Models¶

A ModelConfig stores request settings and translates provider payloads and responses. Endpoints and inference-capable routers cache configurations by model name. Changes apply to later requests for that name.

Getting a Configuration¶

import intuit;

auto ep = new OpenAI("http://localhost:1234");
ModelConfig cfg = ep.config("llama3");

ep.config(modelName) creates a configuration if the name isn't cached. It doesn't check whether the server offers that model. ep.configs() returns cached configurations without making a request. See Endpoints for discovery.

On an inference-capable router, router.config() returns the active model's configuration. router.config(modelName) gets or creates another configuration without activating it. Catalog-only routers don't support either operation.

The concrete endpoint determines the configuration type:

OpenRouter also uses SystemOneModelConfig, which inherits the common completion and embedding behavior. Zen selects ModelConfig or ClaudeModelConfig by model ID. Provider guides explain the differences.

Completion Parameters¶

These omission rules describe the base OpenAI-compatible payload builder. Claude and Qwen have additional rules.

Assign fields directly before making a request:

cfg.temperature = 0.7;
cfg.topP = 0.9;
cfg.maxTokens = 1024;
cfg.stop = ["END"];
cfg.seed = 42;

Completion result = completions(ep, cfg.name, "Write a short greeting.");

The base completion and embedding builders merge params last, so it can override generated fields:

import std.json : JSONValue;

cfg.params = JSONValue.emptyObject;
cfg.params["frequency_penalty"] = JSONValue(0.2);

Qwen adds its own fields after the base builder runs. Claude's completion builder doesn't merge params. Decision builders accept params but write their required model, input/state, and question fields afterward. Common sampling and completion limits aren't added to decision requests.

Structured Output¶

setResponseSchema builds the OpenAI-style response_format object from a schema name, JSON Schema object, and strictness flag (default true):

import std.json : JSONValue;

JSONValue schema = JSONValue.emptyObject;
schema["type"] = JSONValue("object");
schema["properties"] = JSONValue.emptyObject;
schema["properties"]["answer"] = JSONValue.emptyObject;
schema["properties"]["answer"]["type"] = JSONValue("string");
schema["required"] = JSONValue([JSONValue("answer")]);
schema["additionalProperties"] = JSONValue(false);

cfg.setResponseSchema("answer_schema", schema);

The model and server must support structured output. The base builder sends response_format, while Claude translates the schema to output_config.format. Intuit doesn't validate the returned value against the schema.

Primitive shorthand nodes such as JSONValue("string") in schema positions are normalized to { "type": "string" } before the request is sent. Strings in enum, required, descriptions, and other metadata are preserved.

Read a JSON answer with result.json. Invalid JSON text throws MalformedResponseException. Valid JSON can still violate the requested schema.

Tool Choice¶

Use setToolLiability to control whether the model may call tools:

cfg.setToolLiability("auto");

Accepted values are:

Any other value throws FormatException.

After registering a tool, use setRequiredTool to request that specific tool:

cfg.setRequiredTool(ep.tools.get("getWeather"));

Passing null or a tool without a name clears the tool choice. These settings ask the model to produce tool calls. They don't authorize execution in your application. See Tools for registration, execution policy, and manual handling.

Provider Settings¶

• Anthropic covers system prompts, thinking budgets, and Messages API translation.
• Qwen covers thinking controls, embedding settings, and XML tool calls.
• TypeSafe and OpenRouter cover System One decisions.
• Extending Intuit covers custom payload builders and parsers.