Endpoints¶

An endpoint connects to one provider API through IEndpoint. Requests name a model explicitly. The endpoint caches model settings and tools, but doesn't keep conversation history.

This guide finds a model and makes a short conversation through an OpenAI-compatible server. Start with Quick Start if you haven't installed Intuit or prepared a server.

Creating an Endpoint¶

Create an endpoint with the server's base URL and an optional API key.

import intuit;

auto ep = new OpenAI("http://localhost:1234");

Replace the URL with your running server's address. Use the base URL without /v1 or a trailing slash, because Intuit appends request paths directly. For hosted OpenAI, use new OpenAI("https://api.openai.com", "your-api-key").

Finding a Model¶

Inspect the models advertised by the server.

import std.stdio : writeln;

foreach (cfg; ep.available())
    writeln(cfg.name);

available() makes a network request and adds advertised models to the endpoint's cache. It returns the entire cache, which can also contain names you configured manually.

Choose an ID accepted by your server. The examples use llama3 as a placeholder. You can request a known model without calling available() first.

A Single Request¶

Pass a model name and a prompt to completions. This request doesn't retain conversation history.

Completion result = completions(ep, "llama3", "Why is the sky blue?");
writeln(result.text);

Continuing a Conversation¶

Keep a Context when requests need previous messages.

Context ctx;
ctx.system("You are a helpful assistant.");
ctx.user("What is D?");

Completion first = completions(ep, "llama3", ctx);
ctx.user("How does it compare to C++?");
Completion second = completions(ep, "llama3", ctx);
writeln(second.text);

Intuit appends the first assistant choice after each successful completion. The endpoint doesn't keep a separate copy of your history.

Request Settings¶

Get the cached model configuration and change its settings before making the request.

ModelConfig cfg = ep.config("llama3");
cfg.temperature = 0.7;
cfg.maxTokens = 1024;

The configuration is reused for later requests with that model name. Creating one doesn't verify that the server offers the model or supports every setting.

Providers¶

The interface exposes all request methods, but each provider can reject unsupported operations. Check the provider guide rather than assuming every endpoint supports every method.

Choosing an Endpoint¶

Use OpenAI for a compatible local or hosted server, Qwen for its additional settings and parsing, Claude for Anthropic's Messages API, and TypeSafe for System One decisions.

Intuit doesn't host services or load models. See Quick Start for prerequisites, Responses for embedding requests, and Decisions for typed question/answer requests.

Next Steps¶

• Models for common settings and structured output.
• Context for message types, history, and compaction.
• Responses for completion fields and embedding requests.
• Tools and Decisions for additional request workflows.
• Provider guides for authentication, formats, and supported operations.
• Errors for exception types and request timeouts.