Developer FAQ
Applicable role: Developer Last updated: 2026-08-06
This document serves as an index of frequently asked developer questions, with each question linking to the relevant detailed documentation.
1. Integration & Authentication
Q: What should I enter for the Base URL?
Enter the root domain address https://<api-endpoint>; the SDK or client will append the request path according to the protocol. For details, see API Call Basics.
Q: Where do I create an API Key / how do I get one? Both admin users and member users can create keys on the "API Key" page in the console. When a member user creates a key, the routing groups are limited to the scope authorized by the admin user. See Quick Start.
Q: Getting a 401 error?
Check the Authorization: Bearer header format. See Error Codes & Troubleshooting.
2. Models & Calls
Q: What exactly should I put in the model field?
Enter the model name (e.g., deepseek-v4-flash). You can find available models in "Model Plaza," "Default Model Configuration," or via /v1/models. If the admin user has enabled group identifier routing (an advanced feature, disabled by default), you can copy a call name with a group identifier (e.g., 5MHXZWKA/deepseek-v4-flash) from the "Call Guide" page to target a specific group. See Call Guide & Routing.
Q: What is the difference between a model name and a name with a group identifier? A model name requires your key to be bound to a group: when the target model is within the bound group, it is used directly; only when the model is not in that group does the system fall back to the default model configuration. A name with a group identifier explicitly targets a specific group (requires group identifier routing to be enabled). See Call Guide & Routing.
Q: Why can't I access a certain model? The call name may be incorrect, or the group is not within your account's available scope (configured by the platform according to the admin user's settings). Verify your available scope or copy the correct name from the call guide.
Q: What is the difference between OpenAI format and Messages-compatible format? See API Call Basics, Section 5.
Q: How do I switch models?
Change the model field. For details, see API Call Basics.
3. Billing & Rate Limiting
Q: How is billing calculated? Who gets charged? Billing is based on the pricing of the group/account that the request actually hits (as shown in the "Call Guide"). The actual charge is always deducted from the parent organization's balance.
Q: Is billing different when using a group identifier vs. a model name? No — billing is unrelated to how you write the call name. It is determined by the group/account the request ultimately routes to. See Call Guide & Routing.
Q: What should I do about 429 rate limiting?
Reduce concurrency and use exponential backoff. Distinguish between rate limits (API_KEY_RATE_* — wait for the window to reset) and quota/concurrency limits (refer to the actual error codes returned by the API). See Error Codes & Troubleshooting.
Q: What is the concurrency limit? How can I increase it? The admin user-level concurrency limit is configured uniformly by the platform. Coordinate with your parent organization to request adjustments, and contact the platform if needed.
4. Stability
Q: How do I handle 502 errors? When a model provider returns a 5xx error, the platform uniformly maps it to 502. Use exponential backoff to retry. See Error Codes & Troubleshooting.
Q: What should I do if a call with a group identifier returns fingerprint_routing_disabled / group_not_allowed?
Group identifier routing is an advanced feature (disabled by default) that must be enabled by the admin user before it can be used. fingerprint_routing_disabled means it has not been enabled; group_not_allowed means the group is not within your available scope or is currently unavailable — the platform returns an error directly without fallback. Switch to an available group. See Call Guide & Routing.
Q: How do I distinguish between 500 and 502? 500 indicates a platform-side fault; 502 indicates a model provider-side fault. See Error Codes & Troubleshooting.
Q: The model occasionally outputs Korean/Japanese? This is model language drift. See Model Chinese Output Stability.