3.12.3 AI Engine Configuration
The AI Engine section is the core of model management in RDK Studio. This section provides a complete reference for model entry fields, dual-lane routing, and protocol determination rules. The quick setup process described in Section 2.5: Connecting AI Models is expanded here into detailed documentation.
Built-in Default Model
RDK Studio provides all logged-in users with an officially recommended model. After logging in with a D-Robotics account, you can use this model immediately without configuring an API Key. Requests to this model are routed through D-Robotics' official gateway, which has been optimized specifically for RDK development scenarios and treats internal company accounts and external developer accounts equally.
If developers only need an "out-of-the-box" experience, they can simply use the default model without creating any custom model entries.
Custom Model Entries
External developers or teams who wish to use specific models can create custom model entries. Each entry supports the following fields:
| Field | Required | Description |
|---|---|---|
| Label | Yes | Display name of the entry, used for identification in dropdown menus |
| Provider | Yes | Vendor identifier that determines the API protocol used (see "Protocol Determination" below) |
| Model | Yes | Vendor's model ID (must be exact and case-sensitive) |
| API Key | Yes | Key obtained from the vendor’s console |
| Base URL | No | Endpoint URL; leave blank to use the vendor’s default |
| Temperature | No | Sampling temperature: 0 = deterministic, 1+ = more creative; default is 0.7 |
| Top-P | No | Nucleus sampling threshold; default is 1.0 |
| Reasoning visibility | No | Whether to show reasoning steps to the user: inline / hidden |
Protocol Determination Rules
Studio determines which API protocol to use based on the Provider field (not the URL):
| Provider Field Value | Protocol Used | Authentication Method |
|---|---|---|
openai / qwen / doubao / gemini / deepseek / moonshot / siliconflow / ollama / openai-compatible | OpenAI Completions | Authorization: Bearer <key> |
anthropic / anthropic-compatible | Anthropic Messages | x-api-key: <key> |
Protocol determination does not depend on whether the URL path contains the word anthropic. If you use a reverse proxy to wrap an Anthropic service under a path that doesn’t include "anthropic," you must still set the Provider to anthropic-compatible; otherwise, Studio will send requests using the OpenAI protocol and receive a 401 error.
This logic is implemented in the resolveProtocol() function in server/agent/provider-setup.ts.
Dual-Lane Routing
RDK Studio implements two dedicated model lanes: Thinking and Quick:
| Lane | Trigger Scenario | Recommended Models |
|---|---|---|
| Thinking | Main conversations, complex reasoning, hardware awareness, and tool invocation | Claude Sonnet, qwen-plus, doubao-seed-pro, etc. |
| Quick | Summarizing tool results, file browsing summaries, short Q&A | gpt-4o-mini, qwen-turbo, doubao-seed-lite, etc. |
At the top of the AI Engine section, there are two separate dropdown menus to assign model entries to the Thinking and Quick lanes, respectively. The two lanes can use models from different vendors.
If only the Thinking lane is configured and the Quick lane remains empty, all tasks will be routed through the Thinking lane, significantly increasing token costs (5–10× higher). It is strongly recommended to configure both lanes, using a cost-effective small model for the Quick lane.
Connectivity Testing
Each model entry can be tested independently for connectivity:
| Action | Path |
|---|---|
| Test Entry | Click the Test Connectivity button on the right side of the model entry |
| Success Response | Displays a list of available models, confirming that Provider, Model, API Key, and Base URL are correctly configured |
| Failure Response | Shows specific error codes (e.g., 401, 403, 404, timeout); troubleshoot based on the error code |
Advanced Parameters
| Parameter | Function | Default Value |
|---|---|---|
| Temperature | Sampling temperature: 0 = fully deterministic, higher values = more creative | 0.7 |
| Top-P | Nucleus sampling threshold controlling output diversity | 1.0 |
| Thinking Default On | Whether this model defaults to deep-thinking mode | Follow system setting |
| Reasoning visibility | Whether to display reasoning steps | inline |
Adjusting these parameters affects the model’s response style and quality. It is recommended to keep the default values unless specific adjustments are truly needed.
Configuration Import and Export
Model entries can be imported and exported in bulk. See Section 3.12.5: Configuration Import and Export for details.