Custom models

neuro-ts defaults to gpt-4o because it is cheap and fast enough for one-shot built-in simulation. You can change this globally or per call.

Change the default model

import { configureClient } from 'neuro-ts';

configureClient({
  apiKey: process.env.OPENAI_API_KEY,
  model: 'gpt-4o',
  temperature: 0.1,
  maxTokens: 2048,
});

Per-call override (low level)

The generated wrappers do not yet accept a per-call model parameter (kept intentionally tight to mirror the original signatures). To override per call, drop down to the client:

import { getClient } from 'neuro-ts';
import prompts from 'neuro-ts/prompts';

const entry = prompts['neuro.array.map'];

const result = await getClient().executeFunction({
  functionId: entry.functionId, // 'Array.prototype.map'
  systemPrompt: entry.systemPrompt, // generated system prompt
  signatureHint: entry.overloads, // TypeScript overload signatures
  prompt: 'classify each item as fruit or vegetable',
  instance: ['apple', 'carrot', 'banana'],
  args: { callbackfn: undefined },
  model: 'gpt-4o',
});

signatureHint is the list of TypeScript overload signatures for the method. It is embedded in the system prompt to help the model understand expected parameter and return types. Always source it from prompts.json rather than passing [] — an accurate hint measurably improves output quality for complex signatures.

OpenAI-compatible endpoints

Anything OpenAI-Chat-Completions-compatible works (vLLM, Ollama with the OpenAI shim, Azure OpenAI, OpenRouter, Together). Pass baseURL:

configureClient({
  apiKey: process.env.OPENROUTER_API_KEY,
  baseURL: 'https://openrouter.ai/api/v1',
  model: 'meta-llama/llama-3.3-70b-instruct',
});

configureClient({
  apiKey: process.env.OLLAMA_KEY ?? 'ollama',
  baseURL: 'http://localhost:11434/v1',
  model: 'llama3.3',
});

Smaller or weaker models will struggle to follow the JSON-only output contract embedded in every system prompt. If you see prose coming back, raise maxTokens, lower temperature, or use a stronger model.