LLM API Access Glossary

The base address used by clients when sending API requests. AIFast's OpenAI Compatible Base URL is https://www.aifast.club/v1; if the client automatically appends /v1, it needs to be adjusted according to the final request address.

The path after the Base URL to perform the specific action, such as /models, /chat/completions, or /responses. The request body and return structure may be different for different endpoints.

A secret used to identify an account and authorize requests. Store it in server-side environment variables or a secrets manager, never in frontend code, public repositories, screenshots, or public logs.

Common HTTP authentication methods. The request header is usually written as Authorization: Bearer YOUR_API_KEY; adding Bearer repeatedly or missing spaces may result in 401.

The machine-readable identifier that needs to be filled in the model field in the API request is not necessarily equal to the Chinese name or model series name on the web page. The real values ​​available for the current account should be obtained through the console or /v1/models.

Indicates that the API is compatible with part of the OpenAI SDK's capabilities in terms of address, authentication, or request structure. This does not mean that all models, endpoints, tool calls, and multi-modal functions are completely consistent.

A common text generation API that organizes conversation content in messages arrays. Before accessing, you still need to confirm whether the target model supports extended capabilities such as streaming output, tool calling, and image input.

Use input, output, and event streams to organize the API form of requests and responses. Clients such as Codex may rely on this API, and full compatibility cannot be judged just by the successful call of Chat Completions.

A mechanism by which the server sends events in segments over a persistent HTTP connection. LLM streaming output often uses SSE, and the client needs to parse the data field block by block and handle connection interruptions.

The model returns structured parameters, and it is up to the application to decide whether to execute a function, search, file, or other external tool. The success of ordinary text does not mean that the tool call link is necessarily compatible.

The unit of measurement used by the model when processing text. Input, output, cache and inference tokens may be counted in different ways. The actual usage should be based on the response usage, console bill and price page.

The total capacity boundary for input, history messages, and output that can be processed by a single request. Exceeding model or compatibility layer limits may return 400, truncate content, or cause significant increases in latency.

Limits set by the service on request frequency, concurrency, or usage. After receiving a 429, you should first check quota and concurrency before performing a capped exponential backoff.

The request exceeded client allowed time during the connect, read, or total task phase. Streaming generation and Agent tool tasks usually require separate read timeout and total task duration settings.

A strategy that gradually increases the retry interval, such as waiting for 1, 2, 4 seconds and setting a maximum number of times. Suitable for some 429, 5xx and transient network errors, not suitable for authentication or parameter errors.

An identifier used to locate a single request. When debugging, the request ID, time, model ID, status code, and desensitized error text should also be recorded.