout the process.
You must review and approve the Engineering Implementation Plan created by Alibaba Cloud. You must not reject a confirmed plan without a valid technical reason.
You must participate in key node reviews and trial run validations at the agreed-upon times.
Alibaba Cloud does not design reinforcement learning reward functions. You must provide the reward function.
3.4 Prerequisites for LLM computing power optimization technical service
You must submit a service request at least 15 calendar days in advance. If large-scale computing power migration is involved, you must submit the request 30 days in advance.
You must provide model code, training scripts, data samples, and target hardware specifications.
You must grant remote access permissions to allow the Alibaba Cloud team to enter your environment for diagnostics.
You must clearly define performance optimization goals, such as the expected QPS and latency threshold.
You must cooperate in completing migration drills and performance validation.
3.5 Prerequisites for LLM technical supplementary service
You must submit a service request at least 15 calendar days in advance. If large-scale computing power migration is involved, you must submit the request 30 days in advance to allow for a feasibility assessment and resource preparation.
You must provide the necessary access permissions to non-production environments, remote access channels, and contact information for key stakeholders.
You must provide model code, training scripts, data samples, and target hardware specifications.
You must clearly define performance optimization goals, such as the expected QPS and latency threshold.
You must appoint a project manager with decision-making authority to act as the primary contact.
You must purchase Alibaba Cloud computing power or use Alibaba Cloud MaaS services.
You must cooperate in completing migration drills and performance validation.
After Alibaba Cloud submits the implementation plan, you must confirm the plan in writing (including by email) within five working days.
3.6 Prerequisites for AI product training service
You must submit a service request at least 15 calendar days in advance to allow for a feasibility assessment and resource preparation.
You must provide the necessary access permissions to non-production environments, remote access channels, and contact information for key stakeholders.
You must appoint a relevant person who understands the internal use of AI products to act as the primary contact. This person is responsible for arranging the training time and venue and coordinating the participants.
You must purchase Alibaba Cloud AI products, such as Qoder, Meoo, JVS Crew, and MuleRun.
After Alibaba Cloud submits the training plan, you must confirm the plan and schedule in writing (including by email) within five working days.
3.7 Division of responsibilities
3.7.1 Customer and Alibaba Cloud
The customer is responsible for purchasing the LLM technical service.
Both parties are responsible for agreeing on and confirming the specific business objectives and scope of the service.
3.7.2 Customer
Defining business objectives.
Providing the site, devices, necessary non-production environments, remote access channels, permissions, and other resources required for Alibaba Cloud to deliver the service.
Cooperating with Alibaba Cloud to investigate existing algorithm model training configurations and computing power utilization. This includes participating in the implementation of specific plans, such as creating deployment and migration plans.
Reviewing the implementation plan created by Alibaba Cloud and confirming it in writing or by email. You must not reject the technical suggestions or plans provided by Alibaba Cloud without a valid technical reason.
Executing specific migration, deployment, and optimization plans.
Acting as the O&M entity and being responsible for related O&M work.
Authorizing Alibaba Cloud to perform 24/7 monitoring and analysis of computing service performance.
3.7.3 Alibaba Cloud
Organizing the project and establishing an expert team for plan implementation.
Understanding your business objectives and scope, creating an implementation plan, and obtaining your written confirmation, which can be provided by email.
Providing the service catalog items specified in this work statement, such as business objectives, deployment plans, evaluation plans, and optimization plans. This includes providing feasible suggestions to ensure the normal use of computing resources.
3.7.4 Completion criteria
Alibaba Cloud submits the Service Acceptance Report and obtains customer acceptance.
3.8 Service items
Service content: Alibaba Cloud provides the following services to meet your business objectives:
3.8.1 Agent building technical service items
This service provides agent building technical services, including the following:
Scenario requirement investigation: Provides LLM scenario investigation and analysis services, including the following:
Interview the customer to confirm their business pain points and Large Language Model (LLM) application requirements.
Investigating available resources such as existing knowledge bases, data assets, and API operations.
Each package supports one business scenario and three investigation meetings.
Technical status investigation: Provides analysis services for the current status of LLM-related technologies, including the following:
Analyzing your current computing resources, network environment, and technical architecture.
