Key Concepts and Terminology#
Helps you quickly understand professional terms related to AI large modelsThis document uses plain language to explain various technical terms, making them easy to understand even without a technical background.
Core Concepts#
๐ค Large Language Model (LLM)#
Plain Explanation: A large model is like a "super-intelligent AI brain" that can understand and generate human language by learning from massive amounts of text data, completing various intelligent tasks.
โ Massive Scale: Contains billions or even trillions of parameters (think of them as "knowledge points")
โ Comprehensive Capabilities: Can handle text, code, translation, Q&A, and many other tasks
โ Continuous Learning: Constantly improves intelligence through training
๐ค Token#
Plain Explanation: A token is the smallest unit for AI to process text, similar to a "measurement unit" for text.
Token Calculation#
| Language | Token Calculation | Example |
|---|
| Chinese | Usually 1-2 characters = 1 token | "Hello world" โ 4 tokens |
| English | Usually 1 word = 1-2 tokens | "Hello World" โ 2 tokens |
| Code | Calculated based on characters and structure | print("Hello") โ 4-5 tokens |
Because AI service billing is typically based on token usage. The longer the input and output text, the more tokens consumed, and the higher the cost.
๐ API (Application Programming Interface)#
Plain Explanation: An API is a "bridge" for software communication, allowing your application to call Nexconn's AI services.
Analogy#
๐ช API is like a restaurant "menu"You (the application) order from the menu (API) to the kitchen (AI service), then receive your dish (AI response).Practical Purpose#
Through APIs, you can integrate AI capabilities into your website, app, or software without developing complex AI systems yourself.API Types Supported by Nexconn:OpenAI API: Compatible with OpenAI interface standards
Anthropic API: Compatible with Anthropic (Claude) interface standards
Standard REST API: Universal network interface protocol
๐ง Inference#
Plain Explanation: Inference is the process where an AI model "thinks" through input questions or instructions and provides answers.
Inference Process Example#
Question: "Help me write a thank-you letter"1.
Input Stage โ User submits question
2.
Inference Stage โ AI analyzes request and generates content
3.
Output Stage โ Generates complete thank-you letter text
Inference Service provides this AI "thinking" and "answering" capability as a cloud service.
๐ฏ Prompt#
Plain Explanation: A prompt is the "instruction" or "question" you give to the AIโthe content you input.
Comparison Examples#
| Type | Example | Evaluation |
|---|
| โ Poor Prompt | "Write something" | Too vague, AI doesn't know what to write |
| โ
Good Prompt | "Please help me write a 300-word product introduction for a smartwatch, highlighting health monitoring features" | Clear and specific, AI can accurately understand the requirement |
Tips for Writing Good Prompts:Clearly define the task objective
Provide sufficient background information
Specify output format or style
Give specific constraints
Advanced Features#
๐ค MCP (Model Context Protocol)#
Plain Explanation: MCP is a technology that gives AI "superpowers" by adding various tool capabilities.
Analogy#
If you think of AI as a smart assistant, then MCP equips this assistant with various "toolboxes":๐ Search Toolbox: Enables AI to search the web for latest information
๐ Data Toolbox: Enables AI to analyze and process data
๐ File Toolbox: Enables AI to read and edit files
๐ง Communication Toolbox: Enables AI to send emails or messages
AI can retrieve real-time stock prices (using financial data tools)
AI can help you book restaurants (using reservation system tools)
AI can generate and save documents (using file processing tools)
๐ง Agent#
Plain Explanation: An agent is an "AI assistant" that can autonomously complete complex tasksโit not only answers questions but also proactively uses tools to solve problems.
Regular AI vs Agent Comparison#
| Comparison | Regular AI | Agent |
|---|
| Working Method | Passively answers questions | โ
Proactively executes tasks |
| Capability Range | Mainly text generation | โ
Can use multiple tools |
| Task Complexity | Single-step simple tasks | โ
Multi-step complex tasks |
| Autonomy | Requires explicit instructions | โ
Can autonomously plan steps |
Agent Work Example#
Task: "Help me organize last week's sales data and send a report to my boss"1.
Read last week's sales data from database
2.
Analyze data and generate charts
๐ Web Search#
Plain Explanation: Enables AI to access the internet and retrieve the latest real-time information.
Comparison Examples#
Q: "What's the weather in Beijing today?"
A: "Sorry, I cannot retrieve real-time weather information..."
Q: "What's the weather in Beijing today?"
A: "Beijing is sunny today, temperature 15-25โ, air quality is good."
Query real-time information (weather, news, stocks, etc.)
Get latest materials (latest technical documentation, policies and regulations, etc.)
Verify factual information (check data, confirm message authenticity, etc.)
๐ญ Chain of Thought (Reasoning)#
Plain Explanation: Enables AI to "think step by step" like humans, rather than directly giving answers, thereby improving response quality for complex problems.
Comparison Example#
Question: "Xiao Ming has 15 apples, ate 3, bought 8 more, gave 5 to friendsโhow many does he have left?""Xiao Ming has 15 apples left." (Direct answer, prone to errors)
Let me calculate step by step:4.
