OpenAI¶
OpenAI implements OpenAI-compatible chat completions, embeddings, and decisions. It can connect to a hosted API or a local server such as Ollama or LM Studio. The server and model determine which operations actually work.
Connection¶
import intuit;
auto local = new OpenAI("http://localhost:1234");
auto hosted = new OpenAI("https://api.openai.com", "your-api-key");
The constructor accepts url, an optional key, and an optional display name (default "OpenAI"). A nonempty key is sent as Authorization: Bearer <key>.
Use the server's base URL without /v1 or a trailing slash. Intuit appends paths directly and doesn't normalize URLs. For a proxy mounted under a prefix, include that prefix in the base URL.
| Operation | Path |
|---|---|
| Model discovery | GET /v1/models |
| Chat completions | POST /v1/chat/completions |
| Embeddings | POST /v1/embeddings |
| Decisions | POST /v1/decisions |
Intuit doesn't start a server, download models, or load them into memory. Start the service and make the model available before sending requests.
Models and Requests¶
available() reads model IDs from the response's data array and adds uncached ModelConfig instances. It returns all cached configurations, including manually configured names. It doesn't remove names no longer advertised by the server or follow pagination.
config(modelName) gets or creates a configuration without checking the server. Model IDs must match what your deployment accepts.
See Getting Started for completions, Models for request settings, and Responses for single and batch embeddings. Decision requests have their own guide.
Compatibility¶
OpenAI-compatible describes a payload shape, not identical behavior across servers. Structured output, tools, reasoning, embeddings, and decisions need support from the selected model and deployment. Sending an extra field doesn't make the server implement it.
The parser accepts ordinary message text, structured text/reasoning content, and native function tool calls. It preserves raw responses for provider fields outside the shared types. Qwen XML tool calls require the Qwen adapter.
Streaming and the OpenAI Responses API aren't implemented. legacyCompletions(ep, payload) still sends /v1/completions, but it's deprecated in favor of chat completions.