Clarifying the access capabilities and interface specifications of enterprise knowledge sources, databases, and application systems.
Each package supports a technical integration assessment for three main systems.
RAG retrieval pipeline design and implementation, including the following:
This service provides vectorization optimization and RAG implementation for enterprise knowledge bases. This includes optimization plans for document ingestion and cleaning for common formats such as PDF, Word, and TXT. It also includes suggestions for segmentation, metadata extraction, and vector index configuration, along with RAG retrieval pipeline integration and performance validation.
Each package supports up to 50 standard documents, with each document being no more than 10 pages.
Agent technical pipeline and Context Engineering plan design, including the following:
This service delivers an Agent Pipeline Design Plan, which includes the agent orchestration pipeline, node design, model selection, evaluation methods, and standards.
This service designs the MCP tools or skills that the agent engineering pipeline depends on and defines the input and output standards for the tools.
Each package supports up to three agent nodes and two MCP tool or skill calls.
Agent and Context Engineering building and implementation, including the following:
This service includes lightweight memory engineering design and implementation. This does not involve memory scheduling, structured memory storage, memory compression, or memory organization.
This service designs prompts for the nodes in the orchestration plan and performs evaluation and optimization.
This service builds and connects the engineering pipeline based on Alibaba Cloud MaaS products, the agent orchestration plan, and Context Engineering. It also completes the integration with your existing tools or skills.
This service provides agent performance validation and a high availability deployment plan for the agent. It also supports you in completing production-level agent deployment and integration.
Provides project management services during project execution.
Each package supports one pipeline design plan, up to three agent nodes, two MCP tool calls, and 50 standard documents of no more than 10 pages each.
3.8.2 Agent optimization technical service items
1) Agent optimization consulting basic package
This package provides basic agent optimization consulting services, including the following:
Scenario requirement investigation: Provides LLM scenario investigation and analysis services, including the following:
Interviewing and confirming your business pain points and LLM application requirements.
Investigating available resources such as existing knowledge bases, data assets, and API operations.
Each package supports one business scenario and three investigation meetings.
Technical status investigation: Provides analysis services for the current status of LLM-related technologies, including the following:
Analyzing your current computing resources, network environment, and technical architecture.
Clarifying the access capabilities and interface specifications of enterprise knowledge sources, databases, and application systems.
Each package supports a technical integration assessment for five main systems.
Agent orchestration plan optimization design: Provides agent orchestration plan design services, including the following:
This service delivers an Agent Pipeline Design Plan, which includes the agent orchestration pipeline, node design, model selection, evaluation methods, and standards.
Each package supports one pipeline design plan with no more than five nodes.
In-depth optimization of Context Engineering: Provides in-depth, scenario-oriented prompt optimization services, including the following:
Rewriting prompts for compatibility with the Qwen model.
Prompt performance evaluation for up to five models and 1,000 data entries.
Each package supports up to three rounds of prompt optimization. Each round evaluates up to five models and 1,000 data entries.
Provides project management services during project execution.
2) Agent optimization implementation basic package
We offer agent services built on Alibaba Cloud Model Studio, which include the following:
This service provides agent performance optimization, testing, and publishing based on Model Studio. It supports calling your existing tools through the Model Studio MCP Server. Note that you must complete the tool integration yourself.
This service provides agent plan performance validation based on Model Studio and a high availability deployment plan for the agent. It also supports you in completing production-level agent implementation.
This service supports vectorization optimization and RAG optimization for enterprise knowledge bases. This includes optimization plans for document ingestion and cleaning for common formats such as PDF, Word, and TXT. It also includes suggestions for segmentation, metadata extraction, and vector index configuration, along with RAG retrieval pipeline integration and performance validation.
Provides project management services during project execution.
Each package supports up to five agent nodes, five MCP tool calls, and 100 standard documents of no more than 10 pages each. The basic package does not include the design of Multi-Agent, long-term or short-term memory, ReAct, CoT, Reflection, or Task Planning agents. To include these features, you must purchase a supplementary package.