Gave 5 to friends: 20 - 5 = 15
Reduces calculation errors
More rigorous logical reasoning
More transparent answer process
๐ฌ Context Memory#
Plain Explanation: AI can "remember" previous conversation content, enabling coherent multi-turn conversations.
Conversation Example#
๐ค You: "I want to travel to Japan"
๐ค AI: "Japan is a beautiful country! When are you planning to go?"
๐ค AI: "Spring is perfect for visiting Japanโyou can enjoy cherry blossoms! I recommend Tokyo, Kyoto, and Osaka."
๐ค You: "What do I need to prepare?"
๐ค AI: "For traveling to Japan, you need to prepare: 1. Passport and visa 2. Clothing suitable for spring 3. Japanese yen cash 4. Translation software..."
More natural conversation experience
No need to repeat background information
Supports complex multi-turn interactions
Billing Terms#
๐ฐ Pay-As-You-Go#
Meaning: Charged based on actual resource usage (token count)โpay for what you use.Analogy: Like utility billsโyou pay for the electricity you actually use; no usage means no charge.โ Controllable costs, suitable for scenarios with unstable usage
โ No prepayment required, lowers barrier to entry
โ Flexible scaling, use as needed
๐ฆ Resource Package#
Meaning: Purchase a certain amount of token quota in advance and enjoy discounted rates.Analogy: Like buying an annual gym membershipโpurchasing at once is more cost-effective than paying each time.โ Users with high usage frequency
โ Users seeking cost-effectiveness
โ Enterprises with stable usage needs
๐ Unified Token#
Meaning: A universal token quota that can be shared across multiple different AI models. Nexconn's AI large model unified resource package token count is based on DeepSeek-V3.1 as the baseline. When using other models, the token deduction will be proportionally adjusted based on different prices (i.e., more expensive models have higher deduction multiples, output context longer than input has higher deduction multiples, cheaper models have lower deduction coefficients). Through a unified deduction standard, you can freely switch between AI models at different price points and enjoy a convenient billing experience. Please log in to the Nexconn Large Model API Console to check token consumption and request records in real time.Analogy: Like recharging a "universal shopping card" that can be used at multiple stores, not limited to just one.โ Flexibly switch between different models
โ Avoid wasting quota on a single model
โ Simplify account management
Technical Architecture Terms#
๐ End-to-End Encryption#
Plain Explanation: Data is encrypted throughout transmissionโonly the sender and receiver can see the content; no one in between can intercept it.
All sensitive business data
๐ก REST API#
Plain Explanation: REST API is a web communication standard that enables different software systems to interact over the internet.
How It Works (Simplified)#
1.
Your application sends a request โ "Please help me generate an article"
2.
Nexconn AI service receives and processes the request
3.
AI returns a response โ Generated article content
4.
Your application displays the result to the user
Standardized: Follows unified specifications
Simple to Use: Easy to understand and implement
Cross-Platform: Supports various programming languages
๐ SDK (Software Development Kit)#
Plain Explanation: An SDK is an official "toolbox" containing ready-made code and tools that make it easier for developers to use AI services.
Comparison Example#
Need to write extensive code for network request encapsulation
Need to handle data format conversion
Need to implement error handling logic
Need to complete authentication and authorization flow
โฑ๏ธ Long development time, prone to errors
๐๏ธ Parameters#
Plain Explanation: Parameters are the "knowledge units" that make up an AI modelโmore parameters usually means a "smarter" model.
| Model Size | Parameter Count | Capability Characteristics |
|---|
| Small Model | Hundreds of millions | Fast response, low cost, suitable for simple tasks |
| Medium Model | Tens of billions | Balances performance and cost |
| Large Model | Hundreds of billions ~ trillions | Most capable, suitable for complex tasks |
Models with more parameters typically have higher computational costs and correspondingly higher API call fees.
๐ Fine-tuning#
Plain Explanation: Based on a general large model, conduct specialized training for specific domains or tasks to make it better at handling specific problems.
Analogy#
Like a general practitioner (general model) who, after specialized training, becomes a cardiologist (fine-tuned model), more professional in the field of heart disease.๐ Enterprise Customer Service: Train AI to specifically answer company product questions
โ๏ธ Legal Assistant: Train AI proficient in legal provisions
๐ฅ Medical Advisor: Train AI familiar with medical knowledge
๐ก๏ธ Temperature#
Plain Explanation: The temperature parameter controls the "creativity level" or "randomness" of AI responses.
Temperature Comparison#
| Temperature Range | Characteristics | Suitable Scenarios |
|---|
| Low Temperature (0-0.3) ๐ง | Stable, conservative, accurate responses | Data analysis, code generation |
| Medium Temperature (0.5-0.7) ๐ค๏ธ | Balanced, natural responses | Daily conversation, Q&A |
| High Temperature (0.8-1.0) ๐ฅ | Creative, diverse, random responses | Creative writing, brainstorming |