3.8.3 LLM post-training technical service items
1) Post-training support package
This service provides supervised fine-tuning (SFT) or reinforcement learning from human feedback (RLHF) technical support for specific task objectives.
This service provides suggestions on data format specifications and sample quality inspection. It does not include fine-tuning data preparation.
This service configures training scripts using lightweight methods such as LoRA and P-Tuning.
This service assists in executing training tasks based on the computing resources that you provide.
Comparative evaluation of output performance.
Provides project management services during project execution.
Each package supports one training task, one model version, one algorithm such as LoRA, three model training iterations, and a data volume of 50,000 samples.
2) LLM evaluation support package
This service designs core metrics based on your evaluation objectives, such as accuracy, relevance, compliance, and response integrity.
This service builds an automated evaluation pipeline that supports batch inference, scoring, and result aggregation and analysis.
Delivers a multi-dimensional evaluation report that includes performance, problem attribution, and subsequent optimization suggestions.
Note: This service does not include evaluation dataset construction. You must provide your own dataset.
Provides project management services during project execution.
Each package supports evaluation for one business scenario. By default, a single package supports the following:
Up to 50,000 evaluation set items
Evaluation models: 3
Up to three evaluation rounds
3.8.4 LLM computing power optimization service items
1) Basic computing power optimization service
This service includes one of the following two options:
Image adaptation:
Adapting the model image based on the target GPU type.
Compiling dependency libraries such as CUDA, PyTorch, and Transformers.
Containerizing and packaging, which includes Dockerfile generation and image delivery.
Performing compatibility testing and identifying abnormal issues.
Inference service deployment:
Deploying the trained model as an online inference service. This supports PAI-EAS, Triton, and TGI.
Debugging API operations and supporting streaming responses.
Deploying the model with quantization (INT8 or GPTQ) based on the inference framework's capabilities to reduce GPU memory usage.
Performing stress testing to assess QPS, latency, and concurrent capabilities.
Delivering an Inference Performance Evaluation Report.
Provides project management services during project execution.
Each package supports one model instance and one type of GPU card.
Inference service deployment is billed separately for each environment, such as development, staging, and production.
2) Training efficiency optimization service
This service includes the content of the basic computing power optimization service, plus the following:
Performing end-to-end performance tuning for the training task.
Optimizing distributed training strategies, such as ZeRO-2/3, DDP, and FSDP.
Optimizing GPU memory usage and improving throughput.
Providing suggestions for Checkpoint management and fault tolerance mechanisms.
Delivering a Training Performance Optimization Suggestion Report.
Provides project management services during project execution.
Each package supports one training task, one model, and one type of GPU card.
3.8.5 LLM technical supplementary service items
Each supplementary service package covers one of the following services.
1) Agent building supplementary service
This service expands the capacity of the "Agent Building Technical Service" package. Each supplementary package supports one of the following:
Agent building consulting extension items (choose any two of the following):
One additional 60-minute investigation meeting
One additional system technical assessment
One additional node in the design capacity
One additional round of prompt optimization for one model
RAG building extension item:
Processing for 100 additional standard documents of 10 pages each
Agent building implementation extension items (choose one of the following):
Multi-Agent collaboration flow orchestration plan design and performance validation
Short-term or long-term memory mechanism design (Session and Vector DB) plan design and performance validation
Tool calling and function routing (Function Calling) plan design and performance validation
Enhanced autonomous decision-making (ReAct or CoT chain-of-thought) plan design and performance validation
Reflection and self-correction mechanism (Reflection) plan design and performance validation
Dynamic task decomposition and planning (Task Planning) plan design and performance validation
A maximum of one agent node and one MCP tool calling
Provides project management services during the execution of the supplementary service.
This service must be purchased with or after the "Agent Building Technical Service". It cannot be used alone to start a new project. It supports flexible combinations to meet lightweight incremental needs or supplement advanced capabilities.
2) Agent optimization supplementary service
This service expands the capacity of any service package in the "Agent Optimization Technical Service". Each supplementary package supports one of the following:
Agent optimization consulting extension items (choose any two of the following):
One additional 60-minute investigation meeting
One additional system technical assessment
One additional node in the design capacity
One additional round of prompt optimization for one model
RAG optimization extension item:
Processing for 100 additional standard documents of 10 pages each
Agent optimization implementation extension items (choose one of the following):
Multi-Agent collaboration flow orchestration plan design and performance validation
Short-term or long-term memory mechanism design (Session and Vector DB) plan design and performance validation
Tool calling and function routing (Function Calling) plan design and performance validation
Enhanced autonomous decision-making (ReAct or CoT chain-of-thought) plan design and performance validation
Reflection and self-correction mechanism (Reflection) plan design and performance validation
Dynamic task decomposition and planning (Task Planning) plan design and performance validation
Up to 1 agent node and 1 MCP tool calling.
Provides project management services during the execution of the supplementary service.
This service must be purchased with or after the "Agent Optimization Technical Service". It cannot be used alone to start a new project. It supports flexible combinations to meet lightweight incremental needs or supplement advanced capabilities.
3) LLM post-training supplementary service
This service expands the capacity of any service package in the "LLM Post-training Technical Service" to accommodate complex or progressive implementation needs. Each supplementary package supports one of the following extensions:
Supplement for the post-training support package (choose one of the following):
One additional model training iteration
Processing capacity for 50,000 additional fine-tuning data entries
Post-training data augmentation plan design
Supplement for the evaluation support package (choose one of the following):
Processing capacity for 50,000 additional evaluation samples
Support for two additional evaluation models
Two additional evaluation cycles
Provides project management services during the execution of the supplementary service.
This service must be purchased with or after the "LLM Post-training Technical Service". It cannot be used alone to start a new project. It supports flexible combinations to meet lightweight incremental needs or supplement advanced capabilities.
4) General computing power optimization supplementary service
This service expands the capacity of any service package in the "Computing Power Optimization Service" to accommodate complex or progressive implementation needs. Each supplementary package supports one of the following extensions:
Supplement for the basic computing power optimization service (choose one of the following):
One additional model type
One additional GPU type
One additional deployment environment
Training efficiency optimization service (choose one of the following):
One additional training task
One additional model
One additional type of GPU card
Provides project management services during the execution of the supplementary service.
This service must be purchased with or after the "LLM Computing Power Optimization Service". It cannot be used alone to start a new project. It supports flexible combinations to meet lightweight incremental needs or supplement advanced capabilities.
3.8.6 AI product training service items
This service provides training for Alibaba Cloud AI products, such as Qoder, Meoo, JVS Crew, and MuleRun. The training includes the following:
General best practices training and workshop services, which include the following:
General best practices training on the end-to-end use of the product.
Best practices training on product use for different roles.
General hands-on product workshop training.
Each package supports eight class hours, with a default of four class hours per day.
4. Service SLA
Technical Manager, Large Model Service.
This service provides an LLM Technical Service Work Plan and an LLM Technical Service Acceptance Report.
5. Service process
5.1 Agent building technical service process
Application deadline: You must submit your application at least 15 calendar days before the agent building technical service start date.
5.2 Agent optimization technical service process
Application deadline: You must submit your application at least 15 calendar days before the agent optimization technical service start date.
5.3 LLM post-training technical service process
Application deadline: You must submit your application at least 15 calendar days before the LLM post-training technical service start date.
5.4 LLM computing power optimization service process
Application deadline: You must submit your application at least 15 calendar days before the LLM computing power optimization service start date.
5.5 AI product training service process
Application deadline: You must submit your application at least 15 calendar days before the AI product training service start date.
6. Acceptance criteria
The service is considered accepted when Alibaba Cloud provides the following deliverables:
Alibaba Cloud delivers the LLM Technical Service Work Plan and the LLM Technical Service Acceptance Report and obtains your written confirmation, which can be provided by email.
The LLM Technical Service Work Plan and the LLM Technical Service Acceptance Report include the following:
An analysis of your LLM requirements and suggestions for LLM technology selection, provided by Alibaba Cloud before the service starts.
Technical plan suggestions for your business implementation, provided by Alibaba Cloud based on your business characteristics and requirements.
7. Completion criteria
The service is complete after the implementation is finished and customer acceptance is confirmed